First Wave M / Phase B piece (GATING): read-only backend that makes the
calibrated-learning / serving-discipline loop inspectable — 'the engine
earns the right to guess', ADR-0175.
The workbench computes NONE of these numbers:
- GET /calibration/classes — per-class gold-tether view from the persisted
practice arena ledger (evals/gsm8k_math/practice/v1/report.json per_class).
Each class's reliability_floor is the engine's own one-sided Wilson
conservative_floor (via ClassTally.reliability); PROPOSE (θ=0.85) / SERVE
(θ=0.99) license verdicts come from core.reliability_gate.license_for.
Failures-first ordering. A test proves the reader's floor equals a direct
conservative_floor() call — no re-implementation.
- GET /serving/metrics — the live correct/refused/wrong counts read unchanged
from the committed train_sample + holdout_dev report.json (currently
4/46/0 and 5/495/0 — wrong=0). Never re-runs a lane.
Honest current state: the committed practice ledger's three classes
(additive/divisive/multiplicative) are all below N_MIN=10, so none has
earned a license yet — the reader shows exactly that, no fake green light.
- workbench/calibration.py: pure readers; imports core.reliability_gate;
EvidenceUnavailableError -> 501 (fail-closed) when the artifact is absent.
- schemas + TS mirrors (CalibrationClass, ServingMetrics); both snapshots
regenerated (deterministic); both drift gates pass.
- trust boundary: read-only over committed artifacts + engine-owned
derivation; no execution, no mutation, no license ever changed.
Verified: 30 Python tests (incl. the no-reimplementation proof + fail-closed),
390 vitest, both schema drift gates, snapshots deterministic.
First R3 (Theater) piece. Reworks the stale artifact-keyed ReplayRoute into
the turn-keyed hero over the #716 sealed-replay backend, and fully retires
the dead W-026 artifact-keyed machinery on both sides (the NOT_YET_MIRRORED
comment anticipated this).
The hero makes determinism *felt*: pick a journaled turn, CORE re-executes
its prompt in a sealed fresh runtime, and the result renders as a hash
verdict (≡ equivalent / ≠ diverged) + the original/replay DigestBadges + a
leaf diff (critical weighted above informational, each with a ≠ glyph). The
'What this proves' card surfaces comparison_basis + origin_state and states
the honest limit: a divergence means nondeterminism OR origin-state
influence, never rendered on its own as a determinism failure.
Retirement (verified zero serving uses):
- Python schemas.py: removed ReplayComparison/ReplayDivergence/ReplayStatus/
ReplayDivergenceSeverity; scripts/dump-enums.py drops the two replay enums;
both snapshots regenerated (deterministic: dump == committed)
- TS: removed ReplayComparison/ReplayDivergence/ReplayDivergenceSeverity;
added TurnReplayComparison/TurnReplayDivergence (+ basis/origin/severity
unions); NOT_YET_MIRRORED now empty (every engine schema mirrored)
- badges: removed ReplayStatusBadge + ReplayDivergenceSeverityBadge +
meta/enums + their enumCoverage cases (hero renders severity inline)
- api: fetchReplayComparison/useReplayComparison -> fetchTurnReplay/
useTurnReplay (/replay/<turnId>)
- deleted ArtifactList, ReplayComparisonPanel, ReplayDiffViewer,
ReplayMetadataTable, old replay.test.tsx
- App route /replay/:artifactId? -> /replay/:turnId?; conformance row
turn-keyed (loading 'Loading turns...', empty 'No turns recorded yet.')
Tests: ReplayRoute.test.tsx (equivalence hero + honesty card, diverged
critical leaf, informational-only label, replace-mode select, j/k spine).
Full vitest 358 green; workbench Python 34 green; both snapshots deterministic.
Follow-up flagged (not in this PR to keep it focused): the artifact query
hooks (useArtifacts/useArtifactDetail/fetchArtifacts*) are now orphaned but
entangled with the still-live artifact EvidenceSubject kind — a separate
cleanup once an Artifacts route or that subject's fate is decided.
Brief: docs/handoff/wave-R3-briefs-2026-06-13.md (all four R3 pieces).
Replaces the W-026 501 stub. Re-executes a journaled prompt in a sealed
fresh runtime (ChatRuntime(no_load_state=True): no checkpoint load, no
checkpoint write, no proposal lineage) and compares the envelope
leaf-by-leaf against the recorded TurnJournalEntry.
Design correction vs the scoping doc (amended in-doc): journaled turns
each ran in a fresh ChatRuntime(), never one continuous session, so
genesis-PREFIX replay would manufacture spurious divergence; shipped
basis is sealed_fresh_runtime_single_turn (O(1)). origin_state:
"unrecorded" — the journal does not record whether the original turn's
runtime loaded a checkpoint, so divergence is reported as nondeterminism
OR origin-state influence, never disambiguated.
- workbench/replay.py: pure comparison + injected executor; every
TurnJournalEntry field classified critical/informational exactly once
(exhaustiveness enforced by test)
- api.py: route wiring under _CHAT_TURN_LOCK; runtime failure -> 500
runtime_unavailable, no comparison may be fabricated
- schemas: additive TurnReplayComparison/TurnReplayDivergence (W-026
artifact-keyed pair retires with the frontend Replay Moment PR)
- tests: 10 obligations incl. tamper-prompt, tamper-leaf precision,
no-execution-no-comparison, no-trace (journal + engine_state bytes),
wall-clock tolerance, sealed-construction proof
- snapshot regenerated; NOT_YET_MIRRORED debt entries for the two new
classes (mirrors land with the frontend PR)
- api-contract-v1.md § Replay rewritten for the turn-keyed shape
Append-only JSONL journal records the exact ChatTurnResult envelope returned by /chat/turn with stable turn_id, trace_hash, all three surfaces, verdicts, and deterministic journal_digest.
GET /trace/turns and GET /trace/{turn_id} serve journal evidence for the Trace route frontend. Read model only; no teaching, pack, or journal-owned engine_state mutation.
Wires teaching/math_proposals/proposals.jsonl into the CORE Workbench
API (ADR-0160) alongside the existing cognition proposal queue:
workbench/schemas.py
- MathReasoningStep, MathProposalSummary, MathProposalDetail,
MathRatifyResult schemas
workbench/readers.py
- MATH_PROPOSALS_JSONL + _DEFAULT_MATH_AUDIT_PATH constants
- teaching/math_proposals added to ALLOWED_ARTIFACT_ROOTS
- _HANDLER_DISPATCH table (vocabulary_addition→LexicalClaim; all
others not yet implemented)
- list_math_proposals(), read_math_proposal(), ratify_math_proposal()
- read_math_proposal() re-runs decompose_audit() to recover full
4-step reasoning trace (canonical_bytes only carries trace_id)
- ratify_math_proposal() raises NotImplementedError with clear
"handler not yet implemented: {change_kind}" for unhandled kinds
workbench/api.py
- GET /math-proposals, GET /math-proposals/{id}
- POST /math-proposals/{id}/ratify → _math_ratify()
(vocabulary_addition→200/routed; unhandled→501 with loud message)
tests/test_adr_0172_w4_workbench_e2e.py — 6 tests:
1. loads from JSONL
2. renders domain:math badge (distinct from cognition /proposals)
3. ratify-vocabulary_addition routes to LexicalClaim (200)
4. ratify-matcher_extension fails loudly (501 "handler not yet
implemented")
5. all 4 trace steps visible in detail response
6. no cross-contamination between math and cognition queues
teaching + runtime suites green (28 + 20 passed).
Brief-gap note: canonical_bytes() excludes proposal_id and serialises
evidence pointers as hashes only. D1 loader derives proposal_id via
sha256(line_bytes) and re-runs decompose_audit() to recover full trace
for read_math_proposal(). This works but means the JSONL cannot be
loaded without the original audit file. If a future wave needs
standalone JSONL loading, C1 should emit a richer format.
