First Model Influence Replay.
The first proof on the orthogonal axis.

01The killer query — show_model_influence(decision_id)

model:                 model_credit_underwriting v1
attribution_frame:     feature_weights
prediction_score:      0.84
decision_threshold:    0.75
score_above_threshold: true (margin: +0.090)

feature_attribution:
  debt_to_income            value=0.42   weight= +0.310
  credit_utilization        value=0.78   weight= +0.220
  employment_tenure_years   value=8.2    weight= +0.180
  payment_history_pct       value=0.96   weight= +0.120
  loan_amount_to_income     value=10.5   weight= -0.210

counterfactual_probes:
  if loan_amount_to_income = 5.0  → score 0.68 → WOULD have changed outcome
  if credit_utilization    = 0.30 → score 0.91 → would NOT have changed outcome
  if debt_to_income        = 0.65 → score 0.61 → WOULD have changed outcome

02The feature attribution, visualized

Sign convention: positive weight pushed toward approval; negative weight pushed away. Sum is not necessarily prediction_score — that depends on the attribution frame. In this proof: feature_weights.

03Counterfactual probes bound at decision time

Three probes signed into the influence record at decision time. Each probe asks: "if this one feature had this other value, what would the score have been, and would the outcome have changed?"

04The money quote

05The scenario — why this scenario carries

This is the strongest possible compounding of Eric's canonical continuity tenant: one tenant, one Claim #84711, five proofs (#15 + #184 + #174 now) all reading the same underlying reality. The model's reasoning survives the dissolution that survived the responsibility that survived the loss that survived the decision.

06The schema (Eric LOCKED Option A — standalone)

ModelInfluenceRecord {
    at_ms,
    influence_id,
    decision_id,                          // back-reference to the Decision
    model_id, model_version,
    attribution_frame,                    // "feature_weights" | "shap"
                                          // | "integrated_gradients" | "attention"
                                          // | "regulator_approved" | custom
    feature_attribution: [
        { feature_name, value, weight }
    ],
    prediction_score,
    decision_threshold,
    counterfactual_probes: [
        { feature_name, alternative_value, resulting_score, would_have_changed_outcome }
    ],
    recognized_by,                        // model_owner (or regulator for approved frames)
    signature
}

Standalone — not embedded on Decision. Reason: tomorrow SHAP / IG / attention / regulator-approved / internal attributions can all attach to the same decision. Decision → ModelInfluenceRecord is 1:N.

Snapshot extension: AuthorityStateSnapshot.model_influences: Vec<ModelInfluenceSnapshot>, sorted by (at_ms, influence_id), skip-if-empty.

07What Survives Model Retirement?

08What this proof IS and IS NOT

09Honest limits (Eric LOCKED — 5, with #5 elevated to money quote)

1. Replayable influence is not Replayable model. We record the influence pattern; we do not (yet) re-run the model from weights. That is #167.
2. Feature attribution is a chosen frame. SHAP, LIME, attention, IG — real, and sometimes disagree. The proof records the frame bound at decision time; it does not claim that frame is the only true read.
3. Counterfactual probes are bound at decision time. Probes years later that weren't pre-signed require model reproduction (#167) or a new probe.
4. Phase E lock open. Per-event signature verification is the open lock from L9; L9.1 closes it.
5. Influence ≠ Causation — see section 04 above.

10Where this proof sits — the orthogonal axis

This is not a rung on the continuity ladder. It is the first proof on a new axis. The corpus now reads in two dimensions:

11Evidence appendix

12Readiness determination

What this unlocks: an AI regulator, an internal model governance team, a fairness auditor, a model risk officer, or a class-action plaintiff alleging AI bias can now ask the deepest question that follows an AI decision — "Why did the model matter?" — and receive a deterministic, signed, byte-identically replayable answer. The first proof on the orthogonal axis is now standing.

What this does not unlock: fairness verdicts. Bias determinations. Causal claims. Model re-execution. Those wait for #167. Influence ≠ Causation.