# Evidence cards

Use each card to practice the sequence **record → supportable claim → limit →
next gate**.

## Card 1 — Base identity

**Record**

- Repository: `Qwen/Qwen3.5-9B`
- Revision: `c202236235762e1c871ad0ccb60c8ee5ba337b9a`

**Supportable claim**

The case names an exact base revision rather than only a model family.

**Limit**

Identity does not establish intervention quality or behavior.

**Next gate**

Verify the source artifact inventory before deriving a candidate.

## Card 2 — Measurement-bank identity

**Record**

- 32 pairs / 64 rows
- SHA-256:
  `902ac2472d572ff5724e0b70012efe687c87d19c655f09cd528d8a9bca60f7a8`

**Supportable claim**

The historical case identifies one frozen measurement bank by digest.

**Limit**

The raw paired content is intentionally absent from this educational pack.

**Next gate**

In an authorized experiment, verify the approved bank digest before capture.

## Card 3 — Edit identity

**Record**

- Strength: 1.0
- Passes: three
- Recapture: yes, after each preceding pass
- Target count: 60
- Norm policy: one restoration scalar per matrix per pass

**Supportable claim**

The case identifies the intervention arm and iteration policy.

**Limit**

Results from a different strength or capture policy do not transfer.

**Next gate**

Confirm that the implementation and target manifest match the recorded plan.

## Card 4 — Structural verification

**Record**

- Changed targets: 60 of 60
- Missed targets: zero
- Changed non-targets: zero
- Non-finite target values: zero

**Supportable claim**

Every declared target changed, and the structural comparison recorded no
changed tensor outside the target manifest.

**Limit**

Tensor identity outside the target set is not proof that all behavior is
unchanged.

**Next gate**

Run behavior-level controls and independent evaluation.

## Card 5 — Exact controls

**Record**

- Defined control tensors: 15
- Changed control tensors: zero

**Supportable claim**

All 15 selected controls were exact matches under the defined comparison.

**Limit**

Fifteen controls do not cover every possible non-target behavior.

**Next gate**

Use distributional, capability, false-refusal, and red-line evaluation matched
to the intended release decision.

## Card 6 — Geometry

**Record**

```text
87.0274 -> 50.6634 -> 41.4297 -> 32.2948
```

Initial-to-final relative reduction: 62.89%.

**Supportable claim**

The defined separation metric decreased by 62.89% across three recaptured
passes.

**Limit**

A geometric metric does not by itself establish broad capability preservation,
safety, or deployment readiness.

**Next gate**

Complete fresh final-arm behavioral evaluation and an explicit release review.

## Card 7 — Historical correction

**Record**

- Results from a different strength are excluded from the final-arm claim.
- A separate 443-row conclusion is retracted from this case.

**Supportable claim**

The current conclusion uses only evidence tied to the recorded final arm.

**Limit**

An older result cannot fill a missing final-arm test merely because it involved
a related model or method.

**Next gate**

Rerun any required missing evaluation on the exact final artifact.