Update README.md

README.md

@@ -24,55 +24,6 @@ canary GUID crosssearch:5157836c-9b68-4c5c-a16f-156cf0edc5df
 
 This resource contains **100 tasks and 4,650 evaluation runs**. The current evaluation comprises **3,300 native-search runs** (11 model configurations, three final attempts per task) and **1,350 uncapped search-backend runs** (three models, 50 tasks, three alternative backends, three attempts).
 
-## Inkling 2 Small — November 8, 2026
-
-Added **300 native-search trials** of `thinkingmachines/Inkling-2-Small:peft:262144` across the same 100 tasks, three trials per task. Inkling 2 Small scored **44.0% pass@1** (132/300) and **23.0% pass³** (23/100). This cohort replaces the 300 original Inkling runs in the current Hugging Face release. The blog continues to show both configurations.
-
-GPT-6.1 Sol graded 294 completed answers with the existing rubric and exact-evidence checks. Three context-limit outcomes and three timeouts are retained as unsuccessful trials with null rubric scores. All 300 slots remain in the pass@1 and pass³ denominators. The original first 100 runs are preserved. Mean rubric score across the 294 valid grades was 82.77%.
-
-The new `rollouts-inkling2-20261108.jsonl.zst` archive includes saved traces, final answers, request/response records, binary artifacts and grading evidence with operational credentials and local author paths removed. `run-index-inkling2-20261108.jsonl.gz` and `grades-inkling2-20261108.jsonl.gz` identify this cohort. The combined index contains 4,650 entries and preserves all 4,350 entries from the other configurations. Use this index to select current runs. Source manifests and row hashes are in `verification-inkling2-20261108.json`.
-
-Agent cost was estimated from saved uncached input, cached input and output usage at [Tinker’s published prices](https://tinker-docs.thinkingmachines.ai/tinker/models/models_and_pricing/), excluding judging and infrastructure. Cost per fully correct answer was $0.32, median agent cost was $0.09, median agent duration was 15.7 minutes, and median output usage was 29,743 tokens.
-
-## GPT-6.1 Sol regrading — November 6, 2026
-
-The earlier **4,650-answer release** was regraded with GPT-6.1 Sol using the original task context, reference answer, rubric and answer. No generation was rerun. The 4,350 retained answers and their original grades are preserved; their current scores are in **run-index.jsonl.gz** and **grades-gpt61-20261106.jsonl.gz**. Valid low and zero grades are included.
-
-The new judge used high reasoning effort, temperature 0 and a 16,000-token response limit. Every accepted grade has the original rubric IDs, a normal completed response and verbatim evidence from the answer. Invalid responses were retained separately and replaced only by distinct completion attempts on the same frozen input. See [the grading protocol](grading-protocol.md).
-
-The independent validation study compared 172 answers from 43 overlapping tasks against two human grades per answer. GPT-6.1 Sol agreed with an individual human on 94.8% of whole-answer pass/fail decisions, compared with 94.2% human–human agreement and 89.0% for Gemini 3.7 Flash. That study used question-only task context; full benchmark grading also supplied the original task context and reference answer.
-
-## November 2026 update
-
-This release includes **1,650 updated runs**: 300 Gemini 4 Argon native-search runs and 1,350 uncapped Exa, Parallel and Firecrawl runs for GPT-5.6 Sol, Kimi K3 and Gemini 3.8 Flash. Each original task/model/condition/replicate cell is represented once in this new cohort. Earlier incomplete attempts are not substituted for complete results, and valid low and zero scores are retained.
-
-The added files begin with `rollouts-20261104-`. They contain the original final answers, rubric judgments, available model reasoning, saved request/response captures and trace files, with operational credentials and local author/account paths removed. The source archives remain unchanged. Search/fetch call-count limits are disabled in the new search-backend cohort; task time, token, context and response-size limits are unchanged. The uncapped search-backend results replace the earlier capped comparison.
-
-
-## Files
-
-- CANARY.txt: evaluation-only notice and training-contamination canary strings.
-- tasks.jsonl.gz: expert questions, domain labels, frozen rubrics and evidence maps.
-- tasks.tar.gz: starting workspace assets, separated from grader-only reference material.
-- rollouts-inkling2-20261108.jsonl.zst: all 300 Inkling 2 Small trial records, losslessly compressed with Zstandard. The uncompressed JSONL row hashes are unchanged.
-- rollouts-*.jsonl.gz: saved agent trace files, API request/response captures, final answers, original rewards and criterion judgments. Protected source archives are exported as readable derivatives. Source archives remain unchanged; no private key is distributed.
-- run-index.jsonl.gz: current GPT-6.1 Sol scores and retained limit outcomes for all 4,650 runs, with original rewards, rollout shards and versioned grade pointers.
-- grades-inkling2-20261108.jsonl.gz: 294 valid Inkling 2 Small grades and six explicit limit outcomes.
-- run-index-inkling2-20261108.jsonl.gz: all 300 Inkling 2 Small slots with original outcome status and shard pointers.
-- analysis-inkling2-20261108.json and verification-inkling2-20261108.json: cohort metrics and exact coverage/checksum evidence.
-- grades-gpt61-20261106.jsonl.gz: current grades, per-item decisions, rationales, verbatim evidence and request/response/answer hashes; sanitized answers are included for convenience.
-- run-index-original-20261105.jsonl.gz: the pre-regrade index for the 4,350 retained runs.
-- analysis-gpt61-20261106.json and verification-gpt61-20261106.json: versioned metrics and complete-coverage checks.
-- CrossSearch-GPT61-Judge-Prompt.txt: the frozen semantic judge prompt used for regrading.
-- run-index-20261104.jsonl.gz: just the 1,650 generation runs added on November 4, now with current and original rewards, including `release_cohort` and `search_call_limit`.
-- judge-models.csv: one row for each of the 415 audited encrypted extension records, with evidence references and recorded judge labels.
-- CrossSearch-Judge-Prompt.txt and CrossSearch-Agent-Prompt.txt: appendix prompt templates, message roles and source notes.
-- grading-protocol.md: scoring protocol and judge-version uncertainty.
-- judge-model-audit.json: recorded judge fields for the 415 encrypted extension records.
-- analysis-summary.json: current GPT-6.1 Sol results, uncapped search comparisons and efficiency values.
-- analysis-summary-20261104.json: the earlier results under the original grading protocol, preserved for comparison.
-- release-manifest.json: final counts, shard checksums, redaction totals and verification status. Check this manifest for completeness.
-
 ## Reading the data
 
     from datasets import load_dataset
@@ -129,7 +80,3 @@ Some exceptionally large traces use `large_record_parts`. Their normal rollout r
     print(current[0]['reward'], current[0]['original_reward'])
 
 Current scores and original scores use the same answer. The versioned grade record links them by run ID and SHA-256 hashes; it does not overwrite the original recorded judgment.
-
-## Reading the Inkling 2 Small archive
-
-The Inkling archive uses lossless Zstandard compression with a 128 MiB window. After downloading it, run `zstd -d -c rollouts-inkling2-20261108.jsonl.zst > inkling2.jsonl`, then read the JSONL with Python. Each row is the complete saved record, including its trace and binary members. The other rollout files continue to use gzip.