publication_metadata.json

{
  "record_type": "publication_metadata",
  "record_version": "1.1.0",
  "evidence_manifest_v1_0_0_sha256": "a3ecb53728523005648b995871b8389f40cfa997a45b7f58935e9cecc335b936",
  "evidence_manifest_v1_1_0_sha256": "4c5328b7ebf2e6f9a037ad0ad09f756c16c794bf2006c465a28ab394c4ee91d4",
  "evidence_manifest_v1_1_0_public_half_sha256": "e2a1e2507c6909f7159848d940f54c6be8c4c703621058e346a75582974a3eb1",
  "evidence_manifest_note": "Each value is the SHA-256 of the bytes of a checksum manifest file, one per retained evidence package. evidence_manifest_v1_0_0_sha256 is the 1.0.0 package, and is the value 1.0.0 published under the name evidence_manifest_sha256: unchanged, and carried here so the older commitment stays readable. evidence_manifest_v1_1_0_sha256 covers the whole 1.1.0 package, both halves; that manifest lists the filenames of private holdout material, so it is never distributed. evidence_manifest_v1_1_0_public_half_sha256 covers the public half of the same 1.1.0 package, which is the half that can be shown to a reviewer, so it is the only one of the three that a holder of the evidence can verify against. All three packages remain private; these hashes are commitments to their contents, not a means of independent verification.",
  "release_id": "CompilingThings/compile-benchmark-v1.1.0",
  "release_title": "CompilingThings Compile Benchmark for MQL5\u00ae",
  "release_repo_slug": "compile-benchmark",
  "metaquotes": {
    "trademark_attribution": "MQL5\u00ae and MetaTrader 5\u00ae are registered trademarks of MetaQuotes Ltd.",
    "no_endorsement": "CompilingThings is an independent project. No affiliation, sponsorship, certification, endorsement, or approval by MetaQuotes Ltd. is claimed.",
    "compiler_use": "Compilation validation was performed using the MetaQuotes Language Compiler supplied with MetaEditor. The compiler is used solely as the MQL5\u00ae compilation and validation instrument.",
    "no_redistribution": "This release does not distribute MetaTrader 5\u00ae, MetaEditor, the MetaQuotes Language Compiler, EX5 protection technology, MetaQuotes documentation, or other MetaQuotes-owned software or proprietary materials.",
    "source_ownership": "The benchmark prompts published in this release were produced by CompilingThings' own generators and are owned by CompilingThings. No model-generated MQL5 source or other model outputs are distributed. No ownership of the MQL5 language, the MetaTrader platform, MetaEditor, or MetaQuotes technology is claimed.",
    "marks_remain": "All MetaQuotes trademarks and MetaQuotes-owned technology remain the property of MetaQuotes Ltd."
  },
  "item_naming": {
    "item_id_purpose": "item_id is a stable public identifier used to join prompts and result rows.",
    "names_are_recoverable": "The original item name of every published item is recoverable: each prompt contains a `// NAME:` line, because that line was part of the prompt the models were given. Publishing the prompts as evaluated therefore publishes the names.",
    "why_not_removed": "Removing the names would make the published prompt differ from the evaluated prompt. Every hash in this release, and the claim that a reader is running the benchmark we ran, depends on those being the same bytes. The names are published rather than the benchmark being weakened.",
    "what_this_discloses": "The names disclose the corpus naming scheme, indicator families, parameter conventions and versioning for these 184 items. The corpus itself, and the names of every item not in this release, remain undistributed."
  },
  "positive_control": {
    "what_it_is": "Before an arm is scored, two fixtures are compiled. One is known to produce a compiler error. One is known to produce a compiler warning and still build.",
    "pass_condition": "The error fixture must report at least one error. The warning fixture must report at least one warning and must score as a headline pass and a strict failure - the one outcome that tells the two metrics apart.",
    "on_failure": "The harness raises and writes no rows for that arm. A scored arm is therefore an arm whose control passed; there is no run in which the control failed and results were kept anyway.",
    "why_every_row_reads_1_and_1": "The recorded counts come from the fixtures, not from the evaluated item, so they are the same on every row of an arm. A row reading 1 and 1 means the control passed for that arm. It is not a defect in that item.",
    "what_it_establishes": "That the compiler and the log scorer could still see an error and a warning during this run. It says nothing about any individual item's verdict."
  },
  "serving": {
    "scope": "The two bf16 local arms of the 1.0.0 release, and nothing else. The three local arms that produced every 1.1.0 row were served at Q8_0 through llama.cpp and are described under serving_v1_1. The frontier arm's serving is described under models.frontier.",
    "engine": "Hugging Face transformers, AutoModelForCausalLM.from_pretrained, generating with model.generate",
    "decoding": "Greedy. do_sample=False and top_p=1.0.",
    "temperature_note": "Both local arms record temperature 0.0 and temperature_applied UNSUPPORTED. Greedy decoding does not consume a temperature, so the requested value was recorded and had no effect. This is why the field reads UNSUPPORTED rather than APPLIED.",
    "seed": "torch.manual_seed is set per generation. Recorded as seed_applied APPLIED.",
    "compute_dtype": "bfloat16, passed as torch_dtype. No quantisation was applied at load time for either local arm.",
    "max_new_tokens": 8192,
    "library_versions_not_established": "The transformers, torch and tokenizer library versions in force during the evaluation run were not recorded in the run parameters and are not established. The merge-time environment is recorded, but that was a different process.",
    "evaluation_hardware": "Both local arms were generated on one AMD Radeon AI PRO R9700 (RDNA 4, gfx1201) under ROCm, in an AMD Ryzen 9 7950X3D host with 64 GB of system memory.",
    "hardware_note": "The device is not recorded in the run parameters; it is stated here from the evaluation host's configuration. An earlier version of this note inferred CPU execution from the per-item timing alone, which was wrong.",
    "compiler_build": 5836,
    "compiler_build_note": "The three 1.0.0 arms compiled against the same compiler build, and so did every arm added in 1.1.0. The build number is published; the compiler binary is not distributed and is not ours to distribute."
  },
  "attested_not_verifiable": {
    "statement": "The following are author attestations. They rely on undistributed evidence and cannot be verified from this package. Each claim is scoped: a claim that names a release, or names its own subject, is about that one alone, and every other claim describes an instrument or a procedure both releases share. 1.0.0 here means the frozen-300 evaluation and the 184-item public set carried forward from the 1.0.0 release; 1.1.0 means the two holdout sets, the Q8_0 bridge and the corpus row-hash manifest this release adds.",
    "claims": [
      "The per-item verdicts are the compiler outcomes we observed. The generated MQL5, the compiler logs and the compiler binary are not distributed, so a reader can recompute the statistics from our rows but cannot confirm the rows against the artifacts that produced them.",
      "The base model identity at the pinned revision.",
      "The tuned arm's adapter hash, epoch, global step, merge tool, merge settings and merged-weight hash.",
      "The determinism result for the two historical merges.",
      "The decontamination collision counts, which are measured against a corpus that is not distributed.",
      "1.0.0 scope - that the 1.0.0 evaluation set was frozen at seed 42 before any result was seen and was not redrawn. The 1.1.0 holdout sets were selected by a different rule, published with its three digest conventions under holdout_sets; those digests commit to manifests that are not distributed, so the selection behind them is an attestation of the same kind.",
      "Everything stated about how the harness behaves: that the positive control aborts an arm rather than writing rows, that the seed is set per generation, that extraction and scoring ran identically on all three arms, and that the scoring rules published here are the rules the harness applied. The harness is not distributed.",
      "1.0.0 scope - the full-300 comparison. Of the 1.0.0 frozen 300 only the clean-184 rows are published, so the full-300 figures and the change between the two strata cannot be recomputed from this package. That limit is on the 1.0.0 file set alone. For the 1.1.0 holdouts every projected result row is published - 1,200 rows in holdout_ea300_results.jsonl and 800 in holdout_nonea200_results.jsonl - and their arithmetic is checkable, with the prompts and item names withheld as what_is_verifiable_instead below states.",
      "1.0.0 scope - which model produced which arm. The public `arm` is derived from the source directory of the sealed rows; the two local arms' own sealed arm field reads the same value on both and discriminates nothing. We bound the tuned rows to the merged artifact internally by byte-identity to the run directory that served it, but a reader holding only this release cannot check that attribution. The same holds in 1.1.0 and for the same reason: PROJECTION_REPORT.json arm_derivation_v1_1 records that the published arm comes from the source extract, while the sealed rows' own arm field reads the same value, B, on every local arm of this release, and C on the frontier arm."
    ],
    "what_is_verifiable_instead": "From the 1.0.0 file set: everything computed from the published rows - both prompt hashes, the pairing, all three rates, the differences, the confidence intervals, both McNemar results and both paired tables. From the files 1.1.0 adds: in holdout_ea300_results.jsonl, holdout_nonea200_results.jsonl and bridge_q8_184_results.jsonl a reader can check the row counts, the arm ids, the headline and truncation counts per arm and the item pairing across arms; the holdout prompt hashes cannot be recomputed and no holdout row can be regenerated, because the prompts are not published. Across both, a reader can run the 184 published prompts against any model and compare, which is the comparison the release exists to enable."
  },
  "licence_scope": {
    "licensed_under_the_benchmark_evaluation_licence": [
      "prompts.jsonl",
      "per_item_results.jsonl",
      "serving_template.json",
      "holdout_ea300_results.jsonl",
      "holdout_nonea200_results.jsonl",
      "bridge_q8_184_results.jsonl",
      "corpus_row_hashes.json"
    ],
    "supporting_files": [
      "README.md",
      "LICENSE",
      "CITATION.cff",
      "SHA256SUMS.txt",
      ".gitattributes",
      "publication_metadata.json",
      "PROJECTION_REPORT.json",
      "verify_public_release.py"
    ],
    "supporting_files_terms": "The supporting files describe and verify the release. They may be reproduced and quoted for the purpose of using, checking or citing this benchmark, including in published results that disagree with ours, and verify_public_release.py may be run as-is to verify a copy of this release. They carry no grant over the benchmark data itself, which is covered by the licence above.",
    "note": "Being listed as published means a file ships in the release. It does not by itself say which terms cover it; this block does."
  },
  "error_counts": {
    "errors": "The number of individual error diagnostics printed in the compile log, counted line by line.",
    "result_line_errors": "The compiler's own final tally, read from its `Result: N errors, M warnings` line.",
    "why_they_differ": "The compiler stops printing individual diagnostics past a limit while its final tally keeps counting. A log can list 99 errors by line and end with `Result: 103 errors`. Both numbers are recorded and neither is corrected against the other.",
    "measured_disagreement": "They disagree on 12 of the 368 local-arm rows of per_item_results.jsonl and on 14 of the 552 local-arm rows of bridge_q8_184_results.jsonl - 26 of the 920 local-arm rows - all in the base arm, all at 99 or 100 printed diagnostics. Both counts are published on 1,104 rows, not 920: those 920 local-arm rows and the 184 frontier rows of per_item_results.jsonl, which carry both counts and disagree on none, in errors or in warnings. The two holdout row files publish `errors` and no `result_line_errors`, so no disagreement is computable on them.",
    "effect_on_the_verdict": "None. A verdict turns on whether there were any errors at all, and both counts are far above zero on every row where they differ."
  },
  "bootstrap_caveat": "The bootstrap interval is reported for completeness and is not independently reproducible from this package: the resampling implementation, its RNG and the percentile convention are not published, and a seed alone does not determine identical endpoints across implementations. The Wilson and Newcombe intervals, the McNemar result and both rates do recompute from the published rows, and those are the intervals the headline rests on.",
  "scoring": {
    "verdicts": {
      "verdict_headline": "TRUE when the compile log reports zero errors AND the .ex5 artifact was produced. This is the headline metric.",
      "verdict_strict": "TRUE when the compile log reports zero errors AND zero warnings AND the .ex5 artifact was produced.",
      "note": "Both are computed for every item that reached compilation, and published side by side. Neither is derived from the other. Where generation failed outright nothing reached the compiler and both fields are null: three rows, all in holdout_ea300_results.jsonl, all on the frontier arm, each with bucket INFRASTRUCTURE-FAILURE and finish_reason `error`. Those are the only null verdicts in the release. A null is not a failure and must not be read as one - it is an item on which nothing about the model was measured, which is why INFRASTRUCTURE-FAILURE is excluded from the denominator above."
    },
    "buckets": {
      "values": [
        "compile-pass",
        "compile-fail",
        "TRUNCATED",
        "INFRASTRUCTURE-FAILURE"
      ],
      "precedence": "Applied in this order, and the order is load-bearing: a generation error gives INFRASTRUCTURE-FAILURE; else truncation gives TRUNCATED; else a compile-side infrastructure reason gives INFRASTRUCTURE-FAILURE; else compile-pass if verdict_headline is true, otherwise compile-fail.",
      "why_that_order": "Generation errors, truncations, and compile infrastructure failures are excluded before assigning compile pass or fail. Each of the three is the harness's failure or the platform's rather than the model's, so only once all three are out does a pass or a fail describe the model.",
      "excluded_from_the_denominator": [
        "TRUNCATED",
        "INFRASTRUCTURE-FAILURE"
      ]
    },
    "extraction": {
      "version": "three_way.v4",
      "step_1": "Truncate the response at the first serving end-token found, if any: <|end|>, <|endoftext|>, </s>, <|im_end|>.",
      "step_2": "Find every fenced code block, accepting both CommonMark fence characters (``` and ~~~). An unterminated final fence still counts as a block, because that is what a truncated generation looks like and dropping it would hide the truncation.",
      "step_3": "If there are no fenced blocks, use the whole response, stripped.",
      "step_4": "Otherwise take the first block that defines a real MQL5 event handler - a recognised handler name with a matching return type and parameter signature. If no block defines one, take the first block. The chosen index is published per row as used_block_index, beside n_fenced_blocks.",
      "why_the_entry_point_rule": "First-wins handed the compiler whatever a response opened with. Measured against the live compiler on an identical EA body: a leading plan block scored 2 errors and a leading block of helper includes scored 1, each published as the model's own compile failure, while the same EA sent alone passed. The cost is asymmetric - only arms that write prose and plans pay it - so the rule exists to stop the extractor manufacturing failures.",
      "applied_identically_to_every_arm": "The same extraction runs on every arm. Stripping fences for some arms and not others is the bias this rule exists to prevent."
    },
    "compile_log_reading": {
      "encoding": "The compiler writes its logs as UTF-16LE with a byte-order mark. They must be decoded before matching. Matching the raw bytes returns zero hits and exits cleanly, which silently collapses the strict metric onto the headline metric with nothing in the output saying so.",
      "diagnostic_rule": "A diagnostic is a printed line matching `(line,col): error N:` or `(line,col): warning N:`, not the bare substring 'error' anywhere in the log.",
      "completion_rule": "A log without the compiler's own `Result: N errors, M warnings` line is an INFRASTRUCTURE-FAILURE, never a model result. That line is the compiler's statement that a compilation ran to completion; a partial log carries zero printed diagnostics and would otherwise read as a clean compile."
    },
    "what_is_not_published": "The harness, the extractor and the scorer are not distributed, and neither is the compiler. These rules describe what they do at the level published here; load-bearing details below that level are not published, so an independent implementation may score individual items differently while following every published rule. A reader also cannot confirm that our published rows were produced by an implementation that follows these rules. Two of these instruments have no retained known-answer validation record: the compile verdict, whose two control fixtures and compiler output were not kept, and the extractor, whose rule is stated above and whose source is retained but for which no demonstration that it selects the right block on a known input and rejects a falsifying one was kept. Every verdict in the release rests on both, and on the holdout rows used_block_index is not published either, so the extractor's choice is not row-auditable there."
  },
  "interval_conventions": {
    "z": 1.959963984540054,
    "z_1_0_0": 1.96,
    "z_note": "Two conventions are in force, one per release, because every 1.0.0 figure is carried forward exactly as 1.0.0 published it rather than recomputed. Every interval in statistics_v1_1 uses z = 1.959963984540054, the two-sided 95% normal quantile at full precision, which is the value the code that produced that block used: all 170 intervals in it reproduce with that z - 108 Wilson intervals and 62 Newcombe difference intervals - and none reproduce with the rounded value. The fields carried forward from 1.0.0 - the headline rates, the headline difference and frontier_result - were computed with the rounded z = 1.96 published here as z_1_0_0, and reproduce only with it. The two values move these endpoints by up to about two parts in 100,000. Nothing is recomputed by this scoping: it records which z produced which block, so a reader matching an endpoint uses the z that belongs to the field's release.",
    "rate_intervals": "Wilson score interval, applied to each arm's successes over its denominator.",
    "difference_interval": "Newcombe's square-and-add MOVER interval for the difference between two paired proportions.",
    "phi": "Newcombe's phi-hat for the paired 2x2 table, the association term the MOVER interval needs. It is published because the interval cannot be recomputed without it. With cells e (both pass), f (first arm only), g (second arm only), h (both fail): A = e*h - f*g and D = sqrt((e+f)(g+h)(e+g)(f+h)). The rule, as computed: phi is A/D, and 0 when D is zero because a marginal is zero. No continuity correction is applied.",
    "mcnemar": "Exact binomial test on the discordant pairs."
  },
  "what_is_not_measured": {
    "specification_adherence": "The verdict is a compiler outcome. It does not establish that the generated program implements the specification, or any part of it. A minimal Expert Advisor that compiles and does nothing scores exactly as a correct one does. This is a stronger limitation than 'compiling is not trading correctly': adherence to the prompt is not measured at all.",
    "why_it_cannot_be_ruled_out_here": "Because the generated source is not distributed, this benchmark does not distinguish semantic implementation from compilable boilerplate.",
    "what_would_settle_it": "Structural or functional conformance checks against each specification, or publication of the generated source so others can judge. Neither is in this release."
  },
  "serving_template_confound": {
    "what_was_done": "The two local arms were served with the same prompt template, the one the tuned model was trained on: <|system|>...<|end|>\\n<|user|>...<|end|>\\n<|assistant|>. The published prompt_sha256 is identical for the two local arms on every item, so this is verifiable from the release rather than asserted.",
    "the_confound": "That template is not the base model's native chat format. The base model ships a ChatML chat template using <|im_start|> and <|im_end|>. Serving an instruction model a format it was not tuned for can depress its output quality, so part of the measured difference could in principle reflect template mismatch rather than capability.",
    "why_it_was_done_this_way": "Holding the prompt byte-identical across arms removes prompt wording as a variable. That choice trades one confound for another: identical inputs, but an input that suits one arm better than the other.",
    "corroborating_observation": "The published rows show a clean split in response convention: every base row records one fenced code block and every tuned row records none. That is consistent with two models following different output conventions, and it is what the extraction rule exists to handle fairly.",
    "template_control": "The base model was run twice over the same 184 items with identical generation settings and identical scoring, once under the shared training-format template and once under its own native ChatML template. Among the 183 items with scoreable outputs under both templates, one verdict changed: it compiled under the shared training template and failed under native ChatML. The remaining item was truncated under the shared-template condition and could not be compared. The observed difference on the 183 jointly scoreable items was one compile pass in favour of the shared template. Under the shared template the control reproduced the published base arm's two passing items.",
    "template_control_rows": "The control's per-item rows are not distributed. They are identified by SHA-256: training-format condition c9e0a70a99f90785ed4e7f9213ad63fecd9ddbf2319c1394885ecc09b82ce866, native ChatML condition 47f8b9eed8c1e0c369af26ca9606f48a72399f9c1cfa52b29918062964c64e33.",
    "template_control_caveats": "The control served a Q8_0 quantisation of the base model on a different inference stack than the published arms, which were served at bf16. The one-item effect is measured in that control configuration; carrying it to the published bf16 arms assumes the template effect transfers across quantisation and serving stack, and that transfer was not measured. Neither control condition is a published arm and neither replaces one. The shared-template Q8 control produced the same two passing item identities as the published base arm among the 183 jointly scoreable items. One item was truncated, so full 184-item replication is not claimed.",
    "status": "Measured in a control configuration: one changed verdict among 183 jointly scoreable items, in the shared template's favour. Transfer to the published bf16 arms was not measured. The headline remains a difference between the two local arms as served, under one shared template."
  },
  "licence": {
    "title": "CompilingThings Benchmark Evaluation Licence v1.0",
    "notice": "Copyright \u00a9 2026 CompilingThings. All rights reserved.\n\nThe grant covers the published benchmark files: prompts.jsonl, per_item_results.jsonl, serving_template.json, holdout_ea300_results.jsonl, holdout_nonea200_results.jsonl, bridge_q8_184_results.jsonl and corpus_row_hashes.json.\n\nYou may: run the 184 published prompts against any model; check, cite and compare the published result rows and the corpus row-hash manifest, including the holdout row files, whose prompts are not published; implement and use the published scoring contract; publish and compare benchmark results, including results that disagree with ours; cite the benchmark and the release identifier.\n\nYou may not: use the prompts for training, fine-tuning, continued pretraining, distillation, or reinforcement learning; incorporate the prompts or result rows into another dataset; create or sell derivative datasets; claim any right to the private training corpus; represent the benchmark as your own work; treat this grant as extending to future releases.\n\nNo other rights are granted. Nothing here grants any right to the private training corpus, the generated MQL5 outputs, the model weights, the production tooling, or any unreleased material. This grant does not extend to future releases or later versions of this benchmark. Commercial licences, evaluation access, research collaboration and partnership are arranged individually under written agreement. Enquiries are welcome via a discussion on the Hugging Face repository.",
    "permitted": [
      "run the 184 published prompts against any model",
      "check, cite and compare the published result rows and the corpus row-hash manifest, including the holdout row files, whose prompts are not published",
      "implement and use the published scoring contract",
      "publish and compare benchmark results, including results that disagree with ours",
      "cite the benchmark and the release identifier"
    ],
    "prohibited": [
      "use the prompts for training, fine-tuning, continued pretraining, distillation, or reinforcement learning",
      "incorporate the prompts or result rows into another dataset",
      "create or sell derivative datasets",
      "claim any right to the private training corpus",
      "represent the benchmark as your own work",
      "treat this grant as extending to future releases"
    ],
    "hugging_face_fields": {
      "license": "other",
      "license_name": "compilingthings-benchmark-evaluation-licence-v1.0",
      "license_link": "LICENSE"
    },
    "published": [
      "this dataset card (README.md)",
      "the 184 benchmark prompts (prompts.jsonl)",
      "the serving template and system prompt (serving_template.json)",
      "sanitised per-item metadata for the three 1.0.0 arms (per_item_results.jsonl)",
      "hash-keyed per-item results on the 300 EA holdout (holdout_ea300_results.jsonl)",
      "hash-keyed per-item results on the 200 non-EA holdout (holdout_nonea200_results.jsonl)",
      "the Q8_0 bridge rows over the 184 public prompts (bridge_q8_184_results.jsonl)",
      "the corpus row-hash manifest (corpus_row_hashes.json)",
      "the projection report (PROJECTION_REPORT.json)",
      "the release metadata, including all aggregate results (publication_metadata.json)",
      "the verification script (verify_public_release.py)",
      "the checksum manifest (SHA256SUMS.txt)",
      "the citation metadata (CITATION.cff)",
      "the line-ending pin that keeps the checksums valid on checkout (.gitattributes)",
      "the licence (LICENSE)"
    ],
    "not_published": [
      "the training corpus",
      "generated MQL5 outputs, all arms",
      "model weights",
      "production tooling",
      "the item-id-to-name mapping"
    ],
    "note": "The grant is deliberately narrow and covers only the published benchmark files. Everything outside that list is reserved, and nothing here extends to a later release.",
    "platform_terms_caveat": "Hosting-platform terms operate separately from this notice. They have not been read by the authors of this metadata and are deliberately not characterised here; only artifacts intended to be public are uploaded anywhere."
  },
  "training_corpus": {
    "rows": 83155,
    "sha256": "5e9881b61f3375d3d575c35950a236375eb5648f5fe55b8d18de321f69659c68",
    "row_schema": [
      "prompt",
      "completion"
    ],
    "measured": "row count and hash computed directly from the corpus file",
    "distributed": false,
    "why_withheld": "The corpus is not distributed with this release. Claims that depend on it are attested, not independently reproducible from the public package."
  },
  "evaluation": {
    "source_pool_records": 9168,
    "source_pool_sha256": "2ca2d669626a44b8a0096ce25873e23d6cbf08f6887295a26179218c7e0d29cf",
    "frozen_n": 300,
    "frozen_seed": 42,
    "selection_sha256": "8684badae73db0b8c69eb2f8c3d62e6b54aed64875430498485d178af3b33ff4",
    "manifest_sha256": "2accc593805d0d480e5bd6777971bf37c540a6b2bc2527283b06efc1554bf1d5",
    "selection_sha256_convention": "SHA-256 of the UTF-8 bytes of json.dumps(sorted(selected_ea_names), separators=(\",\", \":\"), ensure_ascii=True) - the item names alone, sorted, compact-separated and ASCII-escaped, with no path, no seed and no counts, so it is a set identity that is identical across machines. It is not the sha256 of the manifest file.",
    "manifest_sha256_convention": "SHA-256 of the UTF-8 bytes of json.dumps(the whole manifest object with its own manifest_sha256 key removed, sort_keys=True, separators=(\",\", \":\"), ensure_ascii=True), with no trailing newline. It covers the provenance fields as well as the names, and it is not the sha256 of the manifest file on disk, which is a third value.",
    "selection_rule": "score(ea_name) = sha256(utf8(str(seed)) + b\"\\x00\" + utf8(ea_name)); order by (lowercase hex of score ASC, ea_name ASC); take first n; manifest lists them sorted by ea_name. seed is encoded as its decimal characters (--seed 42 hashes b\"42\"). No RNG.",
    "selection_is_deterministic": "No RNG. The manifest states the full rule."
  },
  "decontamination": {
    "against": "the training corpus named above, by SHA-256",
    "methods": [
      {
        "name": "exact",
        "definition": "SHA-256 equality over the UTF-8 encoding of the evaluation item's reference solution and each training row's completion field",
        "collisions": 0
      },
      {
        "name": "normalised-equivalence",
        "definition": "SHA-256 equality after a domain normaliser is applied to both sides. Compares an evaluation item's reference solution against every training row's completion.",
        "normaliser_source_sha256": "789bd40d55a5c38ead9c0951fbe38bf1a4aa0b06bcb2e9fdc6a17907d2486794",
        "normaliser_source_hash_scope": "SHA-256 of the _normalize_code function source after universal-newline normalisation to LF, from the def _normalize_code line through the line immediately before the next top-level def, with no trailing newline. It is not a hash of the whole source file, and it is not a hash of the raw on-disk bytes, which carry CRLF line endings.",
        "normaliser_substitution_count": 13,
        "normaliser_substitution_count_meaning": "the number of regex substitution rules in the normaliser's source at the hash above, counted from that source",
        "normalisation_steps": [
          "strip line and block comments",
          "canonicalise EA name string literals",
          "canonicalise magic numbers",
          "canonicalise timeframe constants",
          "canonicalise applied-price enums",
          "canonicalise indicator period arguments",
          "canonicalise risk, stop-loss, take-profit and trailing multipliers",
          "canonicalise instrument symbol literals",
          "collapse all whitespace runs to a single space"
        ],
        "steps_are_bound_to_the_hash_above": "The step list describes the normaliser at that source hash. If the hash moves, treat the list as unverified until it is re-read.",
        "direction_of_error": "The normaliser is deliberately aggressive: it collapses parameter variation, so two items that differ only in their parameters normalise to the same string. Removing these items therefore removes more than a strict reading requires. The clean stratum is the conservative one.",
        "collisions": 116
      }
    ],
    "frozen_n": 300,
    "clean_n": 184,
    "colliding_n": 116,
    "no_similarity_cutoff": "No similarity threshold decides membership. The two methods above are the whole predicate.",
    "which_side_is_scored": "Both methods above are scored on the reference-solution side: an evaluation item's reference solution against every training row's completion. The specification side - the text actually put to the models - was scanned as well, and only its exact-SHA result is scored, at 0 collisions. The normalised-exact method cannot be scored there, and this is measured rather than assumed: these specifications are entirely `//` comment lines and the normaliser strips comments as its first step, so every specification normalises to the empty string and collides with every other by construction - 300 of 300 rows empty on the frozen 300, one distinct normalised hash, and the same at 184 of 184 on the public 184 against the 220k corpus. The retained scan records the resulting count as an instrument artefact, escalates it rather than scoring it as a stop, and records that which side the stop conditions bind on is unsettled and was not settled there. So the specification-side evidence in this release is the exact-SHA result and nothing else, and the 116 and the 0 in the table above are reference-solution-side figures.",
    "reproduction_gate_passed": true,
    "reproduction_gate": "What that boolean records. On 2026-08-31 the 2026-08-19 normalised-exact scan of the frozen 300 against the 83k corpus was re-run by a separate tool, on an archived copy of the same training file - same sha256, 5e9881b61f3375d3d575c35950a236375eb5648f5fe55b8d18de321f69659c68 - and the re-run was gated on reproducing the audit's published figures: 116 normalised-exact collisions against 116 published, 0 exact against 0 published, and all 30 preserved witnesses present with their normaliser hashes reproduced, 30 of 30. All four conditions held, which is the `true` above. The gate record is retained privately and identified by SHA-256, 0359e3522235bc364daaedf27e540eb98b2428848500884379b1dc74cd098ff4. This does not make the scan reproducible by a reader: neither corpus is distributed, and the archived copy that re-run read no longer exists either, so 2026-08-31 was the last occasion on which it could be run at all.",
    "known_answer_validation": "The scanner's own detection test, and where its record is. Three known-positive fixtures and one falsifying twin were put through the same matcher at the same settings as the scans - threshold 0.7, 256 permutations, shingle k 5, seed 20260812, over a 400-row training sample and 5 evaluation items - and all four results are retained in the evidence package and not distributed: frozen300_selftest_P1_known_positive_verbatim_20260819.json, sha256 1bb4d66d3b708732d763f5b93a94dafcfa76ef2581fd321a4b678e3c888bdfb3, a verbatim duplicate, 5 of 5 exact and 5 of 5 normalised; frozen300_selftest_P2_known_positive_cosmetic_20260819.json, sha256 e8b3327ddb0c55c66c1e41c3b63e3c5afadda53ae26abed0668e4b5f28253933, a cosmetic variant, 0 exact and 5 of 5 normalised; frozen300_selftest_P3_known_positive_crlf_20260819.json, sha256 87666831547d8d8b2d19e6806bdb441596335c84b5fd40f800b55ad7bac036c5, a line-ending variant, 0 exact and 5 of 5 normalised; and frozen300_selftest_TWIN_falsifying_20260819.json, sha256 f075afd1f628a3de8945a030f464967a1700f63f5384fa989f871de2108e9eed, the falsifying twin, 0 exact and 0 normalised. The exact method therefore detects only the verbatim positive, the normalised method detects all three positives, and both reject the twin. Each hash here is a line of the evidence package's public-half manifest, whose own sha256 is published at the top of this document as evidence_manifest_v1_1_0_public_half_sha256; a reader without the package cannot check them.",
    "known_answer_validation_scope": "Which scan each set of fixture records belongs to. Those four are the 2026-08-19 scan's. The two later scan runs each kept four of their own, at the same settings and with the same four results, and both sets are retained in the evidence package and not distributed. From the 2026-09-10 run, which covered the public 184 and the 1,000-item EA pool together: c184hb_selftest_P1_known_positive_verbatim_20260910T182010Z.json, sha256 280ae2c546a977e83528bbb0b5c8bc42442d90fd955921aebed375c294c8b883; c184hb_selftest_P2_known_positive_cosmetic_20260910T182010Z.json, sha256 061252cfe079b40e690d4feb903ce37131039c8b03c335fec32e7350ed2abe1f; c184hb_selftest_P3_known_positive_crlf_20260910T182010Z.json, sha256 602e61be3e7a1cc918725f09c15c0567ee155e1ec5f18a842a840f424dc924e9; and c184hb_selftest_TWIN_falsifying_unrelated_20260910T182010Z.json, sha256 112ede73d8dd62357f4a7f6612fe0f37999638914f2d457f6fa71d76f1c79d77. From the 2026-09-11 scan of the 600-item non-EA pool: holdoutC_selftest_P1_known_positive_verbatim_20260911T074334Z.json, sha256 afca32d2ea0817e17f261023a06ca623c223a6ade9872f6cd98da0f81f5e64ec; holdoutC_selftest_P2_known_positive_cosmetic_20260911T074334Z.json, sha256 e797de3fb0aefa23fcaeea9fe4605754279f03f33667945c195122cf21e40025; holdoutC_selftest_P3_known_positive_crlf_20260911T074334Z.json, sha256 65521e5b647a6db58e91406898e8d34a634d267574a11c99e986df256ddac583; and holdoutC_selftest_TWIN_falsifying_unrelated_20260911T074334Z.json, sha256 3495cd7f6d10def6b8284d2ebe58650560fd6e67ad4772c7f3b5c5091f7fcd20. Each of those eight reads 5 of 5 exact and 5 of 5 normalised on the verbatim positive, 0 exact and 5 of 5 normalised on the cosmetic and the line-ending positive, and 0 and 0 on the twin. Two things differ from the 2026-08-19 four: the training sample behind them is 5 rows rather than 400, and each of the two runs also carried three known positives and a falsifying twin through the same streaming pass as the scan itself, against all 215,541 corpus rows, recorded in the scan record as matcher_validation_corpus_canaries and passing on every check. The public-half extract of the 2026-09-10 result, DERIVED_corpus220k_vs_clean184_ONLY_20260910T182010Z.json, sha256 79fe033926ad91929fdd3144dca99033e478f4e742de0bd9cad6993fe699443d, carries that validation as two boolean fields of its own, matcher_validated_fixtures and matcher_validated_corpus_canaries, both true, and names the four fixture records it summarises. A fixture pass shows the matcher catches a known positive and rejects a falsifying twin on a small sample; it does not establish its behaviour at corpus scale, and the witness cap below is a separate and known defect.",
    "holdout_pool_scans": "Where the two holdout-pool figures come from. The 1,000-item EA pool and the public 184 were scanned against the 220k corpus in one run on 2026-09-10, retained as corpus220k_vs_clean184_holdoutB_20260910T182010Z.json, sha256 6301ae1f8298ffd96358e03f50dacfac57c9ad3211d4d628039e44ca959523b2: on the EA pool, 0 of 1,000 exact and 352 of 1,000 normalised-exact; on the public 184, 0 exact and 1 of 184 normalised-exact. The 600-item non-EA pool, 150 items in each of its four classes, was scanned on 2026-09-11, retained as corpus220k_vs_holdoutC_20260911T074334Z.json, sha256 bbb28f9bf83fab245d476daf1f745b2159fee66e2af3dfa38bfacd534475b582: 0 of 600 exact and 21 of 600 normalised-exact, the 21 being 20 script items and 1 service item, with none in the include and indicator classes. Those two records are where the 352, the 21 and the split of the 21 across classes are measured. Both also record the specification side, at 0 exact and at 1,000 of 1,000 and 600 of 600 normalised, and neither normalised figure is scored, for the reason under which_side_is_scored above. Neither record removes anything itself: each counts its collisions and halts on them, and the removals in the release card's decontamination table are the two draws', not the scanner's. Both files are retained in the evidence package and are not distributed, and their names carry this project's own working labels for the two pools - holdoutB is the 1,000-item EA pool the 300-item EA holdout is drawn from, holdoutC the 600-item non-EA pool behind the 200-item non-EA holdout - so a counterparty asking for one by name is asking for the file named here. Each hash is a line of the full-package evidence manifest whose own sha256 is published at the top of this document as evidence_manifest_v1_1_0_sha256. That manifest is not distributed either, so these hashes cannot be checked against anything a reader holds; they are what a received file can be checked against.",
    "witnesses_preserved": 30,
    "witness_cap_is_a_known_defect": {
      "preserved": 30,
      "collisions_found": 116,
      "statement": "The scanner caps preserved witnesses. Every collision was counted; only a subset was preserved as a witness. The count is complete, the witness set is not."
    }
  },
  "models": {
    "base": {
      "repo_id": "Qwen/Qwen2.5-Coder-14B-Instruct",
      "pinned_revision": "aedcc2d42b622764e023cf882b6652e646b95671",
      "identity_verdict": "PROVEN_AGAINST_HUGGING_FACE_AT_PINNED_REVISION",
      "weight_dtype": "bf16",
      "quantisation_method": "none - weights are stored at bf16 precision, not quantised",
      "serving_quantisation": "None. The model is loaded with torch_dtype=bfloat16 and no quantisation configuration on this path. The rows' `quantisation` field records the string bf16, which names the dtype, not a quantisation method."
    },
    "tuned": {
      "description": "The base above, fine-tuned on the withheld corpus, then merged to bf16 and served directly. No adapter was applied at serving time.",
      "weights_distributed": false,
      "identity_note": "The tuned weights are not published. The per-item rows commit to its outputs by hash; they are not a route to the model.",
      "weight_dtype": "bf16",
      "quantisation_method": "none - weights are stored at bf16 precision, not quantised",
      "adapter": {
        "adapter_model_sha256": "9cc2d9bc1b75dabca474cb49de2c6aa340ee1185ebf463322cc3341ff0769c3d",
        "adapter_config_sha256": "f07d8621efe63994172cc86150fcfcae8bca4959899a90e505258c1334e7d8ce",
        "epoch": 1,
        "epoch_verdict": "PROVEN",
        "epoch_status": "MEASURED",
        "epoch_evidence": [
          {
            "requirement": "trainer_state.json epoch == N.0",
            "literal_value": "\"epoch\": 1.0"
          },
          {
            "requirement": "trainer_state.json global_step == max_steps",
            "literal_value": "\"global_step\": 4694, \"max_steps\": 4694"
          },
          {
            "requirement": "train_config num_train_epochs consistent with N",
            "literal_value": "train_config.json \"epochs\": 1 ; trainer_state.json \"num_train_epochs\": 1"
          }
        ],
        "directory_name_disclaimer": "The directory name checkpoint-4694 was not used as evidence. The three literals above were read from the file contents.",
        "adapter_selection_note": "A second adapter from a later epoch exists and is the same size. The two are told apart by hash only; a check on the directory name would decide nothing.",
        "lora": {
          "r": 64,
          "lora_alpha": 128,
          "lora_dropout": 0.05,
          "target_modules": [
            "down_proj",
            "gate_proj",
            "k_proj",
            "o_proj",
            "q_proj",
            "up_proj",
            "v_proj"
          ],
          "peft_version_that_wrote_the_adapter": "0.18.1"
        }
      },
      "merge": {
        "tool": "merge_adapter.py",
        "tool_sha256": "f426179b80cce7c1c9f63fbdc2f058e95562bad9a24e6d57fb4ded07d5ede473",
        "tool_sha256_source": "re-derived by this generator from the copy sealed in the evidence package, then checked against the value the merge record wrote",
        "dtype": "bfloat16",
        "device": "cpu",
        "base_identity_asserted_against": "Qwen/Qwen2.5-Coder-14B-Instruct",
        "argv_note": "The verbatim argv is four absolute local paths plus the settings above. The paths are withheld; the settings are what determine the output.",
        "started_utc": "2026-08-23T03:00:55Z",
        "finished_utc": "2026-08-23T03:06:35Z",
        "duration_seconds": 339.795
      },
      "merged_weights": {
        "model_safetensors_sha256": "1b8ee53b697e9a600834871a595924b183f8229d6cb8850cd6ba63075ac4c66b",
        "model_safetensors_bytes": 29540134824,
        "dtype": "bfloat16"
      },
      "determinism": {
        "verdict": "MATCH",
        "independent_methods": [
          "tools/merge_adapter.py internal hashing",
          "GNU coreutils sha256sum, run separately after the merge"
        ],
        "environment": "The 2026-08-20 and 2026-08-23 merges ran on the same interpreter and package set: Python 3.14.0, torch 2.13.0+cpu, transformers 5.15.0, peft 0.20.0, accelerate 1.14.0, safetensors 0.8.0. Byte-identity is therefore evidence of determinism within that environment, not across environments.",
        "what_it_establishes": "It establishes that the same base + same adapter + same tool + same package versions + same device and dtype reproduce the same bytes. It does not establish that the artifact is correct for evaluation, and it produces no score.",
        "limitation": "Two merges run in 2026 from the same inputs in the same environment produced byte-identical weights. Neither was run at release time, and neither was run in a different environment. Cross-environment reproduction is not established."
      },
      "quantisation_parity_vs_base_arm": {
        "verdict": "MATCH",
        "target": "bf16 (bfloat16)",
        "comparison_arm_quantisation": "bf16"
      },
      "served_artifact_binding": "The sealed tuned rows were bound to this merged artifact three ways: the sealed results file is byte-identical to the run directory that served it; that run's parameters name the merged model and its fingerprint, which the run's own console log prints; and the only other tuned run candidate produced no results file at all."
    },
    "frontier": {
      "model_id": "gpt-5.6-sol",
      "provider": "openai",
      "access": "vendor chat-completions API; weights are vendor-hosted and not independently pinned by this release",
      "identity_basis": "model identifier and evaluation date",
      "evaluated_on": "2026-09-02",
      "temperature": 1.0,
      "temperature_note": "the run used temperature 1. This arm is a single run of a sampled decoder.",
      "seed": 42,
      "seed_note": "sent via the API's seed parameter; determinism is not guaranteed by the vendor",
      "applied_settings_note": "What temperature_applied and seed_applied record on this arm. All 184 frontier rows of per_item_results.jsonl carry APPLIED in both fields, and the value is the harness's own: this arm's retained run parameters, at run_params_sha256 below, record temperature_applied APPLIED and seed_applied APPLIED, and the rows carry that forward. The harness derives both from the backend and provider recorded for the arm and from nothing else; it reads no part of the service's response. It is not vendor confirmation, and the retained rows hold none: the only field they carry back from the service is the token usage count. So APPLIED here records how the harness classified its own request, not what the service did with it, and whether the service applied either setting is not recorded - which is what seed_note above and the release card both say. The 1.1.0 local rows record UNKNOWN in these two fields, as serving_v1_1.applied_settings_note explains; the difference between APPLIED here and UNKNOWN there is a difference between two harness classifications and is not evidence about the vendor.",
      "max_completion_tokens": 8192,
      "generation_cap_note": "The retained run record stores the 8,192-token generation cap under the harness's generic max_new_tokens field; max_completion_tokens is the API parameter it was sent as. max_new_tokens is not the literal API parameter used for this arm.",
      "results_rows_sha256": "354f695a34a429d9676279c2a39fbe6c47bd3280054db4c643bac51374fbb4f0",
      "run_params_sha256": "1914a9aef38f22cf40b6fead17eae077f584a4a471a5be9b573eaa6b5cfde00b",
      "run_params_n_note": "run_params.json records n=300, while the retained frontier results contain 184 rows matching the published clean-184 benchmark. The stale n field is not used in any reported result.",
      "served_format": "system and user messages as separate API fields; the local serving template cannot be submitted to the API. The frontier rows' prompt_sha256 hashes the canonical system+user rendering, as defined in PROJECTION_REPORT.json."
    },
    "arm_field_warning": "The sealed per-item rows of the two local arms carry an internal `arm` field that reads \"A\" on both. It discriminates nothing and is withheld from the public rows, which carry an arm derived from the source directory instead. The frontier arm's sealed rows read \"C\"."
  },
  "headline": {
    "stratum": "CLEAN 184 \u2014 no normalised-exact training twin",
    "n": 184,
    "base": {
      "successes": 2,
      "total": 184,
      "rate": 0.010869565217391304,
      "rate_pct": "1.09%",
      "ci95": [
        0.0029858430553776835,
        0.0387599683090911
      ],
      "ci95_pct": [
        "0.30%",
        "3.88%"
      ],
      "ci_method": "wilson-score-95"
    },
    "tuned": {
      "successes": 170,
      "total": 184,
      "rate": 0.9239130434782609,
      "rate_pct": "92.39%",
      "ci95": [
        0.8763482447676944,
        0.9541387187164326
      ],
      "ci95_pct": [
        "87.63%",
        "95.41%"
      ],
      "ci_method": "wilson-score-95"
    },
    "difference": {
      "diff": 0.9130434782608695,
      "numerator": "(169 - 1)/184",
      "ci95": [
        0.8540098171850538,
        0.9455338107113938
      ],
      "phi": -0.16760440475893776,
      "method": "newcombe-1998-square-and-add-mover-paired",
      "diff_pp": "91.30 pp",
      "ci95_pp": [
        "85.40 pp",
        "94.55 pp"
      ]
    },
    "mcnemar": {
      "b": 1,
      "c": 169,
      "n_discordant": 170,
      "p_value": 2.2852138271288273e-49,
      "method": "mcnemar-exact-binomial"
    },
    "contingency": {
      "a_both_pass": 1,
      "b_base_only": 1,
      "c_tuned_only": 169,
      "d_neither": 13,
      "n": 184
    },
    "bootstrap": {
      "method": "paired-item-percentile-bootstrap",
      "seed": 20260824,
      "resamples": 20000,
      "n_items": 184,
      "point_estimate": 0.9130434782608695,
      "ci95": [
        0.8695652173913043,
        0.9510869565217391
      ]
    },
    "denominator_rule": "The denominator is every item of the stratum that both local arms produced a scoreable row for. Measured on these rows: truncation 0, infrastructure failure 0, missing output 0 in both local arms, so the strict and inclusive denominators are identical and the headline is the same number under either reading."
  },
  "frontier_result": {
    "successes": 179,
    "total": 184,
    "rate": 0.9728260869565217,
    "rate_pct": "97.28%",
    "ci95": [
      0.937973928367375,
      0.9883384539803052
    ],
    "ci95_pct": [
      "93.80%",
      "98.83%"
    ],
    "ci_method": "wilson",
    "buckets": {
      "truncated": 0,
      "infrastructure_failure": 0
    },
    "paired_vs_tuned": {
      "both_pass": 168,
      "tuned_only_pass": 2,
      "frontier_only_pass": 11,
      "both_fail": 3,
      "n_discordant": 13,
      "mcnemar_p": 0.0224609375,
      "method": "mcnemar-exact-binomial",
      "gap_pp": "4.89 pp",
      "reading": "The frontier arm compiled 11 items the tuned arm missed; the tuned arm compiled 2 items the frontier arm missed; 3 failed under both."
    },
    "derived": "every figure in this block is derived from the sealed frontier and tuned per-item rows"
  },
  "full_300_for_comparison": {
    "contingency": {
      "a_both_pass": 1,
      "b_base_only": 1,
      "c_tuned_only": 283,
      "d_neither": 15,
      "n": 300
    },
    "base": {
      "successes": 2,
      "total": 300
    },
    "tuned": {
      "successes": 284,
      "total": 300
    },
    "difference": 0.94,
    "mcnemar_p": 1.8338173912479853e-83,
    "decontamination_cost": -0.027,
    "reading": "Decontamination moved the difference from 0.9400 to 0.9130, a change of -0.0270. The uplift is smaller on the decontaminated stratum and the direction is unchanged."
  },
  "public_release_contents": {
    "per_item_rows": {
      "rows": 552,
      "sha256": "3584aefb4f4226e9ba9ae7453c274eebab6f30b62d0dd89a23f41202c6ec070d",
      "distribution": "PUBLIC"
    },
    "fields_published": [
      "item_id",
      "verdict_headline",
      "verdict_strict",
      "bucket",
      "ex5_exists",
      "errors",
      "warnings",
      "result_line_errors",
      "result_line_warnings",
      "compile_infra_reason",
      "generation_error",
      "positive_control_errors",
      "positive_control_warnings",
      "truncated",
      "finish_reason",
      "n_fenced_blocks",
      "used_block_index",
      "fence_unterminated",
      "used_block_unterminated",
      "extractor_version",
      "harness_version",
      "seed",
      "seed_applied",
      "temperature",
      "temperature_applied",
      "max_new_tokens",
      "quantisation",
      "quantisation_mismatch",
      "prompt_sha256_public",
      "prompt_sha256",
      "mq5_sha256",
      "ex5_sha256"
    ],
    "fields_withheld": [
      "adapter_path",
      "adapter_sha256",
      "base_model_sha256",
      "category",
      "compile_log_path",
      "compile_log_raw_path",
      "compile_log_raw_sha256",
      "compile_log_sha256",
      "compile_stem",
      "ea_name",
      "elapsed_seconds",
      "ex5_path",
      "generated_code_path",
      "generated_code_sha256",
      "manifest_selection_sha256",
      "manifest_sha256",
      "metaeditor_build",
      "metaeditor_path",
      "metaeditor_sha256",
      "metaeditor_size_bytes",
      "model_id",
      "mq5_path",
      "source_sha256",
      "system_prompt_sha256",
      "usage"
    ],
    "item_id_scheme": "Opaque. item_id is a stable public identifier used to join prompts and result rows; it carries none of the item's name.",
    "ea_name_not_in_rows_because": "The rows key on item_id so that the row schema does not depend on the corpus's naming convention. This is a schema choice, not anonymisation: every published prompt carries a `// NAME:` line holding that item's original name, so all 184 names are recoverable from this release.",
    "generated_mql5_source_distributed": false,
    "prompts_distributed": true,
    "prompts": {
      "rows": 184,
      "sha256": "d58391038d39b22d1865e5d874764201f65918fbf92fe837a465f58daee7cfa7",
      "distribution": "PUBLIC"
    },
    "serving_template": {
      "sha256": "3601f3ff49381282ee7e3f83e21db9f9e3891910e4092c9461a2a798c9b14756",
      "distribution": "PUBLIC"
    }
  },
  "not_distributed": {
    "item_id_to_name_mapping": 184,
    "training_corpus_rows": 83155,
    "generated_mql5_source": "all three arms",
    "tuned_model_weights": true,
    "release_under": "written agreement only"
  },
  "contact": {
    "route": "Hugging Face - open a discussion on the dataset repository",
    "for": "evaluation access, research collaboration, commercial licensing, partnership",
    "identity": "CompilingThings"
  },
  "limitations": [
    "The training corpus is not distributed. Any claim about its contents is attested by hash, not independently reproducible from this package.",
    "The tuned model weights are not distributed.",
    "Generated MQL5 source is not distributed for any arm. The mq5_sha256 and ex5_sha256 values are stable commitments to the withheld artifacts: they permit integrity verification if those artifacts are later disclosed under agreement. mq5_sha256 is present on every row, and on a row where the extractor recovered source it commits to that source; on seven rows it is instead the empty-output sentinel defined under public_release_contents_v1_1, which commits to nothing because no source was recovered. ex5_sha256 is present only on rows where an EX5 artifact was produced and is null on the rest, because a failed compile leaves nothing to hash. They do not let a reader check independently generated output against ours, because two models almost never emit byte-identical source.",
    "The result rows carry an opaque item_id rather than the original item name, but the names are not hidden: every published prompt contains a `// NAME:` line, because that line was in the prompt the models were given. The naming scheme for these 184 items is therefore public. An earlier draft of this card claimed otherwise and was wrong.",
    "The 184 prompts are now public. They should be treated as a public benchmark from this release onward, not as an uncontaminated evaluation set: once published, prompts enter crawls and downstream corpora regardless of licence. A separate private holdout is reserved for future releases.",
    "The compile verdict comes from one compiler on one platform. It measures whether the code compiles, not whether it trades correctly or profitably.",
    "The benchmark is a single frozen draw at seed 42 from a larger pool. It was frozen before any result was seen, and was not redrawn.",
    "The witness set preserved for the normalised-equivalence collisions is capped and therefore incomplete; the collision count itself is complete.",
    "Two historical merges under the same recorded environment produced byte-identical weights: same merge tool, same adapter, same 336 merge targets over 579 base tensors, and an identical model.safetensors hash. A fresh release-time or cross-environment reproduction was not performed.",
    "One base-model authority underlies both the shard hash check and the Hub tree check, so those two are not fully independent of each other."
  ],
  "reproducibility": {
    "what_a_reader_can_verify": [
      "Run the benchmark. The 184 prompts and the serving template are published, so a third party can generate their own arm, compile it, and score it against the same items.",
      "Both published hashes: prompt_sha256_public is sha256 of the prompt as published, and prompt_sha256 is arm-specific - the rendered serving template on local-arm rows, the canonical system+user rendering on frontier rows. verify_public_release.py recomputes both under each arm's rule.",
      "All three rates, both paired tables, both McNemar p-values and every confidence interval except the bootstrap recompute from the published per-item rows alone.",
      "New in 1.1.0: the row counts, arm ids, headline counts, truncation counts and item pairing of holdout_ea300_results.jsonl, holdout_nonea200_results.jsonl and bridge_q8_184_results.jsonl all recompute from those files, and verify_public_release.py checks 9-12 do exactly that. The holdout prompt hashes do not recompute and the holdout rows cannot be regenerated.",
      "The frozen selection rule is fully stated and deterministic, so anyone holding the 1.0.0 evaluation pool at the source hash published under `evaluation` can recompute which of its items were drawn. That is the only draw it holds for: each 1.1.0 holdout was taken as a prefix of its pool's frozen order rather than by that rule, and no source hash is published for either holdout pool, so neither holdout draw can be recomputed from this package."
    ],
    "what_a_reader_cannot_do": [
      "Inspect the generated MQL5 for any arm.",
      "Recompute the decontamination counts, which are measured against a corpus that is not distributed.",
      "Reproduce our tuned arm, whose weights are not distributed. A reader can run their own model against the same items and compare.",
      "Confirm any per-item verdict against the artifact that produced it. The generated MQL5 and the compiler logs are not distributed, so the rows are an attestation and the statistics computed from them are not."
    ],
    "scope_summary": "The public files support new evaluations on the 184 prompts. The corpus, generated source, tuned weights, and original harness are withheld."
  },
  "checksum_manifest": {
    "file": "SHA256SUMS.txt",
    "files_listed": 14,
    "files_present": 15,
    "reconciliation": "every file except SHA256SUMS.txt, which cannot list itself"
  },
  "supersedes_release_id": "CompilingThings/compile-benchmark-v1.0.0",
  "release_v1_1_note": "Version 1.1.0 adds two private holdout sets (300 EA, 200 non-EA), a Q8_0 bridge over the 184 public prompts, and a row-content hash manifest of the 220k corpus. Every 1.0.0 figure, prompt and result row is carried forward unchanged and nothing from 1.0.0 was re-run.",
  "protocol": {
    "title": "MQL5 Dataset Evaluation Protocol \u2014 Pre-Registered, Three-Way Compile Benchmark Against a Frontier Model.md",
    "sha256": "29e7991af7d7a30120c71a329739de3d29bf4cf4d40c3d343f1327bec80f0323",
    "distribution": "retained in the evidence package, not distributed in this release",
    "what_rests_on_it": "Every `pre-registered` reference in this document and in PROJECTION_REPORT.json, the strict-denominator rule under statistics_v1_1.denominator_variants, and the prompt byte-identity requirement reported beside each set's spec_identity. Those citations point here rather than each carrying the identity.",
    "note": "The title is the file's name in the public half of the retained evidence package and the sha256 is that file's, so a counterparty given the package can check the document they receive is the one cited here against the manifest whose own hash is published above. A reader without the package cannot: this is a commitment, like every other withheld-artifact hash in this release."
  },
  "public_set_184_composition": {
    "prompts_total": 184,
    "declare_no_artifact_type_and_carry_a_buy_line": 178,
    "declare_an_artifact_type": {
      "include": 1,
      "indicator": 2,
      "script": 3
    },
    "items_declaring_an_artifact_type": {
      "item_009": "script",
      "item_089": "script",
      "item_093": "indicator",
      "item_102": "script",
      "item_134": "indicator",
      "item_167": "include"
    },
    "base_arm_passing_items": {
      "item_025": "declares no artifact type; carries a `// BUY:` line",
      "item_102": "declares `// TYPE: script`"
    },
    "how_measured": "Read from prompts.jsonl, which ships: an item is counted as declaring a type when its own specification carries a `// TYPE:` header line. No item carries both that header and a `// BUY:` line, and the two counts sum to 184. The passing items are read from per_item_results.jsonl and are the same two in bridge_q8_184_results.jsonl.",
    "system_instruction_served": "All 184 were sent the single Expert Advisor system instruction, in 1.0.0 and again in the Q8_0 bridge. None was sent the include, indicator, script or service instruction, and the six that declare another type were neither re-instructed nor excluded. This is checkable from the release: verify_public_release.py check 6 substitutes each item into the published serving template and reproduces prompt_sha256 for all 184, and the bridge rows' prompt_sha256 equal the 1.0.0 local rows' on all 184 items.",
    "why_it_is_published": "The set is named for its 1.0.0 lineage, and `EA set` is not what its specifications say. It bears on two published facts. One of the exactly two items the base arm compiles here declares `// TYPE: script`, on a set whose base headline is 2 of 184. And the one specification this set shares with the 200-item non-EA holdout, identified by hash under holdout_sets.holdout_nonea200, is item_093, which declares `// TYPE: indicator`: one specification evaluated under two different system instructions, not two separately authored items."
  },
  "holdout_sets": {
    "manifest_digest_conventions": {
      "manifest_file_sha256": "SHA-256 over the bytes of the manifest file itself, the sha256sum convention. This is the value to check a received file against.",
      "manifest_sha256_recorded_by_every_arm": "The manifest's own internal manifest_sha256 field, which every arm's run parameters record: SHA-256 of the UTF-8 bytes of json.dumps(the manifest object with its own manifest_sha256 key removed, sort_keys=True, separators=(\",\", \":\"), ensure_ascii=True), with no trailing newline. It covers the provenance fields as well as the item names, and it is not the file hash above.",
      "manifest_selection_sha256": "The manifest's own internal selection_sha256 field: SHA-256 of the UTF-8 bytes of json.dumps(sorted(selected_ea_names), separators=(\",\", \":\"), ensure_ascii=True) - the item names alone, sorted, with no path, no seed and no counts, so it is a set identity that is identical across machines. The names are sorted for the hash whether or not the manifest lists them sorted; on the 200 non-EA manifest it lists them in class order and the hash is still over the sorted list.",
      "note": "The three are different numbers over different bytes for the same manifest, exactly as model_identities.digest_convention_warning describes for a GGUF. Cite each digest together with its convention."
    },
    "holdout_ea300": {
      "title": "300-item EA holdout: the first 300, in the pool's frozen order, of the 648 items left after removing the 352 of 1,000 with a normalised-exact match to the 220k corpus",
      "n_items": 300,
      "manifest_file_sha256": "75021a9d8db3fc40931de14239e7e279d37967c109dc0de12c4637473022f64a",
      "manifest_sha256_recorded_by_every_arm": "d32a6ca860a6dd7aa0e232396a6b4a1af7cf91140439a4f0d3862cf14f8f89d8",
      "manifest_selection_sha256": "20e326804cc23282aed8a5591448b8ccd6d9c51819d732387ac0f89f55d7494f",
      "prompts_distributed": false,
      "item_names_distributed": false,
      "decontamination_scope": "scanned against the 220k corpus only. No scan of this pool against the 83k corpus exists, so the 83k arm's figures on this set carry no decontamination evidence of their own.",
      "arms": {
        "base": {
          "rows": 300,
          "sealed_results_sha256": "a9aa3ef6748010e2342b14808856fe4c50f8ea3659b9d45d4293dbe0fc7e3871",
          "run_params_sha256": "19fe581c017faad00a27df4169429f8f5025b7264da3da3fabba6600c4c92876"
        },
        "tuned83k": {
          "rows": 300,
          "sealed_results_sha256": "b5b4f863d8286330cb395d9ec699bb596c6d4ee22e26b6d02212b129561238b8",
          "run_params_sha256": "e1b18ce22466c8103a25b96ef402d59493f1dc0101dfaa3f892da0fc610a0185"
        },
        "tuned220k": {
          "rows": 300,
          "sealed_results_sha256": "4b83c2a7f468db83b8544ae0a5cb9d722c952d46790b2b1921cd6b6e8bfd8fe0",
          "run_params_sha256": "8cb402cd797b2ce359f92ffd28dfc912bb0796f47543a0b2ba02e63853371496"
        },
        "frontier": {
          "rows": 300,
          "sealed_results_sha256": "89f614a8b19c9287df52c9ab338dd11260798cd3d825a049333fa9caa7f2e351",
          "run_params_sha256": "b265e7976db2893a5d5f088fea6742db9c84ca83b9967f657d8f942d96fd3e11"
        }
      }
    },
    "holdout_nonea200": {
      "title": "200-item non-EA holdout, 50 each of include file, custom indicator, script and service: the first 50 of each class, in the pool's own order, after removing the 21 of 600 with a normalised-exact match to the 220k corpus",
      "n_items": 200,
      "manifest_file_sha256": "db0e0614eed18143e41931353fd47eb4e86422f673b0b44e23ec2eebff090065",
      "manifest_sha256_recorded_by_every_arm": "59dcb8aebea6049d7fb16681d485995fb7b699c61b13b89a7fe927060a4fe961",
      "manifest_selection_sha256": "655cd809eb68cbd1e3d8d54e5ff883ea525a0f9461728e43e1f5088b89532cc4",
      "prompts_distributed": false,
      "item_names_distributed": false,
      "decontamination_scope": "scanned against the 220k corpus only, on the same terms as the EA holdout.",
      "overlap_with_the_public_184": "One specification appears in both the 184 public set and this set. It is one item of 200 and it is reported, not removed. MEASURED on the published rows: the item_sha256 sets of the two holdouts and the prompt_sha256_public set of the 184 intersect in exactly this one value and nowhere else.",
      "overlap_with_the_public_184_item_sha256": "8d4946c0c860ecdb2d967b590ae14808e5f748273644b741100b73336aad960d",
      "arms": {
        "base": {
          "rows": 200,
          "sealed_results_sha256": "36df1d9a08f30d7b215e07306474ba8322404023788ecab8d6ff51ef0bb06154",
          "run_params_sha256": "eb377ab8361659e6af6bc88d4fe862bfc12a2bf1508717f63c7fcf47b34e9c2c"
        },
        "tuned83k": {
          "rows": 200,
          "sealed_results_sha256": "2570018f7c374bcfd07adc575b7e38389418f53081a4ccd458ac1175c7204007",
          "run_params_sha256": "d008c59615009fd667e070f6e5e1da0169e65b932264d630e54f8b0cbba780ae"
        },
        "tuned220k": {
          "rows": 200,
          "sealed_results_sha256": "8b53cca6a6922c946be720c13e46cb9cd09c36bd1fd8f286c35c4589910ca47d",
          "run_params_sha256": "335922345f30a546f1115541ce380a98bbb108630e0f8eac4b6c8929ffaadaf4"
        },
        "tuned220k_q8_replay": {
          "rows": 200,
          "note": "the stability re-run of the 220k arm on the same 200 items under the same configuration; published as a distinct arm, never merged with or averaged into the first run",
          "public_row_extract_sha256": "010489a93192aa19a17bc56950c14ba829c6e14c2395bd73daad8062c775075d"
        }
      }
    },
    "attestation": "Attested, not independently reproducible from the public package. The holdout prompts and item names are not published, so a third party can check every holdout arithmetic claim against the published rows and can bind a row to a retained artifact by hash, but cannot regenerate a holdout row from this release."
  },
  "model_identities": {
    "digest_convention_warning": "Two different digests are both called \"the sha256 of a GGUF\" in the retained evidence. gguf_q8_0_file_sha256 is sha256 over the whole file, the sha256sum convention. gguf_q8_0_tensor_rollup_sha256 is the whole-file value reported by llama-gguf-hash, which walks tensors and returns a different number for the same file. Neither is wrong; a citation that does not say which one it means is.",
    "row_to_weight_binding": "Every Q8_0 arm in this release records base_model_sha256 as null in its run parameters. The run-time binding between a scored row and the weights that produced it is a recorded model path, not a hash; the GGUF digests below come from the separate merge and quantise logs. This release does not claim that each scored row is hash-bound to its weights.",
    "base": {
      "role": "base Qwen2.5-Coder-14B-Instruct, served at Q8_0",
      "repo_id": "Qwen/Qwen2.5-Coder-14B-Instruct",
      "pinned_revision": "aedcc2d42b622764e023cf882b6652e646b95671",
      "identity_verdict": "PROVEN_AGAINST_HUGGING_FACE_AT_PINNED_REVISION",
      "gguf_bf16_intermediate_sha256": "b3865b9725575f6dfd494c240aa8b414f55b46d72153a0f9b916aec7fc74edc4",
      "gguf_q8_0_file_sha256": "adc9f8c0201a910617c60075e6431eb47f4bff37b3a69b5589cbf1c032a83bc6",
      "gguf_q8_0_tensor_rollup_sha256": "00d4c901753a6524bc8f2d36bea8c98ef84850fb8a5ba22bfbb169f50b7c88ff",
      "weights_distributed": false
    },
    "tuned_83k": {
      "role": "the 1.0.0 fine-tuned model, merged and quantised to Q8_0 for this release",
      "training_corpus_rows": 83155,
      "training_corpus_sha256": "5e9881b61f3375d3d575c35950a236375eb5648f5fe55b8d18de321f69659c68",
      "adapter_model_safetensors_sha256": "9cc2d9bc1b75dabca474cb49de2c6aa340ee1185ebf463322cc3341ff0769c3d",
      "adapter_config_sha256": "f07d8621efe63994172cc86150fcfcae8bca4959899a90e505258c1334e7d8ce",
      "adapter_epoch": 1,
      "adapter_epoch_verdict": "PROVEN",
      "adapter_epoch_evidence": {
        "trainer_state_epoch": "1.0",
        "trainer_state_global_step": 4694,
        "trainer_state_max_steps": 4694,
        "note": "read from the file contents. The checkpoint directory name was not used as evidence."
      },
      "lora": {
        "r": 64,
        "lora_alpha": 128,
        "lora_dropout": 0.05,
        "peft_version_that_wrote_the_adapter": "0.18.1"
      },
      "merged_model_safetensors_sha256": "1b8ee53b697e9a600834871a595924b183f8229d6cb8850cd6ba63075ac4c66b",
      "gguf_bf16_intermediate_sha256": "504789495c2deb382ddae37d5826837baf028e93dc176b0166724284e6a2cf2c",
      "gguf_q8_0_file_sha256": "cc8bf888408c6e5873a530fd933a623515aa0261f666ca05b3e35b752ff1d199",
      "gguf_q8_0_tensor_rollup_sha256": "e4c6aac606cf8336501e159e170487ae74ec1cab1a47887046ad48ff8f09cc3e",
      "weights_distributed": false
    },
    "tuned_220k": {
      "role": "the same base fine-tuned on the 220k corpus, merged and quantised to Q8_0",
      "training_corpus_rows": 215541,
      "training_corpus_sha256": "3c2f57db5986eaac49f7ae5ed6732e1ae702b339b1354bf3d94e33c7b611e023",
      "adapter_model_safetensors_sha256": "4ea3ebaf9ca45f1ed99c22451d6c6fdb14280513fc76a1acff5468ecd805ba75",
      "adapter_config_sha256": "a1a1cefb0d31a3007b2a0b98e777f007b1620610f0f802994ea0cb63b27d97b6",
      "adapter_epoch_claimed": 1,
      "adapter_epoch_verdict": "NOT ESTABLISHED FROM TRAINER STATE",
      "adapter_epoch_note": "The merge record for this adapter states verbatim: \"EPOCH AND STEP ARE NOT ESTABLISHED. This record pins the adapter's bytes; no epoch and no global_step was read, and the directory name is not evidence of either.\" The one-epoch description of this model is the plan it was trained under, not a trainer-state reading. The 83k adapter's epoch is proven from trainer state and that evidence is recorded above.",
      "lora": {
        "r": 64,
        "lora_alpha": 128,
        "lora_dropout": 0.05,
        "peft_version_that_wrote_the_adapter": "0.18.1"
      },
      "merged_model_safetensors_sha256": "a42ea6eb126286544778a805afab06d774619f4915273b83296d8a603fd5d86b",
      "gguf_bf16_intermediate_sha256": "9f6acbcc1dec43bbf2c06d05eb619e1a0e887c8d6cfaf895b3ded36040a02242",
      "gguf_q8_0_file_sha256": "443cb811bc06cf5ed04696233f7cc8fdafb4ad75d18f3f693d8a3848a9be149c",
      "merge_utc": {
        "started": "2026-09-04T07:28:04Z",
        "finished": "2026-09-04T07:30:46Z"
      },
      "convert_utc": {
        "started": "2026-09-04T07:37:07Z",
        "finished": "2026-09-04T07:39:41Z"
      },
      "quantise_utc": {
        "started": "2026-09-04T07:40:28Z",
        "finished": "2026-09-04T07:41:12Z"
      },
      "weights_distributed": false
    },
    "quantisation_parity": {
      "target": "Q8_0 on every local arm",
      "imatrix_used": false,
      "imatrix_rationale": "plain Q8_0 with no calibration data; calibration would make two same-format GGUFs non-comparable",
      "intermediate_outtype": "bf16, lossless from a bfloat16 source",
      "extra_quantize_flags": "none",
      "why_it_matters": "an A/B quantisation mismatch would not degrade the comparison, it would replace the question: B minus A would measure the serving path and read exactly like a fine-tuning result."
    }
  },
  "serving_v1_1": {
    "scope": "The three local arms of 1.1.0 - base, tuned 83k, tuned 220k - on every set they were run on: the 300 EA holdout, the 200 non-EA holdout, the non-EA stability re-run and the Q8_0 bridge over the 184 public prompts.",
    "engine": "llama.cpp llama-server, Q8_0 GGUF, Vulkan backend, 8 parallel slots, 69,632-token context, all layers offloaded",
    "llama_cpp_build": 10441,
    "llama_cpp_commit": "0177dcc73",
    "llama_cpp_build_note": "Clang 20.1.8, Windows x86_64",
    "launch_line": "llama-server.exe -m <gguf> -np 8 -c 69632 -ngl 999 --device Vulkan0 --host 127.0.0.1 --port 8088",
    "operating_system": "Windows 11 Pro 10.0.26200",
    "graphics_driver": "AMD 32.0.31036.15 (2026-08-12)",
    "decoding": "temperature 0.0, top-k 1 and seed 42 sent with every request; greedy",
    "max_new_tokens": 8192,
    "concurrency": 8,
    "decoding_note": "the harness records what it sent. Whether the server applied every setting as sent is not independently verified by this release.",
    "applied_settings_note": "Every 1.1.0 local row records temperature_applied and seed_applied as UNKNOWN: all 552 rows of bridge_q8_184_results.jsonl, on all three arms. The holdout row files publish neither field. UNKNOWN is the harness's third value beside APPLIED and UNSUPPORTED and it means the harness makes no claim either way - it classifies applied-status from the generation path recorded in the run parameters, the llama-server runs that produced these rows recorded none, and nothing in that serving path reports back whether a setting took effect. It is not the 1.0.0 arms' UNSUPPORTED, which records a setting that was read and had no effect. What was requested is published on the same rows - temperature 0.0 and seed 42 - and what was sent is under decoding above; whether the server applied it is the gap decoding_note names. The 1.0.0 frontier rows in per_item_results.jsonl read APPLIED in both fields instead; what that value records there, and what it does not, is under models.frontier.applied_settings_note.",
    "evaluation_hardware": "one AMD Radeon AI PRO R9700 in an AMD Ryzen 9 7950X3D host with 64 GB of system memory. The device is not recorded in the run parameters; it is stated from the evaluation host's configuration.",
    "environment_variables": "none were set for the server. This is recorded, not inferred from a log.",
    "library_versions_not_established": "The Python-side library versions in force during the 1.1.0 runs were not recorded and are not established. The server build above is pinned; the client stack is a gap.",
    "frontier_arm": {
      "scope": "the 300 EA holdout only; no frontier arm ran on the non-EA holdout",
      "evaluated_on": "2026-09-11",
      "temperature": 1.0,
      "seed_requested": 42,
      "concurrency": 1,
      "max_completion_tokens": 8192,
      "note": "served through the vendor API, which applied its own chat framing to the same system-prompt text and item text. A single run of a sampled decoder: a re-run may differ item for item."
    }
  },
  "frontier_evaluations": {
    "model_id": "gpt-5.6-sol",
    "provider": "openai",
    "identity_basis": "model name and evaluation date; the weights are vendor-hosted and are not independently pinned by this release",
    "public_184": {
      "evaluated_on": "2026-09-02",
      "temperature": 1.0,
      "seed_requested": 42,
      "max_completion_tokens": 8192
    },
    "holdout_ea300": {
      "evaluated_on": "2026-09-11",
      "temperature": 1.0,
      "seed_requested": 42,
      "concurrency": 1,
      "max_completion_tokens": 8192
    },
    "holdout_nonea200": {
      "evaluated": false,
      "note": "no frontier arm was run on the non-EA holdout, and no frontier figure for that set is estimated or implied"
    },
    "asymmetry": "the frontier arm sampled at temperature 1.0 one request at a time; the local arms ran greedy at concurrency 8. The head-to-head is published with that asymmetry stated rather than adjusted for."
  },
  "training_corpus_220k": {
    "rows": 215541,
    "sha256": "3c2f57db5986eaac49f7ae5ed6732e1ae702b339b1354bf3d94e33c7b611e023",
    "distributed": false,
    "why_withheld": "the corpus is not distributed. Claims that depend on it are attested, not independently reproducible from the public package."
  },
  "corpus_row_hash_manifest": {
    "file": "corpus_row_hashes.json",
    "row_hashes": 219867,
    "file_sha256": "4dbabd017b4c50f803a83ec981c99b53b9a2b2247ed1000bde4f571f21e1bcfd",
    "source_manifest_internal_sha256": "4c10a18620126e2379bca8c5c03d33c298fa5f4abc7b31665ff99893cfc18647",
    "source_manifest_internal_sha256_convention": "The private source manifest's own internal manifest_sha256 field, not the sha256 of any file: SHA-256 of the UTF-8 bytes of json.dumps(the manifest object with its own manifest_sha256 key removed, sort_keys=True, separators=(\",\", \":\"), ensure_ascii=True). It covers the construction counts and internal labels this projection drops as well as the row hashes it keeps, so it cannot be recomputed from corpus_row_hashes.json and is a commitment to the source, not a check on this file.",
    "projection_note": "This file is a projection of a retained extract, not a copy of it. The extract's header carried construction counts and internal labels that are not published, and each of its entries carried a positional index into a private build; what survives is the schema version, the committed row count, the hash rule, the trained-corpus pin and the 219,867 row hashes, carried across in the extract's own order with no value altered. The extract is CRLF and this file is LF, like every other file in the release, because .gitattributes pins LF and a CRLF file would be rewritten on checkout and stop matching its manifest hash. The retained extract's own sha256 is 3660bdfb621a9e222fdb2485b906223933edbe8f6e46f7fdd98aac7eef552490.",
    "contents_published": false,
    "names_published": false,
    "relation_to_the_trained_corpus": "The 219,867 hashes are the corpus before length filtering. The corpus the 220k arm trained on is the length-filtered pin of 215,541 rows. The relation is a declared filter - quarantine if and only if the rendered token length exceeds the maximum, and admitted rows are never truncated - recorded in the split manifest. It was not verified row by row and no row-level mapping between the two sets exists in this release.",
    "entry_shape": "Every entry carries one key, sha256, and nothing else. The order is the retained extract's order and carries no published meaning.",
    "who_can_verify_it": "The length threshold is not published and this release did not check the subset row by row, so the check that holds is one-directional: every row of the 215,541-row trained corpus hashes to a member of this manifest, and the 4,326 remaining hashes are the rows the length rule dropped. A counterparty holding the corpus can make that check; no one else can."
  },
  "public_release_contents_v1_1": {
    "holdout_ea300_results.jsonl": {
      "rows": 1200,
      "arms": {
        "base": 300,
        "tuned83k": 300,
        "tuned220k": 300,
        "frontier": 300
      },
      "distribution": "PUBLIC"
    },
    "holdout_nonea200_results.jsonl": {
      "rows": 800,
      "arms": {
        "base": 200,
        "tuned83k": 200,
        "tuned220k": 200,
        "tuned220k_q8_replay": 200
      },
      "distribution": "PUBLIC"
    },
    "bridge_q8_184_results.jsonl": {
      "rows": 552,
      "arms": {
        "base": 184,
        "tuned83k": 184,
        "tuned220k": 184
      },
      "distribution": "PUBLIC"
    },
    "corpus_row_hashes.json": {
      "row_hashes": 219867,
      "distribution": "PUBLIC"
    },
    "holdout_row_fields_published": [
      "arm",
      "item_sha256",
      "prompt_sha256",
      "category",
      "verdict_headline",
      "verdict_strict",
      "bucket",
      "errors",
      "warnings",
      "truncated",
      "finish_reason",
      "mq5_sha256",
      "ex5_exists",
      "ex5_sha256"
    ],
    "holdout_row_fields_withheld": [
      "adapter_path",
      "adapter_sha256",
      "base_model_sha256",
      "compile_infra_reason",
      "compile_log_path",
      "compile_log_raw_path",
      "compile_log_raw_sha256",
      "compile_log_sha256",
      "compile_stem",
      "ea_name",
      "elapsed_seconds",
      "ex5_path",
      "extractor_version",
      "fence_unterminated",
      "generated_code_path",
      "generated_code_sha256",
      "generation_error",
      "harness_version",
      "manifest_selection_sha256",
      "manifest_sha256",
      "max_new_tokens",
      "metaeditor_build",
      "metaeditor_path",
      "metaeditor_sha256",
      "metaeditor_size_bytes",
      "model_id",
      "mq5_path",
      "n_fenced_blocks",
      "positive_control_errors",
      "positive_control_warnings",
      "quantisation",
      "quantisation_mismatch",
      "result_line_errors",
      "result_line_warnings",
      "seed",
      "seed_applied",
      "source_sha256",
      "system_instruction",
      "system_instruction_sha256",
      "system_prompt_sha256",
      "temperature",
      "temperature_applied",
      "usage",
      "used_block_index",
      "used_block_unterminated"
    ],
    "holdout_row_fields_withheld_scope": "Measured, not curated: every key present on any sealed row of the eight arms behind the two holdout files, less the fourteen this release publishes. The sealed `arm` field is withheld too and is not in the list, because the published `arm` is derived from the source extract and the two are different fields - see PROJECTION_REPORT.json arm_derivation_v1_1. The prompt text itself is not a row field in any of them and is withheld with the item names.",
    "mq5_sha256_empty_output_sentinel": {
      "value": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
      "meaning": "the SHA-256 of zero bytes. Where mq5_sha256 holds it, no source was generated or recovered for that row and the row binds no generated MQL5 artifact: it is not a commitment and three rows sharing it are not three copies of one output.",
      "rows": 7,
      "where": "all seven are frontier rows of holdout_ea300_results.jsonl - the three generation failures, which carry null verdicts, and the four generations stopped at the 8,192-token ceiling, which are scored as failures. No row of any other file carries it, and every one of the 55 truncated local-arm rows in this release carries a real source hash, so truncation alone does not produce it.",
      "what_those_rows_do_bind": "item_sha256 and prompt_sha256, which identify the item and the request put to the arm, and the row's own finish_reason and bucket, which record how the generation ended. The four truncated rows also carry errors 1 and warnings 0, which is the compile result recorded against a source that is not there rather than evidence of a generated artifact; the three generation failures carry null in errors, in warnings and in both verdict fields."
    },
    "bridge_row_fields_published": [
      "arm",
      "item_id",
      "prompt_sha256_public",
      "verdict_headline",
      "verdict_strict",
      "bucket",
      "ex5_exists",
      "errors",
      "warnings",
      "result_line_errors",
      "result_line_warnings",
      "compile_infra_reason",
      "generation_error",
      "positive_control_errors",
      "positive_control_warnings",
      "truncated",
      "finish_reason",
      "n_fenced_blocks",
      "used_block_index",
      "fence_unterminated",
      "used_block_unterminated",
      "extractor_version",
      "harness_version",
      "seed",
      "seed_applied",
      "temperature",
      "temperature_applied",
      "max_new_tokens",
      "quantisation",
      "quantisation_mismatch",
      "prompt_sha256",
      "mq5_sha256",
      "ex5_sha256"
    ],
    "field_inventory_is_in": "PROJECTION_REPORT.json"
  },
  "statistics_v1_1": {
    "source": {
      "sha256": "027f3775194a7c895bd8de38f4d6dec9a630ce40cbf48b8f50ba0ca175acdb1d",
      "note": "Full-precision statistics as the statistician computed them, identified by the sha256 above. Every per-item name list has been removed and every field whose value was a filesystem path has been dropped: the holdout item names are not published. Some prose was rewritten for publication - internal set labels replaced by this release's own terms, emphasis normalised - and no figure was recomputed, rounded or re-derived. One value differs from the source deliberately: the Wilson lower bound of the base arm's 0-of-50 indicator cell on the 200-item non-EA set, which the source carries as 6.938893903907228e-18 and this release publishes as 0.0, in the primary block and in its secondary_spec_conformance mirror. The exact bound at 0 successes is zero and the sibling cells print 0.0; that value is binary floating-point residue. It is the only leaf in this block whose value was changed and the change is 7e-18."
    },
    "generated_utc": "2026-09-12T07:58:29+00:00",
    "nothing_rerun": "no arm was re-run, no item re-scored, no sample re-drawn",
    "read_only": true,
    "verdict_field": "verdict_headline",
    "secondary_field": "verdict_strict",
    "denominator_rule": "The denominator is every item present in the frozen manifest and present, by the item's name in the sealed rows, in every arm being compared. An item is excluded only when (a) it is in the contamination exclusion list for that set (the `clean` pool; the `full` pool keeps it), or (b) its preserved row shows truncation, an infrastructure compile failure, or a generation error (the `strict` variant; the `inclusive` variant keeps it and scores it as a failure). Both variants are computed and published side by side for every rate and every paired difference. The strict exclusion is scoped: a per-arm rate drops only that arm's own truncated, infrastructure or generation-error items, and a paired test drops an item truncated in either of the two arms being compared, so the base arm's truncations do not shrink a tuned-vs-tuned pair.",
    "denominator_variants": {
      "inclusive": "every aligned item counts; truncated and infrastructure items are scored as failures",
      "strict": "truncated, infrastructure and generation-error items are removed from the denominator (the pre-registered protocol, identified under the top-level `protocol` field: never silently scored as a model failure)"
    },
    "bootstrap": {
      "seed": 20260911,
      "resamples": 20000
    },
    "precision_checks": {
      "full_184_vs_clean_183": {
        "per_arm": {
          "base_q8": {
            "full_strict_kn": "2/182 (1.1%)",
            "clean_strict_kn": "2/181 (1.1%)",
            "delta_pct_points": 0.006071276789508806
          },
          "tuned83k_q8": {
            "full_strict_kn": "175/184 (95.1%)",
            "clean_strict_kn": "174/183 (95.1%)",
            "delta_pct_points": -0.026728439059155562
          },
          "tuned220k_q8": {
            "full_strict_kn": "169/184 (91.8%)",
            "clean_strict_kn": "168/183 (91.8%)",
            "delta_pct_points": -0.044547398431930674
          },
          "frontier_gpt56sol": {
            "full_strict_kn": "179/184 (97.3%)",
            "clean_strict_kn": "178/183 (97.3%)",
            "delta_pct_points": -0.014849132810638821
          }
        },
        "max_abs_delta_pct_points": 0.044547398431930674,
        "MATERIAL_DISAGREEMENT": false,
        "note": "Denominator ruling: on the public 184 the release publishes the full-184 figures - the 220k arm at 169/184 - and the clean-183 figures beside them are a precision check only. The one contamination collision is reported, not removed."
      },
      "replay_19_subset_vs_full_184": {
        "subset_n": 19,
        "subset_same_items_in_full_run_kn": "18/19 (94.7%)",
        "subset_replay_kn": "18/19 (94.7%)",
        "full_set_kn": "175/184 (95.1%)",
        "n_verdict_flips": 0,
        "note": "a 19-item cell is thin and indicative only; it is a stability instrument, not an estimate of the arm's rate"
      },
      "holdout_300_prefix_vs_clean_648": {
        "STATUS": "not computable",
        "why": "the 300 EA holdout is the first 300 of the clean 648 in frozen order; the remaining 348 items are held in reserve, unrun and unpublished, so no sample-vs-full-set comparison exists for this set. Reported as not computable rather than silently omitted.",
        "reserve_items_unrun": 348
      },
      "legacy_184_vs_holdout_300_same_arms": {
        "FLAG": "not a subset-vs-full precision check - these are different item sets. Published so a reader can see the two sets side by side with counts.",
        "per_arm": {
          "base_q8": {
            "legacy_184_full_strict_kn": "2/182 (1.1%)",
            "holdout_300_full_strict_kn": "1/297 (0.3%)",
            "delta_pct_points": -0.7622007622007623
          },
          "tuned83k_q8": {
            "legacy_184_full_strict_kn": "175/184 (95.1%)",
            "holdout_300_full_strict_kn": "281/300 (93.7%)",
            "delta_pct_points": -1.4420289855072355
          },
          "tuned220k_q8": {
            "legacy_184_full_strict_kn": "169/184 (91.8%)",
            "holdout_300_full_strict_kn": "282/300 (94.0%)",
            "delta_pct_points": 2.152173913043484
          },
          "frontier_gpt56sol": {
            "legacy_184_full_strict_kn": "179/184 (97.3%)",
            "holdout_300_full_strict_kn": "286/293 (97.6%)",
            "delta_pct_points": 0.32831280605431346
          }
        },
        "max_abs_delta_pct_points": 2.152173913043484
      },
      "nonea_c200_200_of_600_vs_the_full_non_ea_pool": {
        "STATUS": "not computable",
        "why": "the 200 non-EA holdout is a derived subset of the 600-item non-EA pool (per class, in the pool's own order, minus the pinned normalised-exact completion collisions, first 50 of each class). The remaining 400 pool items have never been run on any arm, so there is no full-set figure to compare the 200-item draw against. Reported as not computable rather than silently omitted. No re-draw was made and none may be.",
        "reserve_items_unrun": 400,
        "draw_rule": "the first 50 items of each of the four non-EA classes, in the pool's own order, after removing the items carrying a pinned normalised-exact completion collision with the training corpus."
      },
      "nonea_c200_per_class_dispersion": {
        "FLAG": "not a subset-vs-full precision check - it is the spread of the four 50-item class cells around the 200-item aggregate for the same arm. Published because a balanced aggregate can hide a dead class.",
        "per_arm": {
          "base_q8": {
            "aggregate_strict_kn": "56/186 (30.1%)",
            "per_class_strict_kn": {
              "include": "20/43 (46.5%)",
              "indicator": "0/45 (0.0%)",
              "script": "13/48 (27.1%)",
              "service": "23/50 (46.0%)"
            },
            "max_abs_class_minus_aggregate_pct_points": 30.107526881720432,
            "worst_class": "indicator"
          },
          "tuned83k_q8": {
            "aggregate_strict_kn": "135/187 (72.2%)",
            "per_class_strict_kn": {
              "include": "38/42 (90.5%)",
              "indicator": "49/50 (98.0%)",
              "script": "19/45 (42.2%)",
              "service": "29/50 (58.0%)"
            },
            "max_abs_class_minus_aggregate_pct_points": 29.97029114676173,
            "worst_class": "script"
          },
          "tuned220k_q8": {
            "aggregate_strict_kn": "168/188 (89.4%)",
            "per_class_strict_kn": {
              "include": "31/39 (79.5%)",
              "indicator": "50/50 (100.0%)",
              "script": "39/49 (79.6%)",
              "service": "48/50 (96.0%)"
            },
            "max_abs_class_minus_aggregate_pct_points": 10.63829787234043,
            "worst_class": "include"
          }
        }
      },
      "EA_holdout_300_vs_nonEA_C200_same_arms": {
        "FLAG": "not a precision check - different item sets and different task families, EA against non-EA. Published with counts so a reader can see both sets side by side; it is not a paired comparison.",
        "per_arm": {
          "base_q8": {
            "EA_holdout_300_strict_kn": "1/297 (0.3%)",
            "nonEA_C200_strict_kn": "56/186 (30.1%)",
            "delta_pct_points": 29.770826545020096
          },
          "tuned83k_q8": {
            "EA_holdout_300_strict_kn": "281/300 (93.7%)",
            "nonEA_C200_strict_kn": "135/187 (72.2%)",
            "delta_pct_points": -21.47415329768272
          },
          "tuned220k_q8": {
            "EA_holdout_300_strict_kn": "282/300 (94.0%)",
            "nonEA_C200_strict_kn": "168/188 (89.4%)",
            "delta_pct_points": -4.638297872340431
          }
        }
      }
    },
    "q8_bridge": {
      "title": "Part 5 \u2014 Q8 protocol bridge (83k BF16 vs 83k Q8_0 on the same 184 items)",
      "CAVEAT": "The bf16 arm is the 1.0.0 tuned arm as published, identified by the 1.0.0 weight hashes; this release did not re-verify those weights. That identity is recorded, not verified, and the run is labelled arm 'A' in its own run parameters, so the bridge is indicative rather than attested.",
      "bf16_results_sha256": "bbeadfe8ce96e92bf7183e0eddbb94ad087711ccb39a386118b7e61bd3378dd8",
      "bf16_rows_total": 300,
      "n_items_joined": 184,
      "q8_pass_kn": "175/184 (95.1%)",
      "bf16_pass_kn": "170/184 (92.4%)",
      "marginal_item_difference": 5,
      "confounds_named": "the BF16 arm ran on a different serving backend (local transformers) from the Q8 arm (llama-server, -np 8, concurrency 8). A difference here is not attributable to quantisation alone.",
      "paired": {
        "inclusive": {
          "arm_x": "83k Q8_0",
          "arm_y": "83k BF16",
          "n_pairs": 184,
          "contingency": {
            "a_both_pass": 169,
            "b_x_pass_y_fail": 6,
            "c_x_fail_y_pass": 1,
            "d_neither": 8
          },
          "x_rate_kn": "175/184 (95.1%)",
          "y_rate_kn": "170/184 (92.4%)",
          "diff_x_minus_y": 0.02717391304347826,
          "diff_pct_points": 2.717391304347826,
          "mcnemar": {
            "test": "McNemar exact (binomial, two-sided)",
            "b": 6,
            "c": 1,
            "n_discordant": 7,
            "p_value": 0.125
          },
          "ci95_diff_newcombe": [
            -0.0026328202453298165,
            0.0628229813724678
          ],
          "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
          "ci95_diff_bootstrap": {
            "method": "paired bootstrap over items (percentile, 2.5/97.5)",
            "resamples": 20000,
            "seed": 20260911,
            "lo": 0.0,
            "hi": 0.05434782608695652
          },
          "dominance": "NO dominance - each arm fails items the other passes, in both directions"
        },
        "strict": {
          "arm_x": "83k Q8_0",
          "arm_y": "83k BF16",
          "n_pairs": 184,
          "contingency": {
            "a_both_pass": 169,
            "b_x_pass_y_fail": 6,
            "c_x_fail_y_pass": 1,
            "d_neither": 8
          },
          "x_rate_kn": "175/184 (95.1%)",
          "y_rate_kn": "170/184 (92.4%)",
          "diff_x_minus_y": 0.02717391304347826,
          "diff_pct_points": 2.717391304347826,
          "mcnemar": {
            "test": "McNemar exact (binomial, two-sided)",
            "b": 6,
            "c": 1,
            "n_discordant": 7,
            "p_value": 0.125
          },
          "ci95_diff_newcombe": [
            -0.0026328202453298165,
            0.0628229813724678
          ],
          "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
          "ci95_diff_bootstrap": {
            "method": "paired bootstrap over items (percentile, 2.5/97.5)",
            "resamples": 20000,
            "seed": 20260911,
            "lo": 0.0,
            "hi": 0.05434782608695652
          },
          "dominance": "NO dominance - each arm fails items the other passes, in both directions"
        }
      }
    },
    "replay_19": {
      "title": "Release-stability instrument \u2014 19-item replay, 83k Q8_0",
      "replay_results_sha256": "558ec975001179ec1a2772bf6fe197d2c56359ea2561192816985569ce026ca6",
      "n_items_joined": 19,
      "replay_pass_kn": "18/19 (94.7%)",
      "reference_pass_kn": "18/19 (94.7%)",
      "verdict_flips": [],
      "n_verdict_flips": 0,
      "generation_bytes_identical_kn": "18/19 (94.7%)",
      "generation_bytes_differing": 1,
      "meaning": "One of 19 generation hashes differed under the same recorded settings, and no verdict flipped; this release does not isolate why. On a row where generation produced source, the output hash is provenance of the observed run, not a promise that a re-generation reproduces it."
    },
    "secondary_spec_conformance_flag": "Every secondary_spec_conformance block in this statistics section is self-declared and non-independent: the verdict_strict figures in it come from this project's own verifier, not from an external judgement. Each block points back here rather than repeating the caveat.",
    "sets": {
      "legacy_184": {
        "title": "Part 1 \u2014 legacy 184-item public benchmark (historical comparability)",
        "manifest_file_sha256": "6d09240ae5a42493516272ab9645fb38afc3381d29ad0bdbd1432b1d63c8da24",
        "per_arm": {
          "full": {
            "base_q8": {
              "inclusive": {
                "k": 2,
                "n": 184,
                "rate_pct": 1.0869565217391304,
                "kn": "2/184 (1.1%)",
                "ci95": [
                  0.002985905599769044,
                  0.03875918555276116
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 2,
                  "n": 184,
                  "kn": "2/184 (1.1%)",
                  "ci95": [
                    0.002985905599769044,
                    0.03875918555276116
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 2,
                "n": 182,
                "rate_pct": 1.098901098901099,
                "kn": "2/182 (1.1%)",
                "ci95": [
                  0.003018774375770715,
                  0.039175571656886184
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 2,
                  "n": 182,
                  "kn": "2/182 (1.1%)",
                  "ci95": [
                    0.003018774375770715,
                    0.039175571656886184
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned83k_q8": {
              "inclusive": {
                "k": 175,
                "n": 184,
                "rate_pct": 95.1086956521739,
                "kn": "175/184 (95.1%)",
                "ci95": [
                  0.9096677238294343,
                  0.9740562476632323
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 175,
                  "n": 184,
                  "kn": "175/184 (95.1%)",
                  "ci95": [
                    0.9096677238294343,
                    0.9740562476632323
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 175,
                "n": 184,
                "rate_pct": 95.1086956521739,
                "kn": "175/184 (95.1%)",
                "ci95": [
                  0.9096677238294343,
                  0.9740562476632323
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 175,
                  "n": 184,
                  "kn": "175/184 (95.1%)",
                  "ci95": [
                    0.9096677238294343,
                    0.9740562476632323
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned220k_q8": {
              "inclusive": {
                "k": 169,
                "n": 184,
                "rate_pct": 91.84782608695652,
                "kn": "169/184 (91.8%)",
                "ci95": [
                  0.8698640135719263,
                  0.9499762973309088
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 169,
                  "n": 184,
                  "kn": "169/184 (91.8%)",
                  "ci95": [
                    0.8698640135719263,
                    0.9499762973309088
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 169,
                "n": 184,
                "rate_pct": 91.84782608695652,
                "kn": "169/184 (91.8%)",
                "ci95": [
                  0.8698640135719263,
                  0.9499762973309088
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 169,
                  "n": 184,
                  "kn": "169/184 (91.8%)",
                  "ci95": [
                    0.8698640135719263,
                    0.9499762973309088
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "frontier_gpt56sol": {
              "inclusive": {
                "k": 179,
                "n": 184,
                "rate_pct": 97.28260869565217,
                "kn": "179/184 (97.3%)",
                "ci95": [
                  0.9379747965705482,
                  0.988338281982006
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 176,
                  "n": 184,
                  "kn": "176/184 (95.7%)",
                  "ci95": [
                    0.9165644224588859,
                    0.9778068257987526
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 179,
                "n": 184,
                "rate_pct": 97.28260869565217,
                "kn": "179/184 (97.3%)",
                "ci95": [
                  0.9379747965705482,
                  0.988338281982006
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 176,
                  "n": 184,
                  "kn": "176/184 (95.7%)",
                  "ci95": [
                    0.9165644224588859,
                    0.9778068257987526
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            }
          },
          "clean": {
            "base_q8": {
              "inclusive": {
                "k": 2,
                "n": 183,
                "rate_pct": 1.092896174863388,
                "kn": "2/183 (1.1%)",
                "ci95": [
                  0.003002250027948822,
                  0.038966266329115845
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 2,
                  "n": 183,
                  "kn": "2/183 (1.1%)",
                  "ci95": [
                    0.003002250027948822,
                    0.038966266329115845
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 2,
                "n": 181,
                "rate_pct": 1.1049723756906078,
                "kn": "2/181 (1.1%)",
                "ci95": [
                  0.003035481630522502,
                  0.03938713757365047
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 2,
                  "n": 181,
                  "kn": "2/181 (1.1%)",
                  "ci95": [
                    0.003035481630522502,
                    0.03938713757365047
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned83k_q8": {
              "inclusive": {
                "k": 174,
                "n": 183,
                "rate_pct": 95.08196721311475,
                "kn": "174/183 (95.1%)",
                "ci95": [
                  0.9091883431019155,
                  0.9739133038315165
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 174,
                  "n": 183,
                  "kn": "174/183 (95.1%)",
                  "ci95": [
                    0.9091883431019155,
                    0.9739133038315165
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 174,
                "n": 183,
                "rate_pct": 95.08196721311475,
                "kn": "174/183 (95.1%)",
                "ci95": [
                  0.9091883431019155,
                  0.9739133038315165
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 174,
                  "n": 183,
                  "kn": "174/183 (95.1%)",
                  "ci95": [
                    0.9091883431019155,
                    0.9739133038315165
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned220k_q8": {
              "inclusive": {
                "k": 168,
                "n": 183,
                "rate_pct": 91.80327868852459,
                "kn": "168/183 (91.8%)",
                "ci95": [
                  0.8691766392132034,
                  0.9496994333977973
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 168,
                  "n": 183,
                  "kn": "168/183 (91.8%)",
                  "ci95": [
                    0.8691766392132034,
                    0.9496994333977973
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 168,
                "n": 183,
                "rate_pct": 91.80327868852459,
                "kn": "168/183 (91.8%)",
                "ci95": [
                  0.8691766392132034,
                  0.9496994333977973
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 168,
                  "n": 183,
                  "kn": "168/183 (91.8%)",
                  "ci95": [
                    0.8691766392132034,
                    0.9496994333977973
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "frontier_gpt56sol": {
              "inclusive": {
                "k": 178,
                "n": 183,
                "rate_pct": 97.26775956284153,
                "kn": "178/183 (97.3%)",
                "ci95": [
                  0.93764444316215,
                  0.9882742533195696
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 175,
                  "n": 183,
                  "kn": "175/183 (95.6%)",
                  "ci95": [
                    0.9161212614482003,
                    0.9776846478723037
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 178,
                "n": 183,
                "rate_pct": 97.26775956284153,
                "kn": "178/183 (97.3%)",
                "ci95": [
                  0.93764444316215,
                  0.9882742533195696
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 175,
                  "n": 183,
                  "kn": "175/183 (95.6%)",
                  "ci95": [
                    0.9161212614482003,
                    0.9776846478723037
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            }
          }
        },
        "per_class": null,
        "paired": {
          "full": {
            "tuned83k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 174,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 8
                },
                "x_rate_kn": "175/184 (95.1%)",
                "y_rate_kn": "2/184 (1.1%)",
                "diff_x_minus_y": 0.9402173913043478,
                "diff_pct_points": 94.02173913043478,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 174,
                  "c": 1,
                  "n_discordant": 175,
                  "p_value": 7.350102952753538e-51
                },
                "ci95_diff_newcombe": [
                  0.8854460655822902,
                  0.966085147345344
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9021739130434783,
                  "hi": 0.9728260869565217
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 182,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 173,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 7
                },
                "x_rate_kn": "174/182 (95.6%)",
                "y_rate_kn": "2/182 (1.1%)",
                "diff_x_minus_y": 0.945054945054945,
                "diff_pct_points": 94.50549450549451,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 173,
                  "c": 1,
                  "n_discordant": 174,
                  "p_value": 1.4616682008316695e-50
                },
                "ci95_diff_newcombe": [
                  0.890668012548651,
                  0.9696911761441854
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9065934065934066,
                  "hi": 0.978021978021978
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 168,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 14
                },
                "x_rate_kn": "169/184 (91.8%)",
                "y_rate_kn": "2/184 (1.1%)",
                "diff_x_minus_y": 0.907608695652174,
                "diff_pct_points": 90.76086956521739,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 168,
                  "c": 1,
                  "n_discordant": 169,
                  "p_value": 4.5437000071567326e-49
                },
                "ci95_diff_newcombe": [
                  0.8478098364646453,
                  0.9412821039960407
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.8586956521739131,
                  "hi": 0.9510869565217391
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 182,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 167,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 13
                },
                "x_rate_kn": "168/182 (92.3%)",
                "y_rate_kn": "2/182 (1.1%)",
                "diff_x_minus_y": 0.9120879120879121,
                "diff_pct_points": 91.20879120879121,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 167,
                  "c": 1,
                  "n_discordant": 168,
                  "p_value": 9.033944720111621e-49
                },
                "ci95_diff_newcombe": [
                  0.8524578087835255,
                  0.944927036904818
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.8681318681318682,
                  "hi": 0.9560439560439561
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__tuned83k_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 163,
                  "b_x_pass_y_fail": 6,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "169/184 (91.8%)",
                "y_rate_kn": "175/184 (95.1%)",
                "diff_x_minus_y": -0.03260869565217391,
                "diff_pct_points": -3.260869565217391,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 6,
                  "c": 12,
                  "n_discordant": 18,
                  "p_value": 0.237884521484375
                },
                "ci95_diff_newcombe": [
                  -0.08185163231317105,
                  0.013900964632705883
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.07608695652173914,
                  "hi": 0.010869565217391304
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 163,
                  "b_x_pass_y_fail": 6,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "169/184 (91.8%)",
                "y_rate_kn": "175/184 (95.1%)",
                "diff_x_minus_y": -0.03260869565217391,
                "diff_pct_points": -3.260869565217391,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 6,
                  "c": 12,
                  "n_discordant": 18,
                  "p_value": 0.237884521484375
                },
                "ci95_diff_newcombe": [
                  -0.08185163231317105,
                  0.013900964632705883
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.07608695652173914,
                  "hi": 0.010869565217391304
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 167,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "169/184 (91.8%)",
                "y_rate_kn": "179/184 (97.3%)",
                "diff_x_minus_y": -0.05434782608695652,
                "diff_pct_points": -5.434782608695652,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 12,
                  "n_discordant": 14,
                  "p_value": 0.012939453125
                },
                "ci95_diff_newcombe": [
                  -0.10046019777816709,
                  -0.015470146379077683
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.09239130434782608,
                  "hi": -0.016304347826086956
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 167,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "169/184 (91.8%)",
                "y_rate_kn": "179/184 (97.3%)",
                "diff_x_minus_y": -0.05434782608695652,
                "diff_pct_points": -5.434782608695652,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 12,
                  "n_discordant": 14,
                  "p_value": 0.012939453125
                },
                "ci95_diff_newcombe": [
                  -0.10046019777816709,
                  -0.015470146379077683
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.09239130434782608,
                  "hi": -0.016304347826086956
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned83k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 173,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 6,
                  "d_neither": 3
                },
                "x_rate_kn": "175/184 (95.1%)",
                "y_rate_kn": "179/184 (97.3%)",
                "diff_x_minus_y": -0.021739130434782608,
                "diff_pct_points": -2.1739130434782608,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 6,
                  "n_discordant": 8,
                  "p_value": 0.2890625
                },
                "ci95_diff_newcombe": [
                  -0.05925509430454449,
                  0.010795792178232518
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.05434782608695652,
                  "hi": 0.005434782608695652
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 184,
                "contingency": {
                  "a_both_pass": 173,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 6,
                  "d_neither": 3
                },
                "x_rate_kn": "175/184 (95.1%)",
                "y_rate_kn": "179/184 (97.3%)",
                "diff_x_minus_y": -0.021739130434782608,
                "diff_pct_points": -2.1739130434782608,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 6,
                  "n_discordant": 8,
                  "p_value": 0.2890625
                },
                "ci95_diff_newcombe": [
                  -0.05925509430454449,
                  0.010795792178232518
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.05434782608695652,
                  "hi": 0.005434782608695652
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            }
          },
          "clean": {
            "tuned83k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 173,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 8
                },
                "x_rate_kn": "174/183 (95.1%)",
                "y_rate_kn": "2/183 (1.1%)",
                "diff_x_minus_y": 0.9398907103825137,
                "diff_pct_points": 93.98907103825137,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 173,
                  "c": 1,
                  "n_discordant": 174,
                  "p_value": 1.4616682008316695e-50
                },
                "ci95_diff_newcombe": [
                  0.8848375909442479,
                  0.9658980334464432
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9016393442622951,
                  "hi": 0.9726775956284153
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 181,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 172,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 7
                },
                "x_rate_kn": "173/181 (95.6%)",
                "y_rate_kn": "2/181 (1.1%)",
                "diff_x_minus_y": 0.9447513812154696,
                "diff_pct_points": 94.47513812154696,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 172,
                  "c": 1,
                  "n_discordant": 173,
                  "p_value": 2.906631622225263e-50
                },
                "ci95_diff_newcombe": [
                  0.8900805879545658,
                  0.9695222065436557
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9060773480662984,
                  "hi": 0.9779005524861878
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 167,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 14
                },
                "x_rate_kn": "168/183 (91.8%)",
                "y_rate_kn": "2/183 (1.1%)",
                "diff_x_minus_y": 0.907103825136612,
                "diff_pct_points": 90.7103825136612,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 167,
                  "c": 1,
                  "n_discordant": 168,
                  "p_value": 9.033944720111621e-49
                },
                "ci95_diff_newcombe": [
                  0.8470049904851806,
                  0.9409568734283333
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.8579234972677595,
                  "hi": 0.9508196721311475
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 181,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 166,
                  "c_x_fail_y_pass": 1,
                  "d_neither": 13
                },
                "x_rate_kn": "167/181 (92.3%)",
                "y_rate_kn": "2/181 (1.1%)",
                "diff_x_minus_y": 0.9116022099447514,
                "diff_pct_points": 91.16022099447514,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 166,
                  "c": 1,
                  "n_discordant": 167,
                  "p_value": 1.7960978851819555e-48
                },
                "ci95_diff_newcombe": [
                  0.851668647267405,
                  0.944618775580622
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.8674033149171271,
                  "hi": 0.9502762430939227
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__tuned83k_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 162,
                  "b_x_pass_y_fail": 6,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "168/183 (91.8%)",
                "y_rate_kn": "174/183 (95.1%)",
                "diff_x_minus_y": -0.03278688524590164,
                "diff_pct_points": -3.278688524590164,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 6,
                  "c": 12,
                  "n_discordant": 18,
                  "p_value": 0.237884521484375
                },
                "ci95_diff_newcombe": [
                  -0.08228406453623008,
                  0.013972693215261692
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.07650273224043716,
                  "hi": 0.01092896174863388
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 162,
                  "b_x_pass_y_fail": 6,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "168/183 (91.8%)",
                "y_rate_kn": "174/183 (95.1%)",
                "diff_x_minus_y": -0.03278688524590164,
                "diff_pct_points": -3.278688524590164,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 6,
                  "c": 12,
                  "n_discordant": 18,
                  "p_value": 0.237884521484375
                },
                "ci95_diff_newcombe": [
                  -0.08228406453623008,
                  0.013972693215261692
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.07650273224043716,
                  "hi": 0.01092896174863388
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 166,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "168/183 (91.8%)",
                "y_rate_kn": "178/183 (97.3%)",
                "diff_x_minus_y": -0.0546448087431694,
                "diff_pct_points": -5.46448087431694,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 12,
                  "n_discordant": 14,
                  "p_value": 0.012939453125
                },
                "ci95_diff_newcombe": [
                  -0.10098946523101315,
                  -0.015557162942926687
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.09836065573770492,
                  "hi": -0.01639344262295082
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 166,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 3
                },
                "x_rate_kn": "168/183 (91.8%)",
                "y_rate_kn": "178/183 (97.3%)",
                "diff_x_minus_y": -0.0546448087431694,
                "diff_pct_points": -5.46448087431694,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 12,
                  "n_discordant": 14,
                  "p_value": 0.012939453125
                },
                "ci95_diff_newcombe": [
                  -0.10098946523101315,
                  -0.015557162942926687
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.09836065573770492,
                  "hi": -0.01639344262295082
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned83k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 172,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 6,
                  "d_neither": 3
                },
                "x_rate_kn": "174/183 (95.1%)",
                "y_rate_kn": "178/183 (97.3%)",
                "diff_x_minus_y": -0.02185792349726776,
                "diff_pct_points": -2.185792349726776,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 6,
                  "n_discordant": 8,
                  "p_value": 0.2890625
                },
                "ci95_diff_newcombe": [
                  -0.05956767726362802,
                  0.010850721392073048
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.0546448087431694,
                  "hi": 0.00546448087431694
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 183,
                "contingency": {
                  "a_both_pass": 172,
                  "b_x_pass_y_fail": 2,
                  "c_x_fail_y_pass": 6,
                  "d_neither": 3
                },
                "x_rate_kn": "174/183 (95.1%)",
                "y_rate_kn": "178/183 (97.3%)",
                "diff_x_minus_y": -0.02185792349726776,
                "diff_pct_points": -2.185792349726776,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 2,
                  "c": 6,
                  "n_discordant": 8,
                  "p_value": 0.2890625
                },
                "ci95_diff_newcombe": [
                  -0.05956767726362802,
                  0.010850721392073048
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.0546448087431694,
                  "hi": 0.00546448087431694
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            }
          }
        },
        "harness_bucket_counts_per_arm": {
          "base_q8": {
            "compile-fail": 180,
            "TRUNCATED": 2,
            "compile-pass": 2
          },
          "tuned83k_q8": {
            "compile-pass": 175,
            "compile-fail": 9
          },
          "tuned220k_q8": {
            "compile-pass": 169,
            "compile-fail": 15
          },
          "frontier_gpt56sol": {
            "compile-pass": 179,
            "compile-fail": 5
          }
        },
        "exclusion_counts_by_class_per_arm": {
          "scored": {
            "base_q8": 181,
            "tuned83k_q8": 183,
            "tuned220k_q8": 183,
            "frontier_gpt56sol": 183
          },
          "contaminated": {
            "base_q8": 1,
            "tuned83k_q8": 1,
            "tuned220k_q8": 1,
            "frontier_gpt56sol": 1
          },
          "truncated": {
            "base_q8": 2
          }
        },
        "exclusion_buckets_by_artifact_class": null,
        "denominator_pools": {
          "full": 184,
          "clean": 183
        },
        "item_alignment": {
          "join_key": "ea_name",
          "per_arm_unique_names": {
            "base_q8": 184,
            "tuned83k_q8": 184,
            "tuned220k_q8": 184,
            "frontier_gpt56sol": 184
          },
          "intersection_n": 184,
          "per_arm_names_missing_from_intersection": {},
          "duplicate_ea_names_per_arm": {},
          "ALIGNED": true,
          "prompt_identity": {
            "items_with_identical_prompt_sha256_across_all_arms": 0,
            "items_disagreeing_with_the_majority_prompt_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 184
            },
            "ALL_ARMS_BYTE_IDENTICAL_PROMPTS": false,
            "note": "a differing prompt_sha256 means the arm was served a differently assembled prompt string; the head-to-head with that arm is confounded and is reported as such"
          },
          "spec_identity": {
            "items_with_identical_spec_sha256_across_all_arms": 184,
            "items_disagreeing_with_the_majority_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 0
            },
            "rows_missing_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 0
            },
            "ALL_ARMS_IDENTICAL_SPEC": true,
            "note": "spec_sha256 identity is the pairing proof (the same item was put to every arm); prompt_sha256 identity is the separate byte-identity requirement of the pre-registered protocol (see the top-level `protocol` field) and is reported beside it"
          },
          "manifest_selection_sha256_identical_across_arms": true
        },
        "instrument_validation": {
          "base_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 4,
              "row_errors_field": 4,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 4,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned83k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 1,
              "row_errors_field": 1,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 1,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned220k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 7,
              "row_errors_field": 7,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 7,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "frontier_gpt56sol": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 1,
              "row_errors_field": 1,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 1,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          }
        },
        "artifact_class_counts": null,
        "category_field": "The `category` key is present on every row of this set and its value is the empty string on 100% of rows, so no artifact-class breakdown is computable from the rows. It is computable from the published files: prompts.jsonl carries each item's own specification, and six of the 184 declare an artifact type there on an explicit `// TYPE:` line. They are listed under the top-level public_set_184_composition field.",
        "category_key_present_on_all_rows": true,
        "inclusive_scoring_anomalies": {
          "base_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 184,
            "inclusive_k_under_the_PREREGISTERED_rule": 2,
            "inclusive_k_if_the_row_verdict_were_honoured": 2,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned83k_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 184,
            "inclusive_k_under_the_PREREGISTERED_rule": 175,
            "inclusive_k_if_the_row_verdict_were_honoured": 175,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned220k_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 184,
            "inclusive_k_under_the_PREREGISTERED_rule": 169,
            "inclusive_k_if_the_row_verdict_were_honoured": 169,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "frontier_gpt56sol": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 184,
            "inclusive_k_under_the_PREREGISTERED_rule": 179,
            "inclusive_k_if_the_row_verdict_were_honoured": 179,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          }
        }
      },
      "holdout_300": {
        "title": "Part 2 \u2014 300-item EA holdout (clean-648 prefix-300, frozen order)",
        "manifest_file_sha256": "75021a9d8db3fc40931de14239e7e279d37967c109dc0de12c4637473022f64a",
        "per_arm": {
          "full": {
            "base_q8": {
              "inclusive": {
                "k": 1,
                "n": 300,
                "rate_pct": 0.3333333333333333,
                "kn": "1/300 (0.3%)",
                "ci95": [
                  0.0005886577218511581,
                  0.018636693694531237
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 1,
                  "n": 300,
                  "kn": "1/300 (0.3%)",
                  "ci95": [
                    0.0005886577218511581,
                    0.018636693694531237
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 1,
                "n": 297,
                "rate_pct": 0.3367003367003367,
                "kn": "1/297 (0.3%)",
                "ci95": [
                  0.0005946062361728496,
                  0.018822460957932597
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 1,
                  "n": 297,
                  "kn": "1/297 (0.3%)",
                  "ci95": [
                    0.0005946062361728496,
                    0.018822460957932597
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned83k_q8": {
              "inclusive": {
                "k": 281,
                "n": 300,
                "rate_pct": 93.66666666666667,
                "kn": "281/300 (93.7%)",
                "ci95": [
                  0.9032086875144784,
                  0.9590831176160847
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 281,
                  "n": 300,
                  "kn": "281/300 (93.7%)",
                  "ci95": [
                    0.9032086875144784,
                    0.9590831176160847
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 281,
                "n": 300,
                "rate_pct": 93.66666666666667,
                "kn": "281/300 (93.7%)",
                "ci95": [
                  0.9032086875144784,
                  0.9590831176160847
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 281,
                  "n": 300,
                  "kn": "281/300 (93.7%)",
                  "ci95": [
                    0.9032086875144784,
                    0.9590831176160847
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned220k_q8": {
              "inclusive": {
                "k": 282,
                "n": 300,
                "rate_pct": 94.0,
                "kn": "282/300 (94.0%)",
                "ci95": [
                  0.9071605548470915,
                  0.9617136304753079
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 281,
                  "n": 300,
                  "kn": "281/300 (93.7%)",
                  "ci95": [
                    0.9032086875144784,
                    0.9590831176160847
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 282,
                "n": 300,
                "rate_pct": 94.0,
                "kn": "282/300 (94.0%)",
                "ci95": [
                  0.9071605548470915,
                  0.9617136304753079
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 281,
                  "n": 300,
                  "kn": "281/300 (93.7%)",
                  "ci95": [
                    0.9032086875144784,
                    0.9590831176160847
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "frontier_gpt56sol": {
              "inclusive": {
                "k": 286,
                "n": 300,
                "rate_pct": 95.33333333333333,
                "kn": "286/300 (95.3%)",
                "ci95": [
                  0.923202639772079,
                  0.9720010663176659
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 282,
                  "n": 300,
                  "kn": "282/300 (94.0%)",
                  "ci95": [
                    0.9071605548470915,
                    0.9617136304753079
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 286,
                "n": 293,
                "rate_pct": 97.61092150170649,
                "kn": "286/293 (97.6%)",
                "ci95": [
                  0.9515155343949656,
                  0.9883801294668967
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 282,
                  "n": 293,
                  "kn": "282/293 (96.2%)",
                  "ci95": [
                    0.9340355857729857,
                    0.9789096647881782
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            }
          }
        },
        "per_class": null,
        "paired": {
          "full": {
            "tuned83k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 280,
                  "c_x_fail_y_pass": 0,
                  "d_neither": 19
                },
                "x_rate_kn": "281/300 (93.7%)",
                "y_rate_kn": "1/300 (0.3%)",
                "diff_x_minus_y": 0.9333333333333333,
                "diff_pct_points": 93.33333333333333,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 280,
                  "c": 0,
                  "n_discordant": 280,
                  "p_value": 1.0295115178936058e-84
                },
                "ci95_diff_newcombe": [
                  0.8967515102434487,
                  0.9558761830532138
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9033333333333333,
                  "hi": 0.96
                },
                "dominance": "X strictly dominates Y (c=0)"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 297,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 278,
                  "c_x_fail_y_pass": 0,
                  "d_neither": 18
                },
                "x_rate_kn": "279/297 (93.9%)",
                "y_rate_kn": "1/297 (0.3%)",
                "diff_x_minus_y": 0.936026936026936,
                "diff_pct_points": 93.6026936026936,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 278,
                  "c": 0,
                  "n_discordant": 278,
                  "p_value": 4.118046071574423e-84
                },
                "ci95_diff_newcombe": [
                  0.8996577628152559,
                  0.9580903377627212
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9057239057239057,
                  "hi": 0.9629629629629629
                },
                "dominance": "X strictly dominates Y (c=0)"
              }
            },
            "tuned220k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 281,
                  "c_x_fail_y_pass": 0,
                  "d_neither": 18
                },
                "x_rate_kn": "282/300 (94.0%)",
                "y_rate_kn": "1/300 (0.3%)",
                "diff_x_minus_y": 0.9366666666666666,
                "diff_pct_points": 93.66666666666667,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 281,
                  "c": 0,
                  "n_discordant": 281,
                  "p_value": 5.147557589468029e-85
                },
                "ci95_diff_newcombe": [
                  0.9006397809140065,
                  0.9585132560623
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9066666666666666,
                  "hi": 0.9633333333333334
                },
                "dominance": "X strictly dominates Y (c=0)"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 297,
                "contingency": {
                  "a_both_pass": 1,
                  "b_x_pass_y_fail": 279,
                  "c_x_fail_y_pass": 0,
                  "d_neither": 17
                },
                "x_rate_kn": "280/297 (94.3%)",
                "y_rate_kn": "1/297 (0.3%)",
                "diff_x_minus_y": 0.9393939393939394,
                "diff_pct_points": 93.93939393939394,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 279,
                  "c": 0,
                  "n_discordant": 279,
                  "p_value": 2.0590230357872116e-84
                },
                "ci95_diff_newcombe": [
                  0.9036017469011658,
                  0.9607331722947036
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.9090909090909091,
                  "hi": 0.9663299663299664
                },
                "dominance": "X strictly dominates Y (c=0)"
              }
            },
            "tuned220k_q8__vs__tuned83k_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 264,
                  "b_x_pass_y_fail": 18,
                  "c_x_fail_y_pass": 17,
                  "d_neither": 1
                },
                "x_rate_kn": "282/300 (94.0%)",
                "y_rate_kn": "281/300 (93.7%)",
                "diff_x_minus_y": 0.0033333333333333335,
                "diff_pct_points": 0.3333333333333333,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 18,
                  "c": 17,
                  "n_discordant": 35,
                  "p_value": 1.0
                },
                "ci95_diff_newcombe": [
                  -0.03657662339661758,
                  0.04336632758652774
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.03333333333333333,
                  "hi": 0.043333333333333335
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 264,
                  "b_x_pass_y_fail": 18,
                  "c_x_fail_y_pass": 17,
                  "d_neither": 1
                },
                "x_rate_kn": "282/300 (94.0%)",
                "y_rate_kn": "281/300 (93.7%)",
                "diff_x_minus_y": 0.0033333333333333335,
                "diff_pct_points": 0.3333333333333333,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 18,
                  "c": 17,
                  "n_discordant": 35,
                  "p_value": 1.0
                },
                "ci95_diff_newcombe": [
                  -0.03657662339661758,
                  0.04336632758652774
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.03333333333333333,
                  "hi": 0.043333333333333335
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 270,
                  "b_x_pass_y_fail": 12,
                  "c_x_fail_y_pass": 16,
                  "d_neither": 2
                },
                "x_rate_kn": "282/300 (94.0%)",
                "y_rate_kn": "286/300 (95.3%)",
                "diff_x_minus_y": -0.013333333333333334,
                "diff_pct_points": -1.3333333333333333,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 12,
                  "c": 16,
                  "n_discordant": 28,
                  "p_value": 0.5715881884098053
                },
                "ci95_diff_newcombe": [
                  -0.04983361208865994,
                  0.02242048353135713
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.05,
                  "hi": 0.02
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 293,
                "contingency": {
                  "a_both_pass": 270,
                  "b_x_pass_y_fail": 6,
                  "c_x_fail_y_pass": 16,
                  "d_neither": 1
                },
                "x_rate_kn": "276/293 (94.2%)",
                "y_rate_kn": "286/293 (97.6%)",
                "diff_x_minus_y": -0.034129692832764506,
                "diff_pct_points": -3.4129692832764507,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 6,
                  "c": 16,
                  "n_discordant": 22,
                  "p_value": 0.052478790283203125
                },
                "ci95_diff_newcombe": [
                  -0.06860584179765444,
                  -0.0024060942614670816
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.06484641638225255,
                  "hi": -0.0034129692832764505
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned83k_q8__vs__frontier_gpt56sol": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 300,
                "contingency": {
                  "a_both_pass": 268,
                  "b_x_pass_y_fail": 13,
                  "c_x_fail_y_pass": 18,
                  "d_neither": 1
                },
                "x_rate_kn": "281/300 (93.7%)",
                "y_rate_kn": "286/300 (95.3%)",
                "diff_x_minus_y": -0.016666666666666666,
                "diff_pct_points": -1.6666666666666667,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 13,
                  "c": 18,
                  "n_discordant": 31,
                  "p_value": 0.47312965989112854
                },
                "ci95_diff_newcombe": [
                  -0.05486005211734017,
                  0.02075555052375612
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.05333333333333334,
                  "hi": 0.02
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "C  frontier gpt-5.6-sol (vendor-hosted)",
                "n_pairs": 293,
                "contingency": {
                  "a_both_pass": 268,
                  "b_x_pass_y_fail": 7,
                  "c_x_fail_y_pass": 18,
                  "d_neither": 0
                },
                "x_rate_kn": "275/293 (93.9%)",
                "y_rate_kn": "286/293 (97.6%)",
                "diff_x_minus_y": -0.03754266211604096,
                "diff_pct_points": -3.7542662116040955,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 7,
                  "c": 18,
                  "n_discordant": 25,
                  "p_value": 0.043285250663757324
                },
                "ci95_diff_newcombe": [
                  -0.07374828048013808,
                  -0.00374159703780242
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": -0.07167235494880546,
                  "hi": -0.006825938566552901
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            }
          }
        },
        "harness_bucket_counts_per_arm": {
          "base_q8": {
            "compile-fail": 296,
            "TRUNCATED": 3,
            "compile-pass": 1
          },
          "tuned83k_q8": {
            "compile-pass": 281,
            "compile-fail": 19
          },
          "tuned220k_q8": {
            "compile-pass": 282,
            "compile-fail": 18
          },
          "frontier_gpt56sol": {
            "compile-pass": 286,
            "compile-fail": 7,
            "TRUNCATED": 4,
            "INFRASTRUCTURE-FAILURE": 3
          }
        },
        "exclusion_counts_by_class_per_arm": {
          "scored": {
            "base_q8": 297,
            "tuned83k_q8": 300,
            "tuned220k_q8": 300,
            "frontier_gpt56sol": 293
          },
          "truncated": {
            "base_q8": 3,
            "frontier_gpt56sol": 4
          },
          "generation_error": {
            "frontier_gpt56sol": 3
          }
        },
        "exclusion_buckets_by_artifact_class": null,
        "denominator_pools": {
          "full": 300
        },
        "item_alignment": {
          "join_key": "ea_name",
          "per_arm_unique_names": {
            "base_q8": 300,
            "tuned83k_q8": 300,
            "tuned220k_q8": 300,
            "frontier_gpt56sol": 300
          },
          "intersection_n": 300,
          "per_arm_names_missing_from_intersection": {},
          "duplicate_ea_names_per_arm": {},
          "ALIGNED": true,
          "prompt_identity": {
            "items_with_identical_prompt_sha256_across_all_arms": 0,
            "items_disagreeing_with_the_majority_prompt_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 300
            },
            "ALL_ARMS_BYTE_IDENTICAL_PROMPTS": false,
            "note": "a differing prompt_sha256 means the arm was served a differently assembled prompt string; the head-to-head with that arm is confounded and is reported as such"
          },
          "spec_identity": {
            "items_with_identical_spec_sha256_across_all_arms": 300,
            "items_disagreeing_with_the_majority_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 0
            },
            "rows_missing_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0,
              "frontier_gpt56sol": 0
            },
            "ALL_ARMS_IDENTICAL_SPEC": true,
            "note": "spec_sha256 identity is the pairing proof (the same item was put to every arm); prompt_sha256 identity is the separate byte-identity requirement of the pre-registered protocol (see the top-level `protocol` field) and is reported beside it"
          },
          "manifest_selection_sha256_identical_across_arms": true
        },
        "instrument_validation": {
          "base_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 11,
              "row_errors_field": 11,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 11,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned83k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 7,
              "row_errors_field": 7,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 7,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned220k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 2,
              "row_errors_field": 2,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 2,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "frontier_gpt56sol": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 10,
              "row_errors_field": 10,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 10,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          }
        },
        "artifact_class_counts": null,
        "category_field": "The `category` key is present on every row of this set and its value is the empty string on 100% of rows, so no artifact-class breakdown is computable here and none was invented. The key exists; the value is empty.",
        "category_key_present_on_all_rows": true,
        "contaminated_items_excluded_in_clean_pool": [],
        "inclusive_scoring_anomalies": {
          "base_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 300,
            "inclusive_k_under_the_PREREGISTERED_rule": 1,
            "inclusive_k_if_the_row_verdict_were_honoured": 1,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned83k_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 300,
            "inclusive_k_under_the_PREREGISTERED_rule": 281,
            "inclusive_k_if_the_row_verdict_were_honoured": 281,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned220k_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 300,
            "inclusive_k_under_the_PREREGISTERED_rule": 282,
            "inclusive_k_if_the_row_verdict_were_honoured": 282,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "frontier_gpt56sol": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 300,
            "inclusive_k_under_the_PREREGISTERED_rule": 286,
            "inclusive_k_if_the_row_verdict_were_honoured": 286,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          }
        }
      },
      "nonea_c200": {
        "title": "Part 3 \u2014 balanced 200-item non-EA holdout, 50 per class (include / indicator / script / service)",
        "manifest_file_sha256": "db0e0614eed18143e41931353fd47eb4e86422f673b0b44e23ec2eebff090065",
        "per_arm": {
          "full": {
            "base_q8": {
              "inclusive": {
                "k": 56,
                "n": 200,
                "rate_pct": 28.0,
                "kn": "56/200 (28.0%)",
                "ci95": [
                  0.22236889773949026,
                  0.3459230458665071
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 48,
                  "n": 200,
                  "kn": "48/200 (24.0%)",
                  "ci95": [
                    0.1860661599241369,
                    0.30373340979204166
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 56,
                "n": 186,
                "rate_pct": 30.107526881720432,
                "kn": "56/186 (30.1%)",
                "ci95": [
                  0.23972277016842466,
                  0.3704782867572216
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 48,
                  "n": 186,
                  "kn": "48/186 (25.8%)",
                  "ci95": [
                    0.20052355007764888,
                    0.3253966542913804
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned83k_q8": {
              "inclusive": {
                "k": 135,
                "n": 200,
                "rate_pct": 67.5,
                "kn": "135/200 (67.5%)",
                "ci95": [
                  0.6073198511602996,
                  0.7360842846076571
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 126,
                  "n": 200,
                  "kn": "126/200 (63.0%)",
                  "ci95": [
                    0.5612262992405018,
                    0.6938739159014088
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 135,
                "n": 187,
                "rate_pct": 72.19251336898395,
                "kn": "135/187 (72.2%)",
                "ci95": [
                  0.653733203391701,
                  0.7811827761491125
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 126,
                  "n": 187,
                  "kn": "126/187 (67.4%)",
                  "ci95": [
                    0.6036913813439719,
                    0.736905470103653
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            },
            "tuned220k_q8": {
              "inclusive": {
                "k": 168,
                "n": 200,
                "rate_pct": 84.0,
                "kn": "168/200 (84.0%)",
                "ci95": [
                  0.7828592983793289,
                  0.8843258796841297
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 168,
                  "n": 200,
                  "kn": "168/200 (84.0%)",
                  "ci95": [
                    0.7828592983793289,
                    0.8843258796841297
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              },
              "strict": {
                "k": 168,
                "n": 188,
                "rate_pct": 89.36170212765957,
                "kn": "168/188 (89.4%)",
                "ci95": [
                  0.8413986236498567,
                  0.9300717400025025
                ],
                "ci_method": "Wilson score interval (no continuity correction)",
                "secondary_spec_conformance": {
                  "field": "verdict_strict",
                  "k": 168,
                  "n": 188,
                  "kn": "168/188 (89.4%)",
                  "ci95": [
                    0.8413986236498567,
                    0.9300717400025025
                  ],
                  "ci_method": "Wilson score interval",
                  "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                }
              }
            }
          }
        },
        "per_class": {
          "full": {
            "include": {
              "INDICATIVE": true,
              "cell_note": "Per-class cell, n=50, indicative. Cell counts are printed everywhere so nobody over-reads a k-of-n; a per-class CI is given but it settles nothing on its own.",
              "per_arm": {
                "base_q8": {
                  "inclusive": {
                    "k": 20,
                    "n": 50,
                    "rate_pct": 40.0,
                    "kn": "20/50 (40.0%)",
                    "ci95": [
                      0.27608389730256544,
                      0.5381856225241064
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 16,
                      "n": 50,
                      "kn": "16/50 (32.0%)",
                      "ci95": [
                        0.20758216260656845,
                        0.45810297308144077
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 20,
                    "n": 43,
                    "rate_pct": 46.51162790697674,
                    "kn": "20/43 (46.5%)",
                    "ci95": [
                      0.32511057704121815,
                      0.6108435952768602
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 16,
                      "n": 43,
                      "kn": "16/43 (37.2%)",
                      "ci95": [
                        0.24376339009650572,
                        0.5214019084031153
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned83k_q8": {
                  "inclusive": {
                    "k": 38,
                    "n": 50,
                    "rate_pct": 76.0,
                    "kn": "38/50 (76.0%)",
                    "ci95": [
                      0.6258731624205708,
                      0.8570260860300828
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 38,
                      "n": 50,
                      "kn": "38/50 (76.0%)",
                      "ci95": [
                        0.6258731624205708,
                        0.8570260860300828
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 38,
                    "n": 42,
                    "rate_pct": 90.47619047619048,
                    "kn": "38/42 (90.5%)",
                    "ci95": [
                      0.7793489419142897,
                      0.9623377511944261
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 38,
                      "n": 42,
                      "kn": "38/42 (90.5%)",
                      "ci95": [
                        0.7793489419142897,
                        0.9623377511944261
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned220k_q8": {
                  "inclusive": {
                    "k": 31,
                    "n": 50,
                    "rate_pct": 62.0,
                    "kn": "31/50 (62.0%)",
                    "ci95": [
                      0.4815044693099298,
                      0.7413721068980641
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 31,
                      "n": 50,
                      "kn": "31/50 (62.0%)",
                      "ci95": [
                        0.4815044693099298,
                        0.7413721068980641
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 31,
                    "n": 39,
                    "rate_pct": 79.48717948717949,
                    "kn": "31/39 (79.5%)",
                    "ci95": [
                      0.644660319864177,
                      0.8922028176928961
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 31,
                      "n": 39,
                      "kn": "31/39 (79.5%)",
                      "ci95": [
                        0.644660319864177,
                        0.8922028176928961
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                }
              },
              "paired": {
                "tuned83k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 19,
                      "b_x_pass_y_fail": 19,
                      "c_x_fail_y_pass": 1,
                      "d_neither": 11
                    },
                    "x_rate_kn": "38/50 (76.0%)",
                    "y_rate_kn": "20/50 (40.0%)",
                    "diff_x_minus_y": 0.36,
                    "diff_pct_points": 36.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 19,
                      "c": 1,
                      "n_discordant": 20,
                      "p_value": 4.00543212890625e-05
                    },
                    "ci95_diff_newcombe": [
                      0.20631088196171574,
                      0.4866281416931669
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.22,
                      "hi": 0.5
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 20,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 41,
                    "contingency": {
                      "a_both_pass": 19,
                      "b_x_pass_y_fail": 18,
                      "c_x_fail_y_pass": 1,
                      "d_neither": 3
                    },
                    "x_rate_kn": "37/41 (90.2%)",
                    "y_rate_kn": "20/41 (48.8%)",
                    "diff_x_minus_y": 0.4146341463414634,
                    "diff_pct_points": 41.46341463414634,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 18,
                      "c": 1,
                      "n_discordant": 19,
                      "p_value": 7.62939453125e-05
                    },
                    "ci95_diff_newcombe": [
                      0.2352526212275229,
                      0.5626130585283183
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.24390243902439024,
                      "hi": 0.5853658536585366
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 19,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 20,
                      "b_x_pass_y_fail": 11,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 19
                    },
                    "x_rate_kn": "31/50 (62.0%)",
                    "y_rate_kn": "20/50 (40.0%)",
                    "diff_x_minus_y": 0.22,
                    "diff_pct_points": 22.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 11,
                      "c": 0,
                      "n_discordant": 11,
                      "p_value": 0.0009765625
                    },
                    "ci95_diff_newcombe": [
                      0.10248668667863278,
                      0.3242051136202031
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.12,
                      "hi": 0.34
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 11,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 38,
                    "contingency": {
                      "a_both_pass": 20,
                      "b_x_pass_y_fail": 10,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 8
                    },
                    "x_rate_kn": "30/38 (78.9%)",
                    "y_rate_kn": "20/38 (52.6%)",
                    "diff_x_minus_y": 0.2631578947368421,
                    "diff_pct_points": 26.31578947368421,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 10,
                      "c": 0,
                      "n_discordant": 10,
                      "p_value": 0.001953125
                    },
                    "ci95_diff_newcombe": [
                      0.11904686012522916,
                      0.39311537251842865
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.13157894736842105,
                      "hi": 0.42105263157894735
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 10,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__tuned83k_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 26,
                      "b_x_pass_y_fail": 5,
                      "c_x_fail_y_pass": 12,
                      "d_neither": 7
                    },
                    "x_rate_kn": "31/50 (62.0%)",
                    "y_rate_kn": "38/50 (76.0%)",
                    "diff_x_minus_y": -0.14,
                    "diff_pct_points": -14.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 5,
                      "c": 12,
                      "n_discordant": 17,
                      "p_value": 0.143463134765625
                    },
                    "ci95_diff_newcombe": [
                      -0.2892260707501137,
                      0.018293044636248756
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.3,
                      "hi": 0.02
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 17,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 38,
                    "contingency": {
                      "a_both_pass": 26,
                      "b_x_pass_y_fail": 4,
                      "c_x_fail_y_pass": 8,
                      "d_neither": 0
                    },
                    "x_rate_kn": "30/38 (78.9%)",
                    "y_rate_kn": "34/38 (89.5%)",
                    "diff_x_minus_y": -0.10526315789473684,
                    "diff_pct_points": -10.526315789473685,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 4,
                      "c": 8,
                      "n_discordant": 12,
                      "p_value": 0.3876953125
                    },
                    "ci95_diff_newcombe": [
                      -0.28096575476226193,
                      0.07714223086316904
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.2894736842105263,
                      "hi": 0.07894736842105263
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 12,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                }
              },
              "exclusion_classes_within_this_class": {
                "base_q8": {
                  "scored": 43,
                  "truncated": 7
                },
                "tuned83k_q8": {
                  "scored": 42,
                  "truncated": 8
                },
                "tuned220k_q8": {
                  "scored": 39,
                  "truncated": 11
                }
              }
            },
            "indicator": {
              "INDICATIVE": true,
              "cell_note": "Per-class cell, n=50, indicative. Cell counts are printed everywhere so nobody over-reads a k-of-n; a per-class CI is given but it settles nothing on its own.",
              "per_arm": {
                "base_q8": {
                  "inclusive": {
                    "k": 0,
                    "n": 50,
                    "rate_pct": 0.0,
                    "kn": "0/50 (0.0%)",
                    "ci95": [
                      0.0,
                      0.07134759913335872
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 0,
                      "n": 50,
                      "kn": "0/50 (0.0%)",
                      "ci95": [
                        0.0,
                        0.07134759913335872
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 0,
                    "n": 45,
                    "rate_pct": 0.0,
                    "kn": "0/45 (0.0%)",
                    "ci95": [
                      0.0,
                      0.07865159873288836
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 0,
                      "n": 45,
                      "kn": "0/45 (0.0%)",
                      "ci95": [
                        0.0,
                        0.07865159873288836
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned83k_q8": {
                  "inclusive": {
                    "k": 49,
                    "n": 50,
                    "rate_pct": 98.0,
                    "kn": "49/50 (98.0%)",
                    "ci95": [
                      0.8950455641036219,
                      0.9964607407283539
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 48,
                      "n": 50,
                      "kn": "48/50 (96.0%)",
                      "ci95": [
                        0.8653990931249297,
                        0.9889611156723802
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 49,
                    "n": 50,
                    "rate_pct": 98.0,
                    "kn": "49/50 (98.0%)",
                    "ci95": [
                      0.8950455641036219,
                      0.9964607407283539
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 48,
                      "n": 50,
                      "kn": "48/50 (96.0%)",
                      "ci95": [
                        0.8653990931249297,
                        0.9889611156723802
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned220k_q8": {
                  "inclusive": {
                    "k": 50,
                    "n": 50,
                    "rate_pct": 100.0,
                    "kn": "50/50 (100.0%)",
                    "ci95": [
                      0.9286524008666414,
                      1.0
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 50,
                      "n": 50,
                      "kn": "50/50 (100.0%)",
                      "ci95": [
                        0.9286524008666414,
                        1.0
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 50,
                    "n": 50,
                    "rate_pct": 100.0,
                    "kn": "50/50 (100.0%)",
                    "ci95": [
                      0.9286524008666414,
                      1.0
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 50,
                      "n": 50,
                      "kn": "50/50 (100.0%)",
                      "ci95": [
                        0.9286524008666414,
                        1.0
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                }
              },
              "paired": {
                "tuned83k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 0,
                      "b_x_pass_y_fail": 49,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 1
                    },
                    "x_rate_kn": "49/50 (98.0%)",
                    "y_rate_kn": "0/50 (0.0%)",
                    "diff_x_minus_y": 0.98,
                    "diff_pct_points": 98.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 49,
                      "c": 0,
                      "n_discordant": 49,
                      "p_value": 3.552713678800501e-15
                    },
                    "ci95_diff_newcombe": [
                      0.8690597634734527,
                      0.9964607407283539
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.94,
                      "hi": 1.0
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 49,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 45,
                    "contingency": {
                      "a_both_pass": 0,
                      "b_x_pass_y_fail": 44,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 1
                    },
                    "x_rate_kn": "44/45 (97.8%)",
                    "y_rate_kn": "0/45 (0.0%)",
                    "diff_x_minus_y": 0.9777777777777777,
                    "diff_pct_points": 97.77777777777777,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 44,
                      "c": 0,
                      "n_discordant": 44,
                      "p_value": 1.1368683772161603e-13
                    },
                    "ci95_diff_newcombe": [
                      0.855638650532319,
                      0.9960664011171667
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.9333333333333333,
                      "hi": 1.0
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 44,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 0,
                      "b_x_pass_y_fail": 50,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 0
                    },
                    "x_rate_kn": "50/50 (100.0%)",
                    "y_rate_kn": "0/50 (0.0%)",
                    "diff_x_minus_y": 1.0,
                    "diff_pct_points": 100.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 50,
                      "c": 0,
                      "n_discordant": 50,
                      "p_value": 1.7763568394002505e-15
                    },
                    "ci95_diff_newcombe": [
                      0.8990992576628453,
                      1.0
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 1.0,
                      "hi": 1.0
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 50,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 45,
                    "contingency": {
                      "a_both_pass": 0,
                      "b_x_pass_y_fail": 45,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 0
                    },
                    "x_rate_kn": "45/45 (100.0%)",
                    "y_rate_kn": "0/45 (0.0%)",
                    "diff_x_minus_y": 1.0,
                    "diff_pct_points": 100.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 45,
                      "c": 0,
                      "n_discordant": 45,
                      "p_value": 5.684341886080802e-14
                    },
                    "ci95_diff_newcombe": [
                      0.8887698423696228,
                      1.0
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 1.0,
                      "hi": 1.0
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 45,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__tuned83k_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 49,
                      "b_x_pass_y_fail": 1,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 0
                    },
                    "x_rate_kn": "50/50 (100.0%)",
                    "y_rate_kn": "49/50 (98.0%)",
                    "diff_x_minus_y": 0.02,
                    "diff_pct_points": 2.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 1,
                      "c": 0,
                      "n_discordant": 1,
                      "p_value": 1.0
                    },
                    "ci95_diff_newcombe": [
                      -0.053221826578012416,
                      0.1049544358963781
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.0,
                      "hi": 0.06
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 1,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": true
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 49,
                      "b_x_pass_y_fail": 1,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 0
                    },
                    "x_rate_kn": "50/50 (100.0%)",
                    "y_rate_kn": "49/50 (98.0%)",
                    "diff_x_minus_y": 0.02,
                    "diff_pct_points": 2.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 1,
                      "c": 0,
                      "n_discordant": 1,
                      "p_value": 1.0
                    },
                    "ci95_diff_newcombe": [
                      -0.053221826578012416,
                      0.1049544358963781
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.0,
                      "hi": 0.06
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 1,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": true
                  }
                }
              },
              "exclusion_classes_within_this_class": {
                "base_q8": {
                  "scored": 45,
                  "truncated": 5
                },
                "tuned83k_q8": {
                  "scored": 50
                },
                "tuned220k_q8": {
                  "scored": 50
                }
              }
            },
            "script": {
              "INDICATIVE": true,
              "cell_note": "Per-class cell, n=50, indicative. Cell counts are printed everywhere so nobody over-reads a k-of-n; a per-class CI is given but it settles nothing on its own.",
              "per_arm": {
                "base_q8": {
                  "inclusive": {
                    "k": 13,
                    "n": 50,
                    "rate_pct": 26.0,
                    "kn": "13/50 (26.0%)",
                    "ci95": [
                      0.15871527493552393,
                      0.39553157264848837
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 12,
                      "n": 50,
                      "kn": "12/50 (24.0%)",
                      "ci95": [
                        0.1429739139699173,
                        0.3741268375794292
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 13,
                    "n": 48,
                    "rate_pct": 27.083333333333332,
                    "kn": "13/48 (27.1%)",
                    "ci95": [
                      0.1656595762745644,
                      0.40996964982240613
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 12,
                      "n": 48,
                      "kn": "12/48 (25.0%)",
                      "ci95": [
                        0.1492063367379791,
                        0.3878437280950796
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned83k_q8": {
                  "inclusive": {
                    "k": 19,
                    "n": 50,
                    "rate_pct": 38.0,
                    "kn": "19/50 (38.0%)",
                    "ci95": [
                      0.25862789310193596,
                      0.5184955306900703
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 17,
                      "n": 50,
                      "kn": "17/50 (34.0%)",
                      "ci95": [
                        0.2243694922250009,
                        0.478461739497674
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 19,
                    "n": 45,
                    "rate_pct": 42.22222222222222,
                    "kn": "19/45 (42.2%)",
                    "ci95": [
                      0.289687024672703,
                      0.5669921129079686
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 17,
                      "n": 45,
                      "kn": "17/45 (37.8%)",
                      "ci95": [
                        0.25108084672137915,
                        0.5237006551911048
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned220k_q8": {
                  "inclusive": {
                    "k": 39,
                    "n": 50,
                    "rate_pct": 78.0,
                    "kn": "39/50 (78.0%)",
                    "ci95": [
                      0.6475845041874634,
                      0.8724608402978559
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 39,
                      "n": 50,
                      "kn": "39/50 (78.0%)",
                      "ci95": [
                        0.6475845041874634,
                        0.8724608402978559
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 39,
                    "n": 49,
                    "rate_pct": 79.59183673469387,
                    "kn": "39/49 (79.6%)",
                    "ci95": [
                      0.6636299908048744,
                      0.8851815039421786
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 39,
                      "n": 49,
                      "kn": "39/49 (79.6%)",
                      "ci95": [
                        0.6636299908048744,
                        0.8851815039421786
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                }
              },
              "paired": {
                "tuned83k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 5,
                      "b_x_pass_y_fail": 14,
                      "c_x_fail_y_pass": 8,
                      "d_neither": 23
                    },
                    "x_rate_kn": "19/50 (38.0%)",
                    "y_rate_kn": "13/50 (26.0%)",
                    "diff_x_minus_y": 0.12,
                    "diff_pct_points": 12.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 14,
                      "c": 8,
                      "n_discordant": 22,
                      "p_value": 0.28627872467041016
                    },
                    "ci95_diff_newcombe": [
                      -0.0614237176657442,
                      0.29111832884188804
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.06,
                      "hi": 0.3
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 22,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 44,
                    "contingency": {
                      "a_both_pass": 5,
                      "b_x_pass_y_fail": 14,
                      "c_x_fail_y_pass": 8,
                      "d_neither": 17
                    },
                    "x_rate_kn": "19/44 (43.2%)",
                    "y_rate_kn": "13/44 (29.5%)",
                    "diff_x_minus_y": 0.13636363636363635,
                    "diff_pct_points": 13.636363636363637,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 14,
                      "c": 8,
                      "n_discordant": 22,
                      "p_value": 0.28627872467041016
                    },
                    "ci95_diff_newcombe": [
                      -0.06906839603660886,
                      0.3269501414557799
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.06818181818181818,
                      "hi": 0.3409090909090909
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 22,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 11,
                      "b_x_pass_y_fail": 28,
                      "c_x_fail_y_pass": 2,
                      "d_neither": 9
                    },
                    "x_rate_kn": "39/50 (78.0%)",
                    "y_rate_kn": "13/50 (26.0%)",
                    "diff_x_minus_y": 0.52,
                    "diff_pct_points": 52.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 28,
                      "c": 2,
                      "n_discordant": 30,
                      "p_value": 8.67992639541626e-07
                    },
                    "ci95_diff_newcombe": [
                      0.33970846660140913,
                      0.6505168266317711
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.36,
                      "hi": 0.68
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 30,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 47,
                    "contingency": {
                      "a_both_pass": 11,
                      "b_x_pass_y_fail": 26,
                      "c_x_fail_y_pass": 2,
                      "d_neither": 8
                    },
                    "x_rate_kn": "37/47 (78.7%)",
                    "y_rate_kn": "13/47 (27.7%)",
                    "diff_x_minus_y": 0.5106382978723404,
                    "diff_pct_points": 51.06382978723404,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 26,
                      "c": 2,
                      "n_discordant": 28,
                      "p_value": 3.032386302947998e-06
                    },
                    "ci95_diff_newcombe": [
                      0.32348040174335807,
                      0.6461002126173133
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.3404255319148936,
                      "hi": 0.6808510638297872
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 28,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__tuned83k_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 15,
                      "b_x_pass_y_fail": 24,
                      "c_x_fail_y_pass": 4,
                      "d_neither": 7
                    },
                    "x_rate_kn": "39/50 (78.0%)",
                    "y_rate_kn": "19/50 (38.0%)",
                    "diff_x_minus_y": 0.4,
                    "diff_pct_points": 40.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 24,
                      "c": 4,
                      "n_discordant": 28,
                      "p_value": 0.00017999112606048584
                    },
                    "ci95_diff_newcombe": [
                      0.2101100714093677,
                      0.5512558966727544
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.22,
                      "hi": 0.58
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 28,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 44,
                    "contingency": {
                      "a_both_pass": 15,
                      "b_x_pass_y_fail": 19,
                      "c_x_fail_y_pass": 4,
                      "d_neither": 6
                    },
                    "x_rate_kn": "34/44 (77.3%)",
                    "y_rate_kn": "19/44 (43.2%)",
                    "diff_x_minus_y": 0.3409090909090909,
                    "diff_pct_points": 34.09090909090909,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 19,
                      "c": 4,
                      "n_discordant": 23,
                      "p_value": 0.002599477767944336
                    },
                    "ci95_diff_newcombe": [
                      0.14041635147589288,
                      0.5054228250754605
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.1590909090909091,
                      "hi": 0.5227272727272727
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 23,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                }
              },
              "exclusion_classes_within_this_class": {
                "base_q8": {
                  "scored": 48,
                  "truncated": 2
                },
                "tuned83k_q8": {
                  "scored": 45,
                  "truncated": 5
                },
                "tuned220k_q8": {
                  "scored": 49,
                  "truncated": 1
                }
              }
            },
            "service": {
              "INDICATIVE": true,
              "cell_note": "Per-class cell, n=50, indicative. Cell counts are printed everywhere so nobody over-reads a k-of-n; a per-class CI is given but it settles nothing on its own.",
              "per_arm": {
                "base_q8": {
                  "inclusive": {
                    "k": 23,
                    "n": 50,
                    "rate_pct": 46.0,
                    "kn": "23/50 (46.0%)",
                    "ci95": [
                      0.3296965219222435,
                      0.5960112860084253
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 20,
                      "n": 50,
                      "kn": "20/50 (40.0%)",
                      "ci95": [
                        0.27608389730256544,
                        0.5381856225241064
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 23,
                    "n": 50,
                    "rate_pct": 46.0,
                    "kn": "23/50 (46.0%)",
                    "ci95": [
                      0.3296965219222435,
                      0.5960112860084253
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 20,
                      "n": 50,
                      "kn": "20/50 (40.0%)",
                      "ci95": [
                        0.27608389730256544,
                        0.5381856225241064
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned83k_q8": {
                  "inclusive": {
                    "k": 29,
                    "n": 50,
                    "rate_pct": 58.0,
                    "kn": "29/50 (58.0%)",
                    "ci95": [
                      0.4423344176857823,
                      0.7062499664528802
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 23,
                      "n": 50,
                      "kn": "23/50 (46.0%)",
                      "ci95": [
                        0.3296965219222435,
                        0.5960112860084253
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 29,
                    "n": 50,
                    "rate_pct": 58.0,
                    "kn": "29/50 (58.0%)",
                    "ci95": [
                      0.4423344176857823,
                      0.7062499664528802
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 23,
                      "n": 50,
                      "kn": "23/50 (46.0%)",
                      "ci95": [
                        0.3296965219222435,
                        0.5960112860084253
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                },
                "tuned220k_q8": {
                  "inclusive": {
                    "k": 48,
                    "n": 50,
                    "rate_pct": 96.0,
                    "kn": "48/50 (96.0%)",
                    "ci95": [
                      0.8653990931249297,
                      0.9889611156723802
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 48,
                      "n": 50,
                      "kn": "48/50 (96.0%)",
                      "ci95": [
                        0.8653990931249297,
                        0.9889611156723802
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  },
                  "strict": {
                    "k": 48,
                    "n": 50,
                    "rate_pct": 96.0,
                    "kn": "48/50 (96.0%)",
                    "ci95": [
                      0.8653990931249297,
                      0.9889611156723802
                    ],
                    "ci_method": "Wilson score interval (no continuity correction)",
                    "secondary_spec_conformance": {
                      "field": "verdict_strict",
                      "k": 48,
                      "n": 50,
                      "kn": "48/50 (96.0%)",
                      "ci95": [
                        0.8653990931249297,
                        0.9889611156723802
                      ],
                      "ci_method": "Wilson score interval",
                      "FLAG": "self-declared, non-independent; see statistics_v1_1.secondary_spec_conformance_flag"
                    }
                  }
                }
              },
              "paired": {
                "tuned83k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 13,
                      "b_x_pass_y_fail": 16,
                      "c_x_fail_y_pass": 10,
                      "d_neither": 11
                    },
                    "x_rate_kn": "29/50 (58.0%)",
                    "y_rate_kn": "23/50 (46.0%)",
                    "diff_x_minus_y": 0.12,
                    "diff_pct_points": 12.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 16,
                      "c": 10,
                      "n_discordant": 26,
                      "p_value": 0.32693958282470703
                    },
                    "ci95_diff_newcombe": [
                      -0.07617870358394077,
                      0.3039227266392255
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.08,
                      "hi": 0.32
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 26,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B1 tuned 83k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 13,
                      "b_x_pass_y_fail": 16,
                      "c_x_fail_y_pass": 10,
                      "d_neither": 11
                    },
                    "x_rate_kn": "29/50 (58.0%)",
                    "y_rate_kn": "23/50 (46.0%)",
                    "diff_x_minus_y": 0.12,
                    "diff_pct_points": 12.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 16,
                      "c": 10,
                      "n_discordant": 26,
                      "p_value": 0.32693958282470703
                    },
                    "ci95_diff_newcombe": [
                      -0.07617870358394077,
                      0.3039227266392255
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": -0.08,
                      "hi": 0.32
                    },
                    "dominance": "NO dominance - each arm fails items the other passes, in both directions",
                    "n_discordant": 26,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__base_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 23,
                      "b_x_pass_y_fail": 25,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 2
                    },
                    "x_rate_kn": "48/50 (96.0%)",
                    "y_rate_kn": "23/50 (46.0%)",
                    "diff_x_minus_y": 0.5,
                    "diff_pct_points": 50.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 25,
                      "c": 0,
                      "n_discordant": 25,
                      "p_value": 5.960464477539063e-08
                    },
                    "ci95_diff_newcombe": [
                      0.3496662223470193,
                      0.6280461412539109
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.36,
                      "hi": 0.64
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 25,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 23,
                      "b_x_pass_y_fail": 25,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 2
                    },
                    "x_rate_kn": "48/50 (96.0%)",
                    "y_rate_kn": "23/50 (46.0%)",
                    "diff_x_minus_y": 0.5,
                    "diff_pct_points": 50.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 25,
                      "c": 0,
                      "n_discordant": 25,
                      "p_value": 5.960464477539063e-08
                    },
                    "ci95_diff_newcombe": [
                      0.3496662223470193,
                      0.6280461412539109
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.36,
                      "hi": 0.64
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 25,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                },
                "tuned220k_q8__vs__tuned83k_q8": {
                  "inclusive": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 29,
                      "b_x_pass_y_fail": 19,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 2
                    },
                    "x_rate_kn": "48/50 (96.0%)",
                    "y_rate_kn": "29/50 (58.0%)",
                    "diff_x_minus_y": 0.38,
                    "diff_pct_points": 38.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 19,
                      "c": 0,
                      "n_discordant": 19,
                      "p_value": 3.814697265625e-06
                    },
                    "ci95_diff_newcombe": [
                      0.24158549713890842,
                      0.5137079862933835
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.24,
                      "hi": 0.52
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 19,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  },
                  "strict": {
                    "arm_x": "B2 tuned 220k Q8_0",
                    "arm_y": "B1 tuned 83k Q8_0",
                    "n_pairs": 50,
                    "contingency": {
                      "a_both_pass": 29,
                      "b_x_pass_y_fail": 19,
                      "c_x_fail_y_pass": 0,
                      "d_neither": 2
                    },
                    "x_rate_kn": "48/50 (96.0%)",
                    "y_rate_kn": "29/50 (58.0%)",
                    "diff_x_minus_y": 0.38,
                    "diff_pct_points": 38.0,
                    "mcnemar": {
                      "test": "McNemar exact (binomial, two-sided)",
                      "b": 19,
                      "c": 0,
                      "n_discordant": 19,
                      "p_value": 3.814697265625e-06
                    },
                    "ci95_diff_newcombe": [
                      0.24158549713890842,
                      0.5137079862933835
                    ],
                    "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                    "ci95_diff_bootstrap": {
                      "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                      "resamples": 20000,
                      "seed": 20260911,
                      "lo": 0.24,
                      "hi": 0.52
                    },
                    "dominance": "X strictly dominates Y (c=0)",
                    "n_discordant": 19,
                    "COMPUTABLE": true,
                    "DISCORDANCE_THIN": false
                  }
                }
              },
              "exclusion_classes_within_this_class": {
                "base_q8": {
                  "scored": 50
                },
                "tuned83k_q8": {
                  "scored": 50
                },
                "tuned220k_q8": {
                  "scored": 50
                }
              }
            }
          }
        },
        "paired": {
          "full": {
            "tuned83k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 200,
                "contingency": {
                  "a_both_pass": 37,
                  "b_x_pass_y_fail": 98,
                  "c_x_fail_y_pass": 19,
                  "d_neither": 46
                },
                "x_rate_kn": "135/200 (67.5%)",
                "y_rate_kn": "56/200 (28.0%)",
                "diff_x_minus_y": 0.395,
                "diff_pct_points": 39.5,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 98,
                  "c": 19,
                  "n_discordant": 117,
                  "p_value": 5.139554351717117e-14
                },
                "ci95_diff_newcombe": [
                  0.2996261196608533,
                  0.4797735297607776
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.305,
                  "hi": 0.485
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B1 tuned 83k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 180,
                "contingency": {
                  "a_both_pass": 37,
                  "b_x_pass_y_fail": 92,
                  "c_x_fail_y_pass": 19,
                  "d_neither": 32
                },
                "x_rate_kn": "129/180 (71.7%)",
                "y_rate_kn": "56/180 (31.1%)",
                "diff_x_minus_y": 0.40555555555555556,
                "diff_pct_points": 40.55555555555556,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 92,
                  "c": 19,
                  "n_discordant": 111,
                  "p_value": 1.1236121882237838e-12
                },
                "ci95_diff_newcombe": [
                  0.30191683497659605,
                  0.4967304007668792
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.3055555555555556,
                  "hi": 0.5055555555555555
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__base_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 200,
                "contingency": {
                  "a_both_pass": 54,
                  "b_x_pass_y_fail": 114,
                  "c_x_fail_y_pass": 2,
                  "d_neither": 30
                },
                "x_rate_kn": "168/200 (84.0%)",
                "y_rate_kn": "56/200 (28.0%)",
                "diff_x_minus_y": 0.56,
                "diff_pct_points": 56.0,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 114,
                  "c": 2,
                  "n_discordant": 116,
                  "p_value": 1.6339108165695212e-31
                },
                "ci95_diff_newcombe": [
                  0.4824232034252644,
                  0.6248536127419356
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.49,
                  "hi": 0.63
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "A  base Qwen2.5-Coder-14B-Instruct Q8_0",
                "n_pairs": 180,
                "contingency": {
                  "a_both_pass": 54,
                  "b_x_pass_y_fail": 106,
                  "c_x_fail_y_pass": 2,
                  "d_neither": 18
                },
                "x_rate_kn": "160/180 (88.9%)",
                "y_rate_kn": "56/180 (31.1%)",
                "diff_x_minus_y": 0.5777777777777777,
                "diff_pct_points": 57.77777777777778,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 106,
                  "c": 2,
                  "n_discordant": 108,
                  "p_value": 3.6281438664344504e-29
                },
                "ci95_diff_newcombe": [
                  0.4956667524788013,
                  0.6459876482181227
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.5,
                  "hi": 0.65
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            },
            "tuned220k_q8__vs__tuned83k_q8": {
              "inclusive": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 200,
                "contingency": {
                  "a_both_pass": 119,
                  "b_x_pass_y_fail": 49,
                  "c_x_fail_y_pass": 16,
                  "d_neither": 16
                },
                "x_rate_kn": "168/200 (84.0%)",
                "y_rate_kn": "135/200 (67.5%)",
                "diff_x_minus_y": 0.165,
                "diff_pct_points": 16.5,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 49,
                  "c": 16,
                  "n_discordant": 65,
                  "p_value": 5.084767048790177e-05
                },
                "ci95_diff_newcombe": [
                  0.0884623286274878,
                  0.2396122831867154
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.09,
                  "hi": 0.24
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              },
              "strict": {
                "arm_x": "B2 tuned 220k Q8_0",
                "arm_y": "B1 tuned 83k Q8_0",
                "n_pairs": 182,
                "contingency": {
                  "a_both_pass": 119,
                  "b_x_pass_y_fail": 43,
                  "c_x_fail_y_pass": 12,
                  "d_neither": 8
                },
                "x_rate_kn": "162/182 (89.0%)",
                "y_rate_kn": "131/182 (72.0%)",
                "diff_x_minus_y": 0.17032967032967034,
                "diff_pct_points": 17.032967032967033,
                "mcnemar": {
                  "test": "McNemar exact (binomial, two-sided)",
                  "b": 43,
                  "c": 12,
                  "n_discordant": 55,
                  "p_value": 3.3053764732060564e-05
                },
                "ci95_diff_newcombe": [
                  0.09348723933106229,
                  0.24599582265139153
                ],
                "ci95_diff_newcombe_method": "Newcombe (1998) method 10 for the difference of two paired proportions (Wilson intervals + phi)",
                "ci95_diff_bootstrap": {
                  "method": "paired bootstrap over items (percentile, 2.5/97.5)",
                  "resamples": 20000,
                  "seed": 20260911,
                  "lo": 0.09340659340659341,
                  "hi": 0.24725274725274726
                },
                "dominance": "NO dominance - each arm fails items the other passes, in both directions"
              }
            }
          }
        },
        "harness_bucket_counts_per_arm": {
          "base_q8": {
            "compile-fail": 130,
            "compile-pass": 56,
            "TRUNCATED": 14
          },
          "tuned83k_q8": {
            "compile-pass": 135,
            "compile-fail": 52,
            "TRUNCATED": 13
          },
          "tuned220k_q8": {
            "compile-pass": 168,
            "compile-fail": 20,
            "TRUNCATED": 12
          }
        },
        "exclusion_counts_by_class_per_arm": {
          "scored": {
            "base_q8": 186,
            "tuned83k_q8": 187,
            "tuned220k_q8": 188
          },
          "truncated": {
            "base_q8": 14,
            "tuned83k_q8": 13,
            "tuned220k_q8": 12
          }
        },
        "exclusion_buckets_by_artifact_class": {
          "base_q8": {
            "truncated": {
              "indicator": 5,
              "include": 7,
              "script": 2
            }
          },
          "tuned83k_q8": {
            "truncated": {
              "include": 8,
              "script": 5
            }
          },
          "tuned220k_q8": {
            "truncated": {
              "include": 11,
              "script": 1
            }
          }
        },
        "denominator_pools": {
          "full": 200
        },
        "item_alignment": {
          "join_key": "ea_name",
          "per_arm_unique_names": {
            "base_q8": 200,
            "tuned83k_q8": 200,
            "tuned220k_q8": 200
          },
          "intersection_n": 200,
          "per_arm_names_missing_from_intersection": {},
          "duplicate_ea_names_per_arm": {},
          "ALIGNED": true,
          "prompt_identity": {
            "items_with_identical_prompt_sha256_across_all_arms": 200,
            "items_disagreeing_with_the_majority_prompt_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0
            },
            "ALL_ARMS_BYTE_IDENTICAL_PROMPTS": true,
            "note": "a differing prompt_sha256 means the arm was served a differently assembled prompt string; the head-to-head with that arm is confounded and is reported as such"
          },
          "spec_identity": {
            "items_with_identical_spec_sha256_across_all_arms": 200,
            "items_disagreeing_with_the_majority_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0
            },
            "rows_missing_spec_sha256_per_arm": {
              "base_q8": 0,
              "tuned83k_q8": 0,
              "tuned220k_q8": 0
            },
            "ALL_ARMS_IDENTICAL_SPEC": true,
            "note": "spec_sha256 identity is the pairing proof (the same item was put to every arm); prompt_sha256 identity is the separate byte-identity requirement of the pre-registered protocol (see the top-level `protocol` field) and is reported beside it"
          },
          "manifest_selection_sha256_identical_across_arms": true
        },
        "instrument_validation": {
          "base_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 2,
              "row_errors_field": 2,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 2,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned83k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 3,
              "row_errors_field": 3,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 3,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          },
          "tuned220k_q8": {
            "known_positive": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 12,
              "row_errors_field": 12,
              "PASS": true
            },
            "negative_control": {
              "decoded_as": "utf-8/ascii",
              "n_error_lines_found": 0,
              "result_line": "Result: 0 errors, 0 warnings",
              "PASS": true
            },
            "utf16_trap": {
              "raw_is_utf16_le_bom": true,
              "byte_grep_ascii_hits_in_raw": 0,
              "decoded_hits": 12,
              "meaning": "a GNU-grep-style byte match on the .raw.log returns 0 while the decoded log yields the true count; the taxonomy is built on the decoded text"
            }
          }
        },
        "artifact_class_counts": {
          "indicator": 50,
          "include": 50,
          "script": 50,
          "service": 50
        },
        "category_field": "category values observed: ['include', 'indicator', 'script', 'service']",
        "category_key_present_on_all_rows": true,
        "contaminated_items_excluded_in_clean_pool": [],
        "inclusive_scoring_anomalies": {
          "base_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 200,
            "inclusive_k_under_the_PREREGISTERED_rule": 56,
            "inclusive_k_if_the_row_verdict_were_honoured": 56,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned83k_q8": {
            "n_anomalous_rows": 0,
            "rows": [],
            "n_denominator_inclusive": 200,
            "inclusive_k_under_the_PREREGISTERED_rule": 135,
            "inclusive_k_if_the_row_verdict_were_honoured": 135,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          },
          "tuned220k_q8": {
            "n_anomalous_rows": 1,
            "rows": [
              {
                "category": "include",
                "bucket": "TRUNCATED",
                "truncated": true,
                "finish_reason": "length",
                "errors": 0,
                "ex5_exists": true,
                "verdict_headline": true,
                "verdict_strict": true
              }
            ],
            "n_denominator_inclusive": 200,
            "inclusive_k_under_the_PREREGISTERED_rule": 168,
            "inclusive_k_if_the_row_verdict_were_honoured": 169,
            "rule_applied": "an item in an excluded class is scored as a failure under the inclusive denominator (the wording of the pre-registered protocol; see the top-level `protocol` field)"
          }
        },
        "NO_FRONTIER_ARM": "No frontier arm exists on this set: gpt-5.6-sol was never run against the 200 non-EA holdout."
      }
    },
    "replay_c200": {
      "title": "Stability check - the 220k arm re-run on the same 200 non-EA prompts under the same configuration",
      "derived_from": "the published rows in holdout_nonea200_results.jsonl, arms tuned220k and tuned220k_q8_replay",
      "pre_registered_rule": "fixed before the second run started: the first run is the headline regardless of outcome, both runs publish, neither is averaged and neither is selected",
      "first_run": {
        "arm": "tuned220k",
        "passes_all_200": 168,
        "raw_verdict_headline_true": 169,
        "truncated": 12,
        "per_class": {
          "include": {
            "n": 50,
            "passes": 31,
            "truncated": 11
          },
          "indicator": {
            "n": 50,
            "passes": 50,
            "truncated": 0
          },
          "script": {
            "n": 50,
            "passes": 39,
            "truncated": 1
          },
          "service": {
            "n": 50,
            "passes": 48,
            "truncated": 0
          }
        }
      },
      "second_run": {
        "arm": "tuned220k_q8_replay",
        "passes_all_200": 169,
        "raw_verdict_headline_true": 169,
        "truncated": 11,
        "per_class": {
          "include": {
            "n": 50,
            "passes": 31,
            "truncated": 10
          },
          "indicator": {
            "n": 50,
            "passes": 50,
            "truncated": 0
          },
          "script": {
            "n": 50,
            "passes": 40,
            "truncated": 1
          },
          "service": {
            "n": 50,
            "passes": 48,
            "truncated": 0
          }
        }
      },
      "verdict_flips_raw": 2,
      "verdict_flips_under_the_scoring_rule": 1,
      "generated_source_byte_identical_items": 164,
      "n_items": 200,
      "meaning": "Both runs used the same recorded settings; this release does not isolate the cause of the differences. On a row where generation produced source, the output hash is provenance of the observed run, not a promise that a re-generation reproduces it."
    }
  }
}