Datasets:

---

MatrAIx2026

/

PersonaCodeBench

Like 0

Follow

MatrAIx 86

Adds 26 run directories on the 99 selected CPU tasks, one directory per model and reasoning effort: 22 standard runs of GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.1, GPT-5.6 Terra, GPT-6 Sol and GPT-6 Luna (high / xhigh / max), GPT-5.5 (high / xhigh), GLM 5.3 and Grok 4.7 (high), plus 4 directories of a separate, NON-STANDARD experiment (*-limits-2x/, *-limits-3x/). Every recorded attempt is retained (3981 attempts). GPU task tb3/exam-pdf-eval was not run. This PR replaces #8-#17.

Standard runs, scored results (one attempt per task; the later attempt replaces the earlier one):

• gpt-6-astra-high/: 48 pass, 48 fail, 1 invalid, 2 infrastructure_error; valid pass/fail 96/99; 122 attempts recorded. Format-tolerant variant (separately labeled): 50 pass, 48 fail, 1 invalid, 0 infrastructure_error.
• gpt-6-astra-xhigh/: 44 pass, 54 fail, 1 invalid, 0 infrastructure_error; valid pass/fail 98/99; 115 attempts recorded.
• gpt-6-astra-max/: 34 pass, 62 fail, 3 invalid, 0 infrastructure_error; valid pass/fail 96/99; 168 attempts recorded.
• gpt-6.1-sol-high/: 38 pass, 54 fail, 1 invalid, 6 infrastructure_error; valid pass/fail 92/99; 124 attempts recorded. Format-tolerant variant (separately labeled): 41 pass, 56 fail, 1 invalid, 1 infrastructure_error.
• gpt-6.1-sol-xhigh/: 42 pass, 54 fail, 1 invalid, 2 infrastructure_error; valid pass/fail 96/99; 139 attempts recorded. Format-tolerant variant (separately labeled): 43 pass, 55 fail, 1 invalid, 0 infrastructure_error.
• gpt-6.1-sol-max/: 35 pass, 62 fail, 2 invalid, 0 infrastructure_error; valid pass/fail 97/99; 125 attempts recorded.
• claude-fable-5-1-high/: 39 pass, 56 fail, 4 invalid, 0 infrastructure_error; valid pass/fail 95/99; 141 attempts recorded.
• claude-fable-5-1-xhigh/: 29 pass, 65 fail, 5 invalid, 0 infrastructure_error; valid pass/fail 94/99; 150 attempts recorded.
• claude-fable-5-1-max/: 25 pass, 68 fail, 6 invalid, 0 infrastructure_error; valid pass/fail 93/99; 172 attempts recorded.
• gpt-5.6-terra-high/: 27 pass, 70 fail, 2 invalid, 0 infrastructure_error; valid pass/fail 97/99; 118 attempts recorded.
• gpt-5.6-terra-xhigh/: 34 pass, 63 fail, 2 invalid, 0 infrastructure_error; valid pass/fail 97/99; 127 attempts recorded.
• gpt-5.6-terra-max/: 11 pass, 71 fail, 11 invalid, 6 infrastructure_error; valid pass/fail 82/99; 196 attempts recorded. Format-tolerant variant (separately labeled): 12 pass, 80 fail, 4 invalid, 3 infrastructure_error.
• gpt-6-sol-high/: 16 pass, 33 fail, 16 invalid, 34 infrastructure_error; valid pass/fail 49/99; 198 attempts recorded. Format-tolerant variant (separately labeled): 17 pass, 36 fail, 15 invalid, 31 infrastructure_error.
• gpt-6-sol-xhigh/: 14 pass, 28 fail, 20 invalid, 37 infrastructure_error; valid pass/fail 42/99; 222 attempts recorded. Format-tolerant variant (separately labeled): 15 pass, 34 fail, 18 invalid, 32 infrastructure_error.
• gpt-6-sol-max/: 12 pass, 27 fail, 21 invalid, 39 infrastructure_error; valid pass/fail 39/99; 231 attempts recorded. Format-tolerant variant (separately labeled): 14 pass, 35 fail, 18 invalid, 32 infrastructure_error.
• gpt-6-luna-high/: 13 pass, 81 fail, 3 invalid, 2 infrastructure_error; valid pass/fail 94/99; 147 attempts recorded. Format-tolerant variant (separately labeled): 13 pass, 82 fail, 2 invalid, 2 infrastructure_error.
• gpt-6-luna-xhigh/: 9 pass, 80 fail, 5 invalid, 5 infrastructure_error; valid pass/fail 89/99; 164 attempts recorded. Format-tolerant variant (separately labeled): 10 pass, 84 fail, 4 invalid, 1 infrastructure_error.
• gpt-6-luna-max/: 14 pass, 73 fail, 7 invalid, 5 infrastructure_error; valid pass/fail 87/99; 174 attempts recorded. Format-tolerant variant (separately labeled): 15 pass, 77 fail, 4 invalid, 3 infrastructure_error.
• gpt-5.5-high/: 14 pass, 69 fail, 7 invalid, 9 infrastructure_error; valid pass/fail 83/99; 143 attempts recorded. Format-tolerant variant (separately labeled): 14 pass, 73 fail, 3 invalid, 9 infrastructure_error.
• gpt-5.5-xhigh/: 18 pass, 67 fail, 8 invalid, 6 infrastructure_error; valid pass/fail 85/99; 138 attempts recorded. Format-tolerant variant (separately labeled): 18 pass, 68 fail, 7 invalid, 6 infrastructure_error.
• glm-5.3-high/: 22 pass, 69 fail, 7 invalid, 1 infrastructure_error; valid pass/fail 91/99; 146 attempts recorded.
• grok-4.7-high/: 7 pass, 83 fail, 6 invalid, 3 infrastructure_error; valid pass/fail 90/99; 156 attempts recorded.

Non-standard experiment (not comparable to standard runs): GPT-6 Astra and Claude Fable 5.1 at reasoning_effort max with every interaction limit scaled by K = 2 or 3 - the task rows' budgets (episode and verifier seconds, user turns, agent queries per turn, simulator decisions, agent steps), each runtime's own interaction limits, and the per-call API timeouts (agent 600 s -> 1200 / 1800 s, simulator 300 s -> 600 / 900 s). Not scaled: format-repair counts, checker-internal grading deadlines (hash-locked task assets), environment preparation, cleanup grace and transport retries. Code: upstream main 5933dd2 plus a limit-scaling patch selected by SWE_PERSONA_LIMIT_SCALE; each task's interaction clock was checked to equal K x its default budget.

• claude-fable-5-1-max-limits-2x/: 32 pass, 61 fail, 5 invalid, 1 infrastructure_error; valid pass/fail 93/99; 164 attempts recorded.
• gpt-6-astra-max-limits-2x/: 42 pass, 52 fail, 3 invalid, 2 infrastructure_error; valid pass/fail 94/99; 120 attempts recorded.
• claude-fable-5-1-max-limits-3x/: 32 pass, 56 fail, 8 invalid, 3 infrastructure_error; valid pass/fail 88/99; 165 attempts recorded.
• gpt-6-astra-max-limits-3x/: 48 pass, 49 fail, 2 invalid, 0 infrastructure_error; valid pass/fail 97/99; 116 attempts recorded.

Each directory contains manifest.json, summary.json, scores.json (scripts/score.py), attempts.jsonl (every attempt with its round and whether it is the scored one), notes.txt, artifact-index.jsonl (size and SHA-256 of every archive member) and task-artifacts.tar.gz (per attempt: results, run config, clocks, episodes, coding-agent / simulator / judge call logs, judge and checker reports, patches and logs). Extract the archive inside the run directory; manifest artifact_path values name archive members. Directories with a format-tolerant variant also carry summary-format-tolerant.json and scores-format-tolerant.json: format-only breaks (one action wrapped in prose or repeated) tolerated, principle breaks still rejected; not the standard score.

Setup: upstream main 5933dd2, harness 0.5.0rc9. Coding agents, user simulator and judge on Amazon Bedrock (us-west-2): GPT, GLM and Grok models through a local SigV4 forwarder exposing Bedrock's OpenAI-compatible endpoint unchanged; Claude Fable 5.1 through a local LiteLLM proxy translating the OpenAI chat format. Simulator and judge: gpt-6-luna at reasoning_effort medium (README's default is an Azure deployment). Coding agents: max_completion_tokens 128000.

Policies and deviations (details in each notes.txt): a blocked calibration gate was re-sampled with a fresh cache (up to 8 attempts); every invalid or infrastructure_error task and every fail with total_timeout or api_interrupted was rerun once; standard runs also have targeted fix-ups for infrastructure interruptions and judge protocol errors; drained batches continue in "-cont" batches (same attempt). Raw error classifications are preserved; 99 recorded tasks are not 99 valid completions.

Redaction and exclusions: no credentials were present; local machine paths were replaced with "~". Copied runtime and project sources, images, Docker build contexts, workspace snapshots and other binaries were excluded.

Validation: 99 unique task IDs per directory, 3981 attempt records, status totals, 272,881 archive member hashes, presence of every scored attempt in its archive, and sensitive-pattern scans passed locally.