feat(adr-0249): P5 arithmetic lift on real GSM8K + ADR-0249 (arc closure) #70
Loading…
Reference in a new issue
No description provided.
Delete branch "feat/adr-0249-p5-instrument"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
The closing phase of the reader→Hamiltonian compiler arc: it measures the corridor on real GSM8K and lands ADR-0249 (Proposed) with acceptance evidence. P1–P4 are merged (#66–#69); this is P5 + the ratification docs.
P5 — arithmetic-chain lift domain (real data)
evals/generalized_lift_instrument.pygains anarithmetic-chaindomain: the corridor's turn-program executor vs a symbolic fold of the same compiled program, over the real GSM8K dev holdout (evals/gsm8k_math/dev, using each case'sground_truth_graph— never the templated cases).Honest result on the sealed 50-case holdout:
This is the honest-NULL protocol working as intended: the corridor solves every real GSM8K problem it can ingest with zero wrong answers, matches the symbolic fold (the field doesn't beat arithmetic at arithmetic — PARITY is honest, not an inflated lift), and honestly refuses the multi-entity remainder. The deliverable is a decidable measurement + recorded real-data coverage. The stale "no reader-to-Hamiltonian compiler exists" scope note is corrected;
wrong_zero_guard_heldnow binds both exact-regime domains (deductive flagship + arithmetic).ADR-0249 (Proposed — no self-Accept)
docs/adr/ADR-0249-*.mddocuments the whole arc: the composition-as-certified-turn-sequences thesis, the five components, anti-hollow discipline, the Ring-2 non-mutating correction (the residual protocol is zero-bound, so the turn ledger reuses its chain-integrity pattern, not the admission protocol), the three-tier reproducibility bound, Tier-1 scope + recorded frontier, and the honest-measurement done-when.docs/handoff/ADR-0249-Acceptance-Evidence.mdcollects the phase SHAs, results, and in-tree verifications. §10 ruling record awaits your ratification.Governance
Off-serving (A-04); no gate activated (ADR-0246 §3.7 stays uncalibrated); no manifold mutated (I-03); ratified contracts reused, not forked. Every phase smoke-gated green in-worktree.
Tests
tests/test_adr_0249_arithmetic_lift.py— 10/10 (5 new + 5 existing instrument tests still green). Real-GSM8K wrong=0, every-ingested-case-correct, PARITY-with-symbolic-fold, coverage recorded (26/24 tracked), domain joins the full report + scope corrected.[Verification]:
uv run python -m pytest tests/test_adr_0249_arithmetic_lift.py tests/test_generalized_lift_instrument.py -q→ 10 passed;uv run core test --suite smoke -q→ 176 passed.Wires the turn-program executor into the generalized-lift instrument and measures it against the REAL GSM8K dev holdout (evals/gsm8k_math/dev, never the templated cases), using each problem's ground_truth_graph. Baseline is a symbolic fold of the SAME compiled turn program: both paths consume identical problems, the corridor relaxes each step, the baseline folds it numerically. Honest result on the sealed 50-case dev holdout: 26 problems are Tier-1 affine single-accumulator ingestible and solved wrong=0; the corridor matches the symbolic fold on every one (PARITY, delta=0) — the field matches arithmetic, it does not beat it, so this is real-holdout coverage, NOT a manufactured lift. The 24 refused (all not_single_accumulator / multi-entity) are the recorded Tier-2 frontier, never silently dropped. Corrects the now-stale scope-limitation note ('no reader-to-Hamiltonian compiler exists…') — the compiler exists as of this arc; the note now records real-data Tier-1 coverage and the multi-entity frontier. wrong_zero_guard_held extended to bind both exact-regime domains (deductive flagship + arithmetic). 10/10 pins (5 new + 5 existing instrument tests still green). [Verification]: uv run python -m pytest tests/test_adr_0249_arithmetic_lift.py tests/test_generalized_lift_instrument.py -q