Skip to content

P10a: LLM candidates for the experiment, housekeeping, and the dev model pilot - #32

Merged
Ayushxsingh100 merged 7 commits into
mainfrom
feature/p10a
Nov 14, 2026
Merged

Ayushxsingh100 merged 7 commits into
mainfrom
feature/p10a

Conversation

@Ayushxsingh100

Copy link
Copy Markdown
Owner

P10a, the first P10 package: LLM candidates for the experiment, housekeeping, and the dev model pilot. Not reviewed yet. The approval record (C-28) will be the external verdict linked here.

What changed

  • Decisions: C-212 to C-215 and C-217 are Decided. New rows C-218 to C-221 (Decided by the authors) and C-222 (readings, Proposed).
  • Gate status (from the verdict on PR P8b: scoring, metrics and the dev report (P8 gate, M5) #31): P8b, the P8 gate and M5 are passed with limits. The verdict's limits 4 to 8 are not recorded, because their text was not pasted.
  • Validator dail-patch-validator-2: a MODIFY that drops the parent's lifecycle block is now POLICY_VIOLATION (C-221).
  • num_ctx: LLMConfig.num_ctx is optional and sent to Ollama when set.
  • experiments/generation/:
    • generation sees S0 only; the task is the fixed template;
    • the scope is decision 5's;
    • models are pinned by tag and digest, and a different installed digest is refused;
    • every attempt is frozen and listed in a manifest;
    • replay rebuilds each candidate, checks the tree hash, then deletes the build;
    • the oracle export holds configuration only.
  • Extractor: extractor-2 is now a separate version from metrics-draft-1, and its code is pinned by hash. New generation metrics and a pilot report.
  • Concurrency test: the named interleavings now wait for the A_done/B_done markers.
  • Blinding contract: new checks (c) and (d) for the generation package.
  • Pilot: docs/P10A_MODEL_PILOT.md is generated from the frozen records.

Checks

  • dev.py test: 4519 tests OK (7 skipped). Lint is clean. pytest exits 0.
  • 11 of 11 mutants were caught, plus 2 of 2 generation-blinding mutants.
  • The concurrency file passed 50 of 50 runs under CPU load. Caveat: the base file also passed 50 of 50, so the timing failure was not reproduced.

Known issues

  • One pilot call to phi4-mini lasted 11 h, almost certainly because the machine slept during it. Its timing numbers are distorted; the PHASE_GATES P10a section gives the figures without that call.

🤖 Generated with Claude Code

Ayushxsingh100 and others added 7 commits November 13, 2026 19:45
… which scenarios, scope, lifecycle); P8b, the P8 gate and M5 passed with limits per the verdict on PR #31 (Doc 15 §15.3, §25)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…across conditions, digest refusal, freezing, replay rebuild hash, oracle export, extractor versions (Doc 10 §13, §16; Doc 13 §26, §47; C-219 to C-221)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… metrics-draft-1, both on every record and report, the extractor code pinned by hash; the named interleavings wait for A_done/B_done (Doc 13 §32; P8b verdict limits 2, 3)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ecycle block; experiments/generation (S0-only input, the fixed task, the decision-5 scope, pinned models with the digest check and num_ctx, generate and freeze, replay with the rebuild hash check, configuration-only oracle export); generation metrics and the pilot report (Doc 10 §13, §16, §24; Doc 13 §26, §47; C-219 to C-221)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…dels (qwen2.5-coder:7b, granite3.3:8b, phi4-mini:3.8b) with digests and licence checks (C-182, C-218)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rated from the frozen records), licence checks, PHASE_GATES P10a section, BUILD_SEQUENCE row 24, C-222; the pilot renderer sorts decision counts (Doc 13 §26, §33; C-218 to C-222)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… in docs/P10A_MODEL_PILOT.json ("key": "granite3_3-8b", a model key, not a secret)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Ayushxsingh100
Ayushxsingh100 merged commit 839c2d3 into main Nov 14, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant