Move the Amazon Bedrock runs to amazon-bedrock/<model>/ with an index; add 11 runs (medium effort, 2x / 3x limits) and labeled variants
Moves the 26 Amazon Bedrock run directories added in #18 from the repository root into one folder, amazon-bedrock/<model>/<effort>[-limits-<K>x]/, adds 11 runs and labeled variants, and adds an index, amazon-bedrock/README.md, listing every run with its category, scores per benchmark and directory. Other contributors' directories are not touched. This PR also carries #19 (GPT-6 Astra high scored with its second SWE-Live sample) and replaces it.
Layout: one folder per model. Standard runs (README limits) are <effort>/; the non-standard experiment with every interaction limit scaled by 2 or 3 is <effort>-limits-2x/ and <effort>-limits-3x/. Labeled variants are extra files inside a run directory and never its standard score.
Moved without changes (same bytes as on main, or as in #19): 22 runs. Rebuilt with additional attempts: Claude Fable 5.1 xhigh, max, max 2x and max 3x (summary.json and scores.json unchanged; the archive and attempts.jsonl now also hold the timeout re-run attempts).
New standard runs (1x, reasoning_effort medium):
- GPT-6 Astra medium: 38/99
- Claude Fable 5.1 medium: 42/99
- GPT-6.1 Sol medium: 31/99
- GPT-5.6 Terra medium: 21/99
New non-standard runs (every limit x2 / x3, not comparable to standard runs):
- GPT-6 Astra xhigh 2x: 45/99
- GPT-6.1 Sol max 2x: 42/99, 3x: 41/99
- GPT-5.6 Terra max 2x: 18/99, 3x: 26/99
- GPT-6 Sol max 2x: 18/99, 3x: 16/99
New labeled variants:
- format-tolerant (
summary-format-tolerant.json,scores-format-tolerant.json; format-only breaks tolerated through a normalizing proxy, principle breaks still rejected): GPT-6 Astra medium 42/99, GPT-6.1 Sol medium 33/99, GPT-6 Sol max 2x 19/99 and 3x 18/99, GPT-5.6 Terra max 2x 18/99 and 3x 26/99. - timeout re-run (
summary-timeout-rerun.json,scores-timeout-rerun.json): Claude Fable 5.1's scored attempts that timed out (fail with total_timeout, or api_interrupted after an agent call timed out; 170 tasks) were run once more with the same config, code and calibration gate: xhigh 29 -> 30, max 25 -> 26, max 2x 32 -> 32, max 3x 32 -> 33. Other runs' timeouts had one rerun only.
Not included: Grok 4.7 2x and 3x, stopped after 14 and 10 tasks.
Setup and policies as in #18; details in amazon-bedrock/README.md and each run's notes.txt.
Validation: 99 unique task IDs per run; every scored attempt (standard and variants) present in its archive; archive index equal to the archive members for all 37 runs; sensitive-pattern scans passed; the 22 moved runs byte-identical to their current versions.