Skip to content

Latest commit

 

History

History
211 lines (178 loc) · 12 KB

File metadata and controls

211 lines (178 loc) · 12 KB

Model Refresh Status — 2026-07-11

This is a compatibility and infrastructure checkpoint, not a model-quality result. Early runs below used an intentionally dirty implementation checkout; only sections explicitly labeled as clean-revision verification satisfy the current provenance gate.

Implemented contract

  • Central registry: config/model_matrix.toml
  • Default matrix: current_balanced
  • Exact Inspect/provider pins: Inspect 0.3.245, OpenAI SDK 2.45.0, Anthropic SDK 0.116.0
  • Per-model generation profiles with requested/resolved model validation
  • GPT-5.6 structural ModelInfo registered from provider documentation
  • HF schema 0.3 requested/resolved/provider/config/Inspect provenance
  • Limit-exhausted samples rejected by the HPC cell validator and scored export
  • growth_01 uses source-backed defensible ranges rather than hidden exact starting-OD/cadence answers, with consistency required across each run
  • growth_01 separates a 40-turn agent cap from a 160-message hard cap so parallel tool results do not consume the agent-iteration budget
  • transform_01 uses the same separate 40-turn/160-message protection and scores conditions from cultures that produced usable downstream counts
  • clone_01 scores successful digest/ligation conditions, credits corrected retries as troubleshooting, and verifies successful reactions and reported transformant counts against the trajectory

Cayuga verification

The implementation was copied to an isolated Cayuga scratch checkout before verification. Site-specific paths and scheduler identifiers are omitted because they do not affect reproduction; immutable code revisions and check outcomes are recorded below.

Environment setup completed successfully. The final check passed 458 tests, Ruff, shell syntax, registry validation, wheel build, isolated wheel installation, Inspect entry-point import, and GPT-5.6 metadata registration.

The compatibility array ran growth_01, seed 0, serially across the four current-core models. All four providers returned the expected resolved model ID with the registered {max_tokens: 16384, reasoning_effort: medium} profile and Inspect 0.3.245:

Requested model Resolved model Current cell status
openai/gpt-5.6-sol gpt-5.6-sol Valid
openai/gpt-5.6-luna gpt-5.6-luna Valid
anthropic/claude-sonnet-5 claude-sonnet-5 Valid
anthropic/claude-haiku-4-5-20251001 claude-haiku-4-5-20251001 Compatible; original run incomplete

The Haiku retry recorded --message-limit 120 in its cell manifest but again exhausted the message limit. The strengthened validator rejected the cell. Trajectory inspection showed that this was not a repeated-call loop: the model completed a reasonable 30-minute pilot, diagnosed one undersampled fit, reran at 20-minute cadence, and issued the final three fit calls immediately before the message-only limit terminated the sample. Inspect counts each parallel tool result as a separate message, so raising only that limit did not address the task-contract problem.

After aligning the non-answer-bearing prompt and scorer, the Haiku compatibility smoke completed in 63 seconds and passed the strict cell validator. It used 59 messages, 15 assistant turns, and 42 tool calls: all three cultures started at OD600 0.05, every incubation used a 15-minute interval, and all three final fits were analyzable. Re-scoring that transcript with the final consistency rule produced 1.0 on task success, decision quality, troubleshooting, and efficiency. This is a single compatibility smoke from a dirty checkout, not a model-quality estimate.

The final contract check passed 462 tests, Ruff, shell syntax, registry validation, wheel build, isolated installation, and package smoke in the pinned Cayuga environment.

Clean revision verification

The first clean-checkout array exposed two remaining scorer false negatives: colon-only final-answer parsing rejected dash-formatted reports, and the OD600 lower bound rejected scientifically defensible 0.01 and 0.02 starts. The original logs remain unchanged. Replaying their trajectories after the source-backed OD600 0.01-0.10 correction and colon/dash/Markdown-table parser fix produced 1.0 on all four axes for all four models.

During the next clean run, source-path inspection found that the reused virtual environment was editable-installed from an older checkout. That array was cancelled and is invalid for release use. The runner now prefixes the submitted checkout on PYTHONPATH, fails before any API call when the imported src root does not match REPO_ROOT, and records that root in cell manifest schema 1.2.0.

Final code revision 557194792520d27e64c545bed127402061fb9d0c was copied to a fresh isolated checkout. The isolated check passed 467 tests, Ruff, shell syntax, registry validation, wheel build, isolated installation, and package smoke. The compatibility array then completed growth_01, seed 0, serially across the four current core models:

Requested model Resolved model Messages Assistant turns Tool calls Stored overall
openai/gpt-5.6-sol gpt-5.6-sol 47 12 33 1.000
openai/gpt-5.6-luna gpt-5.6-luna 47 12 33 1.000
anthropic/claude-sonnet-5 claude-sonnet-5 55 14 39 1.000
anthropic/claude-haiku-4-5-20251001 claude-haiku-4-5-20251001 59 15 42 1.000

Every cell passed the strict validator with Inspect 0.3.245, the registered generation profile, worktree_dirty=false, manifest schema 1.2.0, the exact commit above, and the expected runtime source root. These four rows are still compatibility smokes from one task and one seed, not comparative model-quality estimates.

Non-growth sentinel verification

The next gate used pcr_01 and screen_01, seed 0, across the same four-model matrix. The initial clean array exposed a screen-parser false negative: three models listed six explicit colony IDs, but the parser read the numeric suffix of white_001 as a total count of one. The source trajectories were correct and remain unchanged. Commit 35b8221d6f7dc856e6f49659575166baf6938aaf fixed ID-list counting; the post-fix check passed 468 tests and the clean array stored full-success screen scores:

Screen model Messages Assistant turns Tool calls Task Decision Troubleshooting Efficiency Overall
openai/gpt-5.6-sol 7 3 2 1.000 1.000 1.000 1.000 1.000
openai/gpt-5.6-luna 7 3 2 1.000 1.000 1.000 1.000 1.000
anthropic/claude-sonnet-5 7 3 2 1.000 1.000 1.000 1.000 1.000
anthropic/claude-haiku-4-5-20251001 9 4 3 1.000 1.000 1.000 1.000 1.000

PCR audit found that common Q5 labels normalized to canonical Q5 in the output but could still receive unsupported-polymerase simulator behavior. Commit cbbebbd128bbf2fde99de1e57507018d72abf9ce canonicalized Q5 and Phusion before simulation; the post-fix check passed 472 tests and the clean PCR array completed cleanly. A final scorer audit then prevented favorable parameters from unrelated failed attempts being combined into decision credit: commit 5229e270297357a7a6adbed3b6aec5a77bcc62d9 restricts PCR decision matching to clean_target_band reactions, and the final check passed 473 tests.

Replaying the clean PCR trajectories under that final scorer contract produced:

PCR model Messages Assistant turns Tool calls Task Decision Troubleshooting Efficiency Overall
openai/gpt-5.6-sol 26 9 15 1.000 0.750 1.000 0.000 0.825
openai/gpt-5.6-luna 22 8 12 1.000 1.000 1.000 0.000 0.900
anthropic/claude-sonnet-5 13 5 6 1.000 1.000 1.000 0.500 0.950
anthropic/claude-haiku-4-5-20251001 17 6 9 1.000 1.000 1.000 0.000 0.900

Sol's PCR decision score is a legitimate protocol-choice difference: its successful reaction used a 75-second extension, outside the cited 40-60 second range. All eight retained screen/PCR cells passed requested/resolved model, generation profile, clean worktree, schema 1.2.0, runtime-source-root, and no-limit-exhaustion gates. These remain one-seed sentinel runs, and neither the initial parser-diagnostic rows nor frozen historical artifacts were rewritten.

Remaining snapshot sentinel verification

The first transform_01 / clone_01 array at clean revision 50f5cdd72e903745fb07d7c1357a69c1ae430ca6 exposed four contract issues while leaving the source logs unchanged: a 40-message-only transform cap terminated a valid batched run; successful dilution, digest, and ligation retries were mixed with failed attempts for decision scoring; reagent filters treated Ampicillin and ampicillin differently; and clone task success did not verify successful reactions or the reported transformant count.

Commit 9a60771a0d1f8c9d72901ddacf065df91f848a3b separated the transform turn/message limits, restricted decision scoring to usable final workflows, made string filters case-insensitive, credited trajectory-resolved clone failures, and strengthened clone task-success reconstruction. Its clean diagnostic array then exposed two report-shape false negatives: Unicode superscript exponents such as 10⁹ were not parsed, and a valid sum of two same-culture, same-dilution plates was compared only with the largest individual plate count. Commit 5168651364ad29529c14b9275b04f794a340f15f accepts those forms while continuing to reject raw-count sums across different dilutions.

One infrastructure retry was also required. The shared Cayuga environment had drifted to Inspect 0.3.222, so the pinned metadata check failed and the evaluation array was cancelled. A new immutable environment restored Inspect 0.3.245, OpenAI 2.45.0, and Anthropic 0.116.0. The cancelled bundle is diagnostic-only and must not be aggregated.

Final revision 5168651364ad29529c14b9275b04f794a340f15f was copied to a fresh isolated checkout. The final check passed 484 tests, Ruff, shell syntax, registry validation, wheel build, isolated installation, and package smoke. The final serial array used RUN_ID 2026_07_12_snapshot_remaining_final_5168651:

Task Model Messages Assistant turns Tool calls Task Decision Troubleshooting Efficiency Overall
transform_01 openai/gpt-5.6-sol 33 8 23 1.000 1.000 1.000 0.500 0.950
transform_01 openai/gpt-5.6-luna 47 10 35 1.000 0.667 1.000 0.000 0.800
transform_01 anthropic/claude-sonnet-5 41 9 30 1.000 0.833 1.000 0.000 0.850
transform_01 anthropic/claude-haiku-4-5-20251001 24 5 17 0.000 0.500 1.000 0.500 0.400
clone_01 openai/gpt-5.6-sol 44 17 25 1.000 1.000 1.000 0.000 0.900
clone_01 openai/gpt-5.6-luna 27 11 14 1.000 1.000 1.000 0.500 0.950
clone_01 anthropic/claude-sonnet-5 28 11 15 1.000 1.000 1.000 0.500 0.950
clone_01 anthropic/claude-haiku-4-5-20251001 29 13 14 1.000 1.000 1.000 0.500 0.950

Haiku's transform task failure is substantive rather than parser-related. Only one count observation reached status=plated; the remaining observations were outside the cited 25-250 range, so the trajectory never produced a valid four-mass measurement set. All eight final cells passed requested/resolved model, generation profile, clean worktree, schema 1.2.0, exact runtime source, Inspect-version, and no-limit-exhaustion gates.

Scale decision

All five snapshot tasks now have a strict one-seed sentinel across the four current-core models. These runs validate compatibility and scorer contracts; they are not comparative quality estimates. Human-baseline and multi-seed work remain intentionally skipped. The next bounded gate should use the same four-model, seed-zero protocol on the newer current wet-lab tasks, while frozen historical results and every cancelled or diagnostic bundle remain excluded from promoted aggregates.