# First Model Influence Replay — the orthogonal axis

**Proof #174 — June 3, 2026 — issued by H33, Inc. — Eric Beans, CEO**

> **"Why did the model matter?"**
> The hero question Eric Beans named for the first proof on H33's orthogonal axis: continuity of computation, distinct from the continuity-of-evidence ladder.

---

## The new axis

| Axis | Question | Status |
|---|---|---|
| Continuity (the ladder, vertical) | "What survives change?" | 5 lines proven |
| **Computation (the branch, orthogonal)** | **"Why did the computation happen?"** | **#174 (this proof) → #167 next** |

Candidate orthogonal endpoint Eric named: **"Reasoning survives systems."**

---

## Money quote (LOCKED Eric Beans)

> **Influence ≠ Causation.**

The proof establishes **that** the model's score depended on these features in this way. It does **not** establish that the model **should have** depended on them.

Parallel structure to #184's "Institutional Memory ≠ Legal Truth." Both quotes define what the corpus does *and does not* claim.

---

## Killer query (LOCKED)

```
asl> show_model_influence("decision_loan_84711_recommendation")

model:                 model_credit_underwriting v1
attribution_frame:     feature_weights
prediction_score:      0.84
decision_threshold:    0.75  (above by +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
```

---

## Scenario timeline

| When | What |
|---|---|
| 2025 | AI risk agent + human credit officer approve $4.2M loan. (#15 substrate.) |
| 2026 | Borrower defaults. Loss + Claim #84711. (#15 substrate.) |
| 2031-03-08 | #15 replay window. Influence record sits AFTER this T → #15 byte-identical. |
| **2031-03-19** | **Model owner signs ModelInfluenceRecord** — retroactive instrumentation pre-dissolution. |
| 2031-04-01 | Enterprise dissolved. (#184 capstone.) |
| 2035-06-01 | Court trustee replays — `show_model_influence` returns the full attribution. |

The same tenant (`tenant_insurance_claim_44962d9b-…`) now carries #15, #184, and #174 simultaneously. Eric's locked **canonical continuity tenant**: compounding proofs against the same underlying reality is much stronger than inventing new demos.

---

## Schema (LOCKED Option A — standalone)

```rust
ModelInfluenceRecord {
    at_ms,
    influence_id,
    decision_id,
    model_id, model_version,
    attribution_frame,
    feature_attribution: Vec<FeatureAttribution>,
    prediction_score, decision_threshold,
    counterfactual_probes: Vec<CounterfactualProbe>,
    recognized_by, signature,
}
```

Standalone because 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` — sorted by `(at_ms, influence_id)`, skip-if-empty. `Eq` dropped from AuthEvent / snapshot / ReplayResult because f64 fields (score, threshold, weights) don't implement Eq.

---

## Audience (Eric LOCKED — primary AI Regulators + ELEVATED Internal Model Governance Teams)

1. **AI Regulators** (SR 11-7 inheritors, EU AI Act, UK FCA)
2. **Internal Model Governance Teams** — *"the people between Legal / Risk / Compliance / Data Science inside banks and insurers. Likely buyers BEFORE regulators arrive."*
3. Model risk officers
4. AI fairness auditors
5. R&W insurers underwriting AI-decided products
6. Class-action plaintiffs alleging AI bias
7. Internal red teams
8. Carryover from prior proofs (trustees, regulators investigating dissolved AI-using firms)

---

## Honest limits (Eric LOCKED — 5, with #5 elevated)

1. Replayable influence is not Replayable model.
2. Feature attribution is a chosen frame; SHAP / LIME / attention / IG can disagree.
3. Counterfactual probes are bound at decision time.
4. Phase E lock open.
5. **Influence ≠ Causation** ← the money quote of the page.

---

## What Survives Model Retirement?

| | What |
|---|---|
| ✓ | model identity |
| ✓ | model influence |
| ✓ | decision (and responsibility chain) |
| ✓ | consequences (loss, claim, lineage) |
| ✓ | responsibility |
| ✗ | model weights (yet) |
| ✗ | model execution (yet) |
| ✗ | full decision reproducibility (yet) |

The ✗ items are #167 Decision Reproducibility's territory — the next proof on the orthogonal axis.

---

## Evidence appendix

| Field | Value |
|---|---|
| state_id at T=2035 | `e72d3c0e71a11ce0aaf1e8c9eb5c720aff49a6238c76976b4f4435b50e43bee2` |
| Tenant | `tenant_insurance_claim_44962d9b-25f5-5622-bd9a-98d5580bb8a2` (canonical continuity tenant) |
| Decision | `decision_loan_84711_recommendation` |
| influence_id | `influence_loan_84711_recommendation_v1` |
| Model | `model_credit_underwriting v1` |
| attribution_frame | `feature_weights` |
| prediction_score | 0.84 |
| decision_threshold | 0.75 (margin +0.090) |
| feature attributions | 5 |
| counterfactual probes | 3 (2 would change, 1 would not) |
| T_INFLUENCE (ms) | 1,932,000,000,000 — 2031-03-19 |
| T_REPLAY (ms) | 2,064,614,400,000 — 2035-06-01 |
| Reconstruction artifact | [reconstruction.json](reconstruction.json) |
| Harness | `tests/model_influence_replay_001.rs` |
| scif-backend SHA | `a27d59872` |

---

## Readiness determination

**First Model Influence Replay: PROVEN IN OPERATION.** The canonical continuity tenant now carries a signed model influence record; `show_model_influence` returns the full attribution + counterfactuals; the chain reconstructs four years post-dissolution.

**What this unlocks:** the first proof on the orthogonal axis. Future AI regulation will require this surface; internal model governance teams will buy it before regulators arrive.

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

---

**Issued by H33, Inc.** · Eric Beans, CEO · 2026-06-03
Independently reconstructable. Inputs: scif-backend @ `a27d59872` · `tests/model_influence_replay_001.rs` · [reconstruction.json](reconstruction.json).