First Decision Reproducibility.
The verb shifts. Reproduce.

01The killer query — reproduce_decision(decision_id)

actor:        princ_credit_risk_agent_001
capability:   recommend_credit
subject:      loan_84711_borrower_principal
policy_ref:   pol_credit_underwriting:1
model:        model_credit_underwriting v1
              (full ModelInfluenceRecord from #174 attached)
threshold:    0.75
responsibility chain: present (actor + supervisor + asset_owner + …)
outcome:      recommend_approve

CONFIDENCE:   82/100  (3 of 5 components fully anchored)

02The replay object — what gets reproduced

03The money quote

04The pattern — three money quotes, one corpus

Eric LOCKED June 3 2026: "You now have a recognizable pattern across the corpus."

Each proof's deepest claim arrives with its own honest limit. The pattern says: H33 produces evidence — not verdicts.

05The Reproduction Confidence — measurable rather than rhetorical

06The schema (Eric LOCKED Option C — two tiny additive fields)

"This is one of the rare cases where two tiny schema additions buy a huge amount of future value. Without anchors, you're inferring. With anchors, you're measuring."

PolicyRegister  + policy_ast_hash: Option<String>
PolicyAmend     + policy_ast_hash: Option<String>
Decision        + inputs_hash:     Option<String>

Both fields are Option<String>, both skip-if-none, both backward-compatible. All four prior canonical-continuity-tenant proofs verified byte-identical state_ids under the extended schema.

Plus the data shapes that compose the helper:

enum ReproductionComponentStatus { FullyAnchored, PartiallyAnchored, PointerOnly, Missing }
struct ReproductionComponent      { component, status, explanation, score, max_score }
struct DecisionReproductionConfidence { total_score, max_score, components[], caption }
struct DecisionReproduction        { decision_*, inputs*, policy_*, model_influence,
                                     decision_threshold, responsibility_chain, outcome,
                                     confidence }

The reproduce_decision(snapshot, decision_id) helper lives in the test harness — it is a composition over existing snapshot fields, keeping the engine surface stable while surfacing the orthogonal-axis affordance.

07The computation axis — where this proof sits

08The canonical continuity tenant — four dimensions, one reality

09What this proof IS and IS NOT

10Honest limits (Eric LOCKED — 5 total)

1. Confidence is a measurement, not a guarantee. 82/100 means 18 points worth of components are anchored partially or by pointer only — and the page names which.
2. Perfect reproduction is often impossible. Stochastic models, randomized policy engines, time-dependent inputs, deprecated dependencies. The proof scores reproducibility measurably, not aspirationally.
3. Reproducibility is not Determinism. Two reproductions may legitimately disagree if the original decision was non-deterministic. The proof captures what was bound at decision time.
4. Phase E lock open. Per-event signature verification remains the standing honest-limit from L9.
5. Reproducibility ≠ Justification — see section 03 above.

11Evidence appendix

12Readiness determination & strategic pause

What this unlocks: an auditor — and every other audience inside that umbrella — can now ask "Can you reproduce this?" and receive a deterministic, scored, byte-identically replayable answer. The category Eric named is now standing: Decision Reproducibility. The pattern across the corpus is now recognizable.

What this does not unlock: justifications, fairness verdicts, causality claims, or guarantees of perfect reproduction. Reproducibility ≠ Justification.