Restore verified peer comparison, efficiency context, and Ollama support

Modelfile

@@ -0,0 +1,16 @@
+FROM ./coneml-810m-alpha-Q8_0.gguf
+
+TEMPLATE """{{- range .Messages }}
+{{- if eq .Role "system" }}System:
+{{ .Content }}
+{{- else if eq .Role "user" }}User:
+{{ .Content }}
+{{- else if eq .Role "assistant" }}Assistant:
+{{ .Content }}<|endoftext|>
+{{- end }}
+{{- end }}Assistant:
+"""
+
+PARAMETER num_ctx 8192
+PARAMETER repeat_penalty 1.15
+PARAMETER stop "<|endoftext|>"

README.md

@@ -65,7 +65,47 @@ Behavioral screens:
 - Everyday-reasoning screen: 11/20 by manual adjudication; the
   phrase-sensitive automatic scorer recorded 6/20.
 
-## 2. Standard benchmark limits
+## 2. Peer comparison and efficiency
+
+The pair was tested against five post-trained peers on the same reduced
+diagnostic screen, through each model's native instruction interface,
+with greedy decoding, repetition penalty 1.15, and matched short-answer
+budgets.
+
+| model | parameters | mixed arithmetic (n=585) | executed functions (n=100) | designated refusals (n=17) |
+|---|---:|---:|---:|---:|
+| ConeML Alpha | 0.81B | 421 (72.0%) | 83 | **13** |
+| ConeML Arithmetic | 0.81B | **442 (75.6%)** | 35 | 11 |
+| Qwen3.5 | 0.8B | 154 (26.3%) | 93 | 0 |
+| Qwen3 | 0.6B | 223 (38.1%) | **98** | 1 |
+| Llama 3.2 Instruct | 1.24B | 343 (58.6%) | 79 | 1 |
+| TinyLlama Chat | 1.1B | 113 (19.3%) | 51 | 0 |
+| SmolLM2 Instruct | 1.7B | 380 (65.0%) | 96 | 1 |
+
+This is not a general leaderboard: the task families match ConeML's
+trained surfaces. On this screen, the ConeML pair led mixed arithmetic;
+Alpha exceeded Llama 3.2 and TinyLlama on function writing but trailed
+Qwen3.5, Qwen3, and SmolLM2. All models recorded zero over-refusals on
+five in-scope contrasts.
+
+Qwen3.5 with thinking enabled reached 451/585 (77.1%), compared with
+ConeML Arithmetic's 442/585 (75.6%), while using at least 17.6 times the
+generated-token budget per item and approximately 24 times the recorded
+wall time. Runtime ratios are implementation-specific.
+
+ConeML's selected base consumed 12.32B token positions—15.2 tokens per
+parameter and approximately 5.99e19 training FLOPs under the `6ND`
+convention. Pretraining took approximately 11 days on one RTX 5090.
+Assuming, rather than claiming to have metered, a 0.70 kW average wall
+draw gives 184.8 kWh; at CHF 0.14–0.30/kWh, that is approximately
+CHF 26–55 of marginal pretraining electricity. This excludes hardware,
+labor, SFT, evaluation, conversion, and emissions accounting.
+
+The complete category table, peer wins, adverse base result, evaluated
+revisions, assumptions, and frozen evidence commitments are in
+[`eval/PEER_COMPARISON.md`](eval/PEER_COMPARISON.md).
+
+## 3. Standard benchmark limits
 
 Zero-shot, greedy decoding with a 256-token generation cap. GSM8K uses
 strict final-number matching; HumanEval reports pass@1 from executed
@@ -80,7 +120,7 @@ These results are weak and define an important boundary: the narrower
 task-family results above do not transfer to broad GSM8K or HumanEval
 performance.
 
-## 3. Supported prompt formats
+## 4. Supported prompt formats
 
 Evaluated decoding is greedy with repetition penalty 1.15, as pinned in
 `generation_config.json`.
@@ -100,7 +140,7 @@ Dialogue generation ends at `<|endoftext|>`. Prompting outside these
 formats, including paraphrases, was not systematically evaluated except
 for the published raw-completion spot check.
 
-## 4. Quantized variants
+## 5. Quantized variants
 
 Q8_0 GGUF was evaluated separately under llama.cpp. That runtime and its
 stop handling differ from the bf16 harness, so the scores are not directly
@@ -116,7 +156,23 @@ comparable and differences cannot be attributed solely to quantization.
 
 Use bf16 when the bf16 profile is required.
 
-## 5. Loading
+For local GGUF inference:
+
+```bash
+llama-cli \
+  -m coneml-810m-alpha-Q8_0.gguf \
+  -p $'Question: What is 7 + 8?\nAnswer:' \
+  -n 32 --repeat-penalty 1.15 --no-conversation --single-turn
+```
+
+For Ollama, download `Modelfile` beside the Q8_0 file, then run:
+
+```bash
+ollama create coneml-810m-alpha -f Modelfile
+ollama run coneml-810m-alpha "What is 7 + 8?"
+```
+
+## 6. Loading with Transformers
 
 ```python
 import torch
@@ -148,7 +204,7 @@ print(answer)
 This is a text-only causal language model. Load it with
 `AutoModelForCausalLM` or `LlamaForCausalLM`, not a multimodal model class.
 
-## 6. Intended use and safety
+## 7. Intended use and safety
 
 This release is intended for research, constrained local inference,
 evaluation-harness development, and experiments whose prompts match the
@@ -159,7 +215,7 @@ financial decisions, autonomous action, or unsandboxed execution of
 generated code. Generated code may be incomplete or unsafe; inspect and
 test it in an isolated environment.
 
-## 7. Known limitations
+## 8. Known limitations
 
 - Not a general code model: complete programs scored 0/10, executable SQL
   1/10, and bash 1/5. The 69% result is for single functions on ConeML's
@@ -173,7 +229,7 @@ test it in an isolated environment.
 - Multi-step GSM-class reasoning remained weak.
 - Frameless prompting was spot-checked, not systematically certified.
 
-## 8. Evidence and reproducibility
+## 9. Evidence and reproducibility
 
 The `eval/` directory publishes:
 
@@ -184,6 +240,9 @@ The `eval/` directory publishes:
 - separate Q8_0 aggregate results; and
 - `PRIVATE_EVIDENCE_SHA256SUMS.txt`: commitments to the complete
   row-level evidence retained by ConeML.
+- `peer-comparison-summary.json`, `PEER_COMPARISON.md`, and
+  `PEER_EVIDENCE_SHA256SUMS.txt`: the peer results, interpretation, and
+  commitments to the retained peer-generation evidence.
 
 Complete proprietary probe rows and generations are retained privately to
 avoid releasing the full measurement instruments. The hashes permit later
@@ -191,7 +250,7 @@ disclosure to be checked against the evidence frozen at release time.
 Public GSM8K and HumanEval rows are not redistributed; only aggregate
 results are published. `SHA256SUMS.txt` covers every shipped file.
 
-## 9. License and commercial enquiries
+## 10. License and commercial enquiries
 
 Released for non-commercial use under CC BY-NC 4.0. Commercial use is not
 granted by this release. See `LICENSE.md`.

SHA256SUMS.txt

@@ -1,14 +1,18 @@
 846a8e2297580b3c8d57f69a07d88833d370f976010b31d9fc9b93e9a3f04464  LICENSE.md
-c16d4402a3123ce06c2684cebc34294c25e1b9463799de8020cd985e2d937c5a  README.md
+dae2b5a811b8d4fd08801460bc80968d8792391bec75a235b6e79d0ba2fd952c  Modelfile
+86ad0bc96b23050d854c8c1ef6353aa1ecc36581f21038b8e418a54de3c37483  README.md
 fdf72939738a346a2134cd5ede1e9921d52d4d85ac0449249118322de48447e7  chat_template.jinja
 a09065a6d6c5ee5c518ed913085e878e7d3807f7d6cb8bcd1d6ce59653bd41b5  coneml-810m-alpha-Q8_0.gguf
 16835ec01c1fa344c5bd3788c7697ab67e68676ebae3fc374b62f7b30749a6bb  coneml-810m-alpha-f16.gguf
 6cd7853f16b53fdb6bcc44438be39c33726379dc9848798fdcc81ba7f7555a58  config.json
 8a56ef50eebf7604d0408b52249591e258a31ff1c9ebf3785f748c0316637c56  conversion.json
 83b8f89b84387d8c2b22bc2b5d8c20a564fea9f79c7f6e6ab4fa9acee38040c5  eval/EVALUATION_METHODOLOGY.md
+3817126000edea9bc2fcb75a783fa73de234590cdfe89e1aea9b569bae809caa  eval/PEER_COMPARISON.md
+73d8100a4b78d7693aabb7a22f3b9948774f7a418d9fa162d19d6ff630baffd4  eval/PEER_EVIDENCE_SHA256SUMS.txt
 b2414d6dddbac7128724d27a7ce54e2fa3586943ece7b94e6c3c5188a1f04484  eval/PRIVATE_EVIDENCE_SHA256SUMS.txt
 0108c09e9b835a0da81c5093b40f0b7937b1553de3abdcb5c207c75888714db2  eval/certification-Q8_0.json
 b4cc4f98ccd8ba91206e4446c38f6b1e5f6e7a07f3c9ec963ee16d6a0619dc71  eval/code-cert-Q8_0.json
+baff252791bdf43653763a53a57024d0075bb9a1f6b35cf1fcc94a9191cc73df  eval/peer-comparison-summary.json
 bff2af0c69ec8cce87cd93a69c22045b4062588e60f1dafb4ba0b85ac68f2c8d  eval/representative-samples.json
 f6ce9e02abad733c0d5ae9fdaa58f290441a24c4f24c7673c4f50ca5150e1d70  eval/summary.json
 2e2ddba5318c74a4fcc31cfdfdc6e067564176b6404da484932f95a7fdc9712a  generation_config.json

eval/PEER_COMPARISON.md

@@ -0,0 +1,128 @@
+# ConeML 810M peer diagnostic and resource context
+
+Date: 2026-07-30
+
+This is an interface-specific diagnostic screen, not a neutral ranking of
+general model capability. Each instruction-tuned model was evaluated
+through its own instruction interface. All models used greedy decoding
+with repetition penalty 1.15. Arithmetic and function-writing used
+matched short-answer generation budgets.
+
+The task families match ConeML's supervised training surfaces. That gives
+the comparison practical value for locating the release's envelope, but
+it also means these results must not be generalized to unrelated tasks.
+Standard GSM8K and HumanEval results remain separately disclosed in the
+model cards.
+
+## Instruct-model results
+
+| measured surface | ConeML Alpha | ConeML Arithmetic | Qwen3.5 0.8B | Qwen3 0.6B | Llama 3.2 1B Instruct | TinyLlama Chat | SmolLM2 1.7B Instruct |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| mixed arithmetic screen (n=585) | 421 (72.0%) | **442 (75.6%)** | 154 (26.3%) | 223 (38.1%) | 343 (58.6%) | 113 (19.3%) | 380 (65.0%) |
+| four core arithmetic lanes (n=225) | 177 (78.7%) | **221 (98.2%)** | 150 (66.7%) | 208 (92.4%) | **221 (98.2%)** | 92 (40.9%) | **221 (98.2%)** |
+| executed single functions (n=100) | 83 | 35 | 93 | **98** | 79 | 51 | 96 |
+| designated-refusal prompts (n=17) | **13** | 11 | 0 | 1 | 1 | 0 | 1 |
+| over-refusals on contrasts (n=5) | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
+
+The honest result is mixed:
+
+- At matched short-answer budgets, the ConeML pair led this post-trained
+  peer group on the broader arithmetic screen.
+- On the four core arithmetic lanes, the Arithmetic variant tied Llama
+  3.2 Instruct and SmolLM2 at this reduced sample size. Its separate
+  full-size internal result is 1,093/1,116 (97.9%).
+- Alpha exceeded Llama 3.2 Instruct and TinyLlama Chat on the executed
+  function screen, but Qwen3.5, Qwen3, and SmolLM2 scored higher.
+- The refusal row measures a trained response policy on designated
+  prompts, not factual correctness or general epistemic calibration.
+
+The peer-harness refusal results above differ from the dedicated ConeML
+probe's 17/17 for each release because the generation loops differ. Both
+measurements are reported rather than merged.
+
+## Qwen3.5 thinking-mode sensitivity
+
+Qwen3.5 0.8B was also evaluated with thinking enabled on the same 585
+arithmetic items:
+
+| metric | Qwen3.5 thinking | Qwen3.5 short-answer | ConeML Arithmetic short-answer |
+|---|---:|---:|---:|
+| accuracy | 451/585 (77.1%) | 154/585 (26.3%) | 442/585 (75.6%) |
+| generated tokens per item | mean 844; median 699 | at most 48 | at most 48 |
+| p95 / generation cap | 1,536 / 1,536 | 48 / 48 | 48 / 48 |
+| truncation rate | 138/585 (23.6%) | approximately 0% | approximately 0% |
+| mean wall time per item | 2.43 s | approximately 0.1 s | approximately 0.1 s |
+| correct answers per 1,000 generated tokens | 0.91 | 5.5 | approximately 15.7 or higher |
+
+Thinking mode recovered Qwen3.5 to statistical parity with ConeML
+Arithmetic on this screen, while using at least 17.6 times the
+per-answer generation budget and approximately 24 times the measured wall
+time. Wall-time ratios are specific to the recorded hardware, batching,
+and implementation. This is an inference-cost comparison, not a claim
+that thinking mode is intrinsically inferior.
+
+## Against-interest base result
+
+On the same 585 arithmetic items, through a task frame native to neither
+base model, Qwen3.5 0.8B Base scored 492/585 (84.1%), while ConeML base
+checkpoint 188 scored 156/585 (26.7%). This result is included because it
+prevents an absolute-superiority reading: high-exposure base models may
+already contain strong task-formatted behavior, and post-training can
+move capability between output surfaces.
+
+## Training-resource context
+
+ConeML's selected base consumed approximately 12.32B token positions:
+15.2 tokens per parameter and approximately 5.99e19 training FLOPs under
+the `6 × parameters × tokens` convention.
+
+| model family | disclosed pretraining tokens | approximate tokens/parameter | approximate training FLOPs vs ConeML |
+|---|---:|---:|---:|
+| ConeML 810M | 12.32B | 15.2 | 1× |
+| TinyLlama 1.1B | 3T | 2,727 | 330× |
+| Llama 3.2 1B | up to 9T | 7,258 | 1,118×, plus distillation |
+| SmolLM2 1.7B | 11T | 6,471 | 1,873× |
+| Qwen3 0.6B | 36T | 60,000 | 2,164× |
+| Qwen3.5 0.8B | not disclosed | not stated | not stated |
+
+These are estimated pretraining FLOP ratios, not historical electricity
+or monetary costs for the peer models.
+
+ConeML pretraining took approximately 11 days on one local RTX 5090. At
+an explicitly assumed average wall draw of 0.70 kW, that corresponds to
+184.8 kWh. Applying an assumed Swiss residential tariff range of
+CHF 0.14–0.30/kWh gives approximately CHF 26–55 of marginal pretraining
+electricity.
+
+| ConeML pretraining unit | estimate |
+|---|---:|
+| average throughput | 12,963 token positions/s |
+| wall time per billion token positions | 21.4 h |
+| energy per billion token positions | 15.0 kWh |
+| energy per million token positions | 15 Wh |
+| estimated wall energy per token position | 0.054 J |
+| marginal electricity per billion token positions | CHF 2.10–4.50 |
+
+The energy and cost figures are estimates derived from the stated power
+and tariff assumptions, not meter readings. They exclude hardware,
+depreciation, labor, supervised fine-tuning, evaluation, conversion,
+datacenter PUE, and carbon intensity. No emissions claim is made.
+
+## Evaluated revisions
+
+- `Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17`
+- `Qwen/Qwen3-0.6B@c1899de289a04d12100db370d81485cdf75e47ca`
+- `unsloth/Llama-3.2-1B-Instruct@5a8abab4a5d6f164389b1079fb721cfab8d7126c`
+- `TinyLlama/TinyLlama-1.1B-Chat-v1.0@fe8a4ea1ffedaf415f4da2f062534de366a451e6`
+- `HuggingFaceTB/SmolLM2-1.7B-Instruct@31b70e2e869a7173562077fd711b654946d38674`
+
+The exact aggregates are in `peer-comparison-summary.json`. Complete
+generation rows are retained privately; their frozen SHA-256 commitments
+are published in `PEER_EVIDENCE_SHA256SUMS.txt`.
+
+## Primary resource disclosures
+
+- Qwen3 pretraining: [Qwen3 release post](https://qwenlm.github.io/blog/qwen3/)
+- Llama 3.2 token count and distillation: [Meta Llama 3.2 model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/MODEL_CARD.md)
+- TinyLlama token count: [TinyLlama model card](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0)
+- SmolLM2 token count: [Hugging Face SmolLM2 model card](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B)

eval/PEER_EVIDENCE_SHA256SUMS.txt

@@ -0,0 +1,31 @@
+96a805aed179f77d9f1959216f9a39e8bf8d50dea5465eeea1eb24e96502a6af  SOURCE_PEER_COMPARISON_V2.md
+a5d6cd4d7a2643b316cefb1dcb6f43296bb69da4064dbc3b3ab4a123bf36b2b1  base-reference/coneml-base188.json
+83643ce110655e708d50f4eb3ee29ab08ab8e003768eb5b5da9a3e545bc9b8ef  base-reference/coneml-base188.rows.jsonl
+ca387426c55ac72d50041e2f082bd4cdc747eedb51deca7929cf7a4acea2c5b6  base-reference/coneml-native-transitive.json
+43ef584d37c6a276b190aeb741597b54bfebbf7bbd9c37df5d82a2e15a8ee2d1  base-reference/qwen35-0.8b-base.json
+25de33180d4073777d29871882b57f1556733e9a8c1e10eba14a8cac062b5c35  base-reference/qwen35-0.8b-base.rows.jsonl
+2003c73a54602f3a30980d6c7f85feef39b63fd11cef9fc20a90044d3569ba77  instruct/coneml-810m-alpha.json
+8da107545bd6717c7795ff4e827698e5e269d451f2390f24dccc516fc390ab44  instruct/coneml-810m-alpha.rows.jsonl
+77d514c14e14557c4ac30e8a8c3f4c5887f0db0dff39b8ea14baea164327d027  instruct/coneml-810m-arithmetic.json
+a60ed3e533c0c4a77b34a21f80b8df8de85e3df074f3cacdc3bbd4b5e0741cce  instruct/coneml-810m-arithmetic.rows.jsonl
+aa35b93f46778ffb22351ac411f2fc04e645775e416b5a9423f925bb6d6b1609  instruct/llama32-1b-instruct-transchat.json
+ee514c35d86f27172a004325b26da915f869164adae59edf8d95b01d958e683f  instruct/llama32-1b-instruct-transchat.rows.jsonl
+f4553ef34844e55364a87f631fac03a3545acbb14cd12dcf182f16151b5e5f8a  instruct/llama32-1b-instruct.json
+d5c0fd8b00b50754ac7dede96081afeaf3143795e005134cf5543483859340bf  instruct/llama32-1b-instruct.rows.jsonl
+a61b56681810410dfc3a8d8e758b8ccd11d01ebcce2d963d6bdbca4644c2cee9  instruct/qwen3-0.6b-transchat.json
+f7c763441edc0d5e37a34902a3cc4efcacc6a35b47056adb9356879a4407cb31  instruct/qwen3-0.6b-transchat.rows.jsonl
+3f34342d79095a5754512ebf57192fdf98a16b364a250712993ac1da0b65e252  instruct/qwen3-0.6b.json
+05847f4187651780008331a85c9cd5fa84d3b636320c9a809fb3c307f1a230d6  instruct/qwen3-0.6b.rows.jsonl
+21984e00a4a6da3d2887138ede25c03e9007fee0be41de462ef58dc58db89554  instruct/qwen35-0.8b-transchat.json
+48a53b6bf3829c207953b1075160df9f6702b1fc2162426577aa69530856e555  instruct/qwen35-0.8b-transchat.rows.jsonl
+fa11fbf67c047ac541ecd38571867f9fe68bfb8aec12b84f2addda8a6d561642  instruct/qwen35-0.8b.json
+74a6debf6ae55db9895f556e654b6905fde871358a94c9c4a2730a418b0e2608  instruct/qwen35-0.8b.rows.jsonl
+72598668bfce72497ac67a4d83d1cad7538c63ec933e8a1d1198876d85dc8a67  instruct/qwen35-thinking-sensitivity.json
+8917677a4c9d29124c5ade229f7080bbd0ccb83471c3d7d39a6f958224596a8f  instruct/smollm2-1.7b-instruct-transchat.json
+deb9eb805475e4ad2d158283b469830138fdf42ee0c9a903bf3611b43a7869ff  instruct/smollm2-1.7b-instruct-transchat.rows.jsonl
+94a3440cdaae14eb27beaf36a1a0103dab4a6d292bc382981f0514a82a93ed0f  instruct/smollm2-1.7b-instruct.json
+1b06370ad7b940ed05b59ffe4acd19035341cfe078ed7b01d59d168abc8d0c71  instruct/smollm2-1.7b-instruct.rows.jsonl
+bdd7cf2867643439e90a3e04cb209e6130839af9422e492bcd2eb1c2cba4556e  instruct/tinyllama-chat-transchat.json
+472aad0c169c4f943fbc9a8367d65fa2d5c41acbbee66d099e16a286f8676d1c  instruct/tinyllama-chat-transchat.rows.jsonl
+3a7c2e7ebcaa78b4b7d9b6d84a5b90a495f260f97287bad3282d54fc84e0c009  instruct/tinyllama-chat.json
+5926a31e926279f8d9459d19cbbe9730a06071bb8b4029c2f9174804ab447943  instruct/tinyllama-chat.rows.jsonl

eval/peer-comparison-summary.json

@@ -0,0 +1,132 @@
+{
+  "schema_version": 1,
+  "date": "2026-07-30",
+  "scope": "Interface-specific diagnostic screen, not a neutral ranking of general model capability. Each instruct model used its native instruction interface. Arithmetic and function-writing used matched short-answer generation budgets.",
+  "decoding": {
+    "strategy": "greedy",
+    "repetition_penalty": 1.15,
+    "arithmetic_n": 585,
+    "core_arithmetic_n": 225,
+    "function_writing_n": 100,
+    "designated_refusal_n": 17,
+    "refusal_contrast_n": 5
+  },
+  "evaluated_revisions": {
+    "ConeML/coneml-810m-alpha": "local release artifact e5df8f2edd773ad079ba00d383ace81b96cbb9579d64c570db505f9efc032b76",
+    "ConeML/coneml-810m-alpha-arithmetic": "local release artifact e1de27f5f91ac43130ce2a57a4bf9a036b84b2e5b589789b5981ecf79489e0d1",
+    "Qwen/Qwen3.5-0.8B": "2fc06364715b967f1860aea9cf38778875588b17",
+    "Qwen/Qwen3-0.6B": "c1899de289a04d12100db370d81485cdf75e47ca",
+    "unsloth/Llama-3.2-1B-Instruct": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
+    "TinyLlama/TinyLlama-1.1B-Chat-v1.0": "fe8a4ea1ffedaf415f4da2f062534de366a451e6",
+    "HuggingFaceTB/SmolLM2-1.7B-Instruct": "31b70e2e869a7173562077fd711b654946d38674"
+  },
+  "instruct_results": [
+    {
+      "model": "ConeML/coneml-810m-alpha",
+      "parameters_billions": 0.8101,
+      "arithmetic_mixed": {"correct": 421, "n": 585, "accuracy": 0.7197},
+      "arithmetic_core_four_lanes": {"correct": 177, "n": 225, "accuracy": 0.7867},
+      "executed_functions": {"correct": 83, "n": 100, "accuracy": 0.83},
+      "designated_refusals": {"correct": 13, "n": 17, "accuracy": 0.7647},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "ConeML/coneml-810m-alpha-arithmetic",
+      "parameters_billions": 0.8101,
+      "arithmetic_mixed": {"correct": 442, "n": 585, "accuracy": 0.7556},
+      "arithmetic_core_four_lanes": {"correct": 221, "n": 225, "accuracy": 0.9822},
+      "executed_functions": {"correct": 35, "n": 100, "accuracy": 0.35},
+      "designated_refusals": {"correct": 11, "n": 17, "accuracy": 0.6471},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "Qwen/Qwen3.5-0.8B",
+      "parameters_billions": 0.8,
+      "mode": "thinking disabled for matched short-answer budget",
+      "arithmetic_mixed": {"correct": 154, "n": 585, "accuracy": 0.2632},
+      "arithmetic_core_four_lanes": {"correct": 150, "n": 225, "accuracy": 0.6667},
+      "executed_functions": {"correct": 93, "n": 100, "accuracy": 0.93},
+      "designated_refusals": {"correct": 0, "n": 17, "accuracy": 0.0},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "Qwen/Qwen3-0.6B",
+      "parameters_billions": 0.6,
+      "arithmetic_mixed": {"correct": 223, "n": 585, "accuracy": 0.3812},
+      "arithmetic_core_four_lanes": {"correct": 208, "n": 225, "accuracy": 0.9244},
+      "executed_functions": {"correct": 98, "n": 100, "accuracy": 0.98},
+      "designated_refusals": {"correct": 1, "n": 17, "accuracy": 0.0588},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "unsloth/Llama-3.2-1B-Instruct",
+      "parameters_billions": 1.24,
+      "arithmetic_mixed": {"correct": 343, "n": 585, "accuracy": 0.5863},
+      "arithmetic_core_four_lanes": {"correct": 221, "n": 225, "accuracy": 0.9822},
+      "executed_functions": {"correct": 79, "n": 100, "accuracy": 0.79},
+      "designated_refusals": {"correct": 1, "n": 17, "accuracy": 0.0588},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+      "parameters_billions": 1.1,
+      "arithmetic_mixed": {"correct": 113, "n": 585, "accuracy": 0.1932},
+      "arithmetic_core_four_lanes": {"correct": 92, "n": 225, "accuracy": 0.4089},
+      "executed_functions": {"correct": 51, "n": 100, "accuracy": 0.51},
+      "designated_refusals": {"correct": 0, "n": 17, "accuracy": 0.0},
+      "over_refusals": {"count": 0, "n": 5}
+    },
+    {
+      "model": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+      "parameters_billions": 1.7,
+      "arithmetic_mixed": {"correct": 380, "n": 585, "accuracy": 0.6496},
+      "arithmetic_core_four_lanes": {"correct": 221, "n": 225, "accuracy": 0.9822},
+      "executed_functions": {"correct": 96, "n": 100, "accuracy": 0.96},
+      "designated_refusals": {"correct": 1, "n": 17, "accuracy": 0.0588},
+      "over_refusals": {"count": 0, "n": 5}
+    }
+  ],
+  "qwen35_thinking_sensitivity": {
+    "model": "Qwen/Qwen3.5-0.8B",
+    "mode": "thinking enabled",
+    "arithmetic_mixed": {"correct": 451, "n": 585, "accuracy": 0.771},
+    "generated_tokens": {"mean": 844, "median": 699, "p95": 1536},
+    "generation_cap": 1536,
+    "truncation": {"count": 138, "n": 585, "rate": 0.2359},
+    "wall_seconds_per_item_mean": 2.43,
+    "correct_per_1000_generated_tokens": 0.913,
+    "comparison": "Qwen thinking reached parity with ConeML Arithmetic's 442/585 while using at least 17.6 times the generated-token budget per item. The recorded wall-time ratio was approximately 24 times. Runtime measurements are hardware- and implementation-specific."
+  },
+  "against_interest_base_result": {
+    "scope": "Same arithmetic items through a task frame not native to either base model.",
+    "ConeML_base_188": {"correct": 156, "n": 585, "accuracy": 0.2667},
+    "Qwen3.5_0.8B_Base": {"correct": 492, "n": 585, "accuracy": 0.841},
+    "interpretation": "The peer base result prevents an absolute-superiority interpretation and illustrates sensitivity to training exposure and interface."
+  },
+  "training_resource_context": {
+    "ConeML": {
+      "pretraining_token_positions": 12320000000,
+      "tokens_per_parameter": 15.2,
+      "estimated_training_flops_6ND": 5.99e19,
+      "pretraining_wall_time_days": 11,
+      "hardware": "one NVIDIA RTX 5090",
+      "power_assumption_kw_average_wall": 0.70,
+      "estimated_pretraining_energy_kwh": 184.8,
+      "assumed_swiss_residential_tariff_chf_per_kwh": [0.14, 0.30],
+      "estimated_marginal_pretraining_electricity_chf": [25.87, 55.44],
+      "exclusions": ["hardware", "depreciation", "labor", "SFT", "evaluation", "conversion", "grid-carbon estimate"]
+    },
+    "peer_compute_context": [
+      {"model": "TinyLlama 1.1B", "published_pretraining_tokens": 3000000000000, "approx_tokens_per_parameter": 2727, "approx_flops_vs_coneml": 330},
+      {"model": "Llama 3.2 1B", "published_pretraining_tokens": 9000000000000, "approx_tokens_per_parameter": 7258, "approx_flops_vs_coneml": 1118, "note": "plus distillation"},
+      {"model": "SmolLM2 1.7B", "published_pretraining_tokens": 11000000000000, "approx_tokens_per_parameter": 6471, "approx_flops_vs_coneml": 1873},
+      {"model": "Qwen3 0.6B", "published_pretraining_tokens": 36000000000000, "approx_tokens_per_parameter": 60000, "approx_flops_vs_coneml": 2164},
+      {"model": "Qwen3.5 0.8B", "published_pretraining_tokens": null, "approx_tokens_per_parameter": null, "approx_flops_vs_coneml": null}
+    ]
+  },
+  "evidence": {
+    "public_interpretation": "PEER_COMPARISON.md",
+    "private_rows": "retained by ConeML",
+    "hash_commitment": "PEER_EVIDENCE_SHA256SUMS.txt"
+  }
+}