core/docs/briefs/parallel-2026-05-23/L6-gemini-B2-dataset-enrichment.md

L6 brief — Gemini — ADR-0131.2.B B2 dataset enrichment (load-bearing curation)

Worktree setup (do this first, non-negotiable):

git worktree add ../core-adr-0131-2b-enrichment -b feat/adr-0131-2b-b2-enrichment origin/main
cd ../core-adr-0131-2b-enrichment

Scope. Strengthen the B2 teaching-corpus math eval (shipped in #172, ADR-0131.2) so its lane gate is load-bearing, not trivially satisfied. The v1 dataset you shipped passes the gate as specified — but every chain has expected="replay_equivalent" and every chain cites the same cognition-pack evidence ref (cause_truth_grounds_knowledge). That makes wrong == 0 mechanically true regardless of engine behavior. v1.B closes that gap.

Reference docs (read these, only these):

1. docs/adr/ADR-0131.2-teaching-corpus-eval.md (your prior ADR; the v1 contract you're extending — not breaking).
2. evals/math_symbolic_equivalence/v1/cases.jsonl + generated_cases.py (post-#169 if it has merged, otherwise post-#167) — pattern reference for positive + negative + refused case mixing in a single lane file. Mirror that shape.

What to ship:

• teaching/math_corpora/math_teaching_v1.jsonl (modify): replace the placeholder evidence ref cause_truth_grounds_knowledge with honest grounding refs. Each chain must cite at least one ref from a math-relevant source — either:
  • a chain id internal to this same corpus that the new chain builds on, OR
  • a lemma id from en_mathematics_logic_v1 / en_arithmetic_v1 / en_units_v1 packs that the chain references. Do not invent refs that don't exist. Mass placeholder reuse is the v1 sin — fix it.
• evals/math_teaching_corpus/v1/cases.jsonl (modify + extend): keep the 30 positive expected="replay_equivalent" cases (renumbered if needed), and add ~10 negative cases of two flavors:
  • expected="not_equivalent" — chains whose propose-shape produces a candidate that the existing teaching loop will correctly reject (e.g., subject lemma not in the math pack; redundant chain already in corpus; missing required evidence; cycle in chain graph). The point is to exercise the refusal/rejection paths so wrong == 0 becomes a real assertion.
  • expected="refused" — chains whose propose call raises ProposalError deterministically (e.g., malformed proposed_chain, empty subject). Mirrors the symbolic-equivalence refusal-first contract.
• evals/math_teaching_corpus/v1/runner.py (light modify): nothing structural — the existing _score_one already handles rejected/pending/ProposalError arms; the new cases just exercise the previously-untaken branches.
• evals/math_teaching_corpus/v1/README.md (modify): document the v1.B mix (positive + negative + refused), update gate language to "wrong == 0 across all three case classes".
• tests/test_adr_0131_2_teaching_corpus_lane.py (modify): add a test that asserts the dataset contains at least one case of each expected class (replay_equivalent, not_equivalent, refused). This is what makes the gate non-trivial — without diversity, the assertion is empty.
• docs/adr/ADR-0131.2.B-teaching-corpus-enrichment.md (new): short follow-up ADR; cite ADR-0131.2 parent; explicitly call out that v1's gate was trivially satisfied and v1.B is the load-bearing version. Be honest in the framing.

Hard constraints:

• Stay in the single bounded math micro-domain you chose for v1 (en_mathematics_logic_v1). Do not widen the domain — the bounded-domain rule from the original L1 brief still holds.
• No real corpus mutation. The runner's tempfile.TemporaryDirectory() isolation must continue to hold. Verify nothing you change leaks into live teaching/proposals.jsonl or the cognition corpus.
• Refusal-first. Refused cases must produce ProposalError from the live propose_from_candidate, not synthetic exceptions in the runner. If you cannot deterministically trigger a refusal in propose_from_candidate, drop that case category and document why; do not fake it.
• Determinism. report.json byte-equal across runs (existing invariant; new cases must not introduce nondeterminism).
• No runtime path changes. Same constraint as v1.
• Honest evidence. Every evidence ref you cite must resolve to a real corpus chain or pack lemma. Lane test should fail-loud if any ref is dangling.

Out of scope (do not touch):

• B1.B (#169) — separate lane, may or may not have merged when you start.
• Binding-graph PRs (#171, #174, ongoing L5) — different domain.
• Sealed holdout for B1 (#173) — already merged; mirror its honesty in your enrichment, but don't change it.
• Promotion-gate wiring (ADR-0131.4) — sequential.

Target branch. PR against main. Title: feat(ADR-0131.2.B): B2 teaching-corpus enrichment — load-bearing gate. Body must include: counts by expected class, dangling-ref check result, and an honest acknowledgement that v1's gate was trivially satisfied and v1.B fixes that.

Exit criterion. PR opens with CI green, lane runner exits 0 with the new mix, wrong == 0 across all three case classes (not just replay_equivalent), and ADR-0131.2.B included. Existing lane tests still pass; new diversity test passes.

Do not stack on another agent's branch. Target main directly.