core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
core/HANDOFF-gpt55-2026-06-20.md

5.4 KiB
Raw Permalink Blame History

HANDOFF — GPT-5.5 — 2026-06-20

Agent and session

  • Agent: GPT-5.5 / Codex
  • Reasoning effort: high architecture review
  • Session entry point: produce the master implementation plan for moving math comprehension from local raw-text organs to substrate-mediated constructional cognition; do not implement runtime code.
  • Baseline: origin/main@60df4fc0, merged audit PR #833.
  • Branch: codex/semantic-substrate-problem-solving-plan

Bootstrap and smoke

Read before editing:

  • GPT55.md
  • AGENTS.md
  • docs/runtime_contracts.md
  • the ADR-0223 audit/session pack and previous handoff
  • current ProblemFrame, contracts, derivation boundaries, pack/compiler, graph-carrier boundaries, governing ADRs, diagnostics, and focused tests

Direct CLI result:

core test --suite smoke -q
zsh: command not found: core

Repository-supported fallback:

uv run python -m core.cli test --suite smoke -q
108 passed, 1 warning in 129.78s

The warning reports a stale engine_state revision. This docs-only task did not clear or mutate engine state.

Modules touched

File Change Summary
docs/analysis/semantic-substrate-problem-solving-master-plan-2026-06-20.md added 14-section master architecture, catalog, PR sequence, immediate PR specification, metrics, retirement, governance, validation, risk, and exact handoff prompt
HANDOFF-gpt55-2026-06-20.md updated this bounded implementation handoff, superseding the audit-session handoff now captured by the merged audit artifacts

No Python, runtime, pack, manifest, benchmark, report, case, sealed, test, or UI file was modified.

Decisions

  1. The transformation boundary is:

    chunk -> exact substrate proposal -> construction candidate -> role obligations
  -> exact bindings -> ProblemFrame -> ContractAssessment -> typed organ input
  -> verified derivation -> existing admission/articulation

  2. Semantic closeness proposes; exact bindings ground; contracts determine; verification admits. CGA score never proves a relation or selects an operator.

  3. The next implementation PR remains proportional-decrease/readiness closure.

  4. Serving migration is deferred until a contract runs from ProblemFrame alone and passes typed derivation, verifier, morphology, confuser, replay, and parity evidence.

  5. Parser retirement is one organ at a time. Migration is incomplete until the raw parser and no-new-legacy allowlist entry are deleted.

  6. ProblemFrame, math state, EpistemicGraph, and PropositionGraph remain purpose-specific. Typed adapters connect them; no universal graph is introduced.

  7. ADR-0223 status is stale (Proposed) despite the merged audit. Architect ratification/status reconciliation is required before the next implementation PR merges.

Invariants reviewed

Invariant Result
versor_condition(F) < 1e-6 Not mechanically perturbed; no field/algebra/runtime code changed.
Exact CGA only Plan requires exact O(N) candidate scans and forbids ANN/HNSW/cosine substitution.
Normalization boundaries No normalization proposed outside approved construction/algebra boundaries.
Wrong-zero serving Every promotion gate retains wrong_ids == []; false-runnable removal counts as progress.
Reviewed mutation Catalog/pack changes remain checksummed, reviewed, data-only; runtime learning is proposal-only.
INV-21/24/29/30 No mutation or open-/closed-world determination path touched or widened.
Graph separation ADR-0144 and ADR-0207 boundaries are explicit in the plan.

No subagents were used.

Exact next PR

branch: codex/problemframe-proportional-change-closure
title:  feat(kernel): close proportional-decrease contracts and make readiness obligation-sound
agent:  Codex or Opus, high reasoning; mandatory XHIGH merge-blocker review

Required outcome:

  • bind decrease_to_fraction for train 0005 from exact facts/spans;
  • represent a typed decrease-delta question target;
  • make fraction_decrease diagnostically runnable for 0005;
  • retain percent_partition runnable for 0046 only under explicit topology;
  • make holdout 0393 non-runnable for forward percent partition;
  • keep serving unchanged and wrong_ids=[].

Forbidden scope:

  • generate/derivation/*
  • generate/math_candidate_graph.py
  • serving changes or raw-text fallback
  • report, case, sealed, pack, or manifest changes
  • case-ID behavior, generic parsing, hidden normalization, or approximate retrieval

Expected diagnostics:

train runnable:   1 -> 2 (0005, 0046)
holdout runnable: 1 -> 0
train serving:    30 correct / 20 refused / 0 wrong
holdout serving:   5 correct / 495 refused / 0 wrong
wrong_ids:        []

Implementation entry prompt

Use the full exact prompt in §14 of docs/analysis/semantic-substrate-problem-solving-master-plan-2026-06-20.md. Before editing, the implementer must state the exact files and explain how the new relation/target/contract types make false readiness unrepresentable. At handoff, list every positive obligation, confuser, metric, and forbidden-diff check proven.

Do not forget

The immediate PR closes diagnostic obligations; it does not complete semantic substrate retrieval. Do not call bounded local construction matching “geometric” or “substrate-mediated.” The reviewed construction catalog and exact CGA proposal seam are later, separate PRs.