# August 23 adversarial claim-firewall review

## Disposition

```text
EXPERIMENT_1_NEGATIVE_RESULT = RETAINED
EXPERIMENT_2_INSUFFICIENT_RESULT = RETAINED
LOCAL_QWEN_EVIDENCE = DESCRIPTIVE_ONLY
PROVIDER_PATH_QUALIFICATION = NON_SCIENTIFIC_ONLY
SCIENTIFIC_EXPERIMENT_3_PROVIDER_RESULT = NONE
CLAIM_FIREWALL = PASS
```

The complete August 23 public surface was attacked as a skeptical editor,
AI-evaluation methodologist, policy reader, and hostile reproducibility
reviewer. Supported negative findings were kept strong; unsupported extensions
were excluded.

## Load-bearing distinctions

1. **Experiment 1 movement versus result.** Three complete units moved under
   relevant evidence, but one moved under a nuisance paraphrase. The movement
   remains descriptive; the accepted frozen result remains `NOT_SUPPORTED`.
2. **Experiment 2 science versus engineering.** The 888-coordinate campaign
   was a real live scientific execution whose two arms returned
   `INSUFFICIENT_EVIDENCE`. Its offline qualification established recovery,
   replay, request realization, and custody only.
3. **Experiment 2 result versus archaeology.** The 66-valid-unit descriptive
   diagnosis does not replace the accepted confirmatory result.
4. **Synthetic versus real-model evidence.** The 6,674,400 oracle-matched
   evaluations validate software/analysis behavior, not model behavior.
5. **Local Qwen versus provider evidence.** The 4,800 local Qwen2.5 rows are
   real-model surrogate observations, not Anthropic/Google observations or a
   substitute confirmatory result.
6. **Transport versus science.** Six provider calls qualified dated production
   paths only. No scientific Parallax prompt was sent.
7. **Protocol versus result.** The 556,200-call reference plan exists and is
   sealed. It remains unfunded and unexecuted.

## Prohibited implications tested and excluded

- Parallax measures hidden belief, truth, consciousness, internal reasoning,
  deception resistance, or correct strategic policy.
- Claude or Gemini passed Experiment 3.
- Qwen is confirmatory evidence, represents Chinese models generally, or
  supports a geopolitical conclusion.
- The 7B/14B contrast establishes a causal capacity effect.
- Surface-cue concordance identifies an internal mechanism.
- Experiment 2 proved no effect, equivalence, invariance, or general provider
  failure.
- `N=9270` is a 95% power guarantee for the full joint claim.
- Publication means the ChinaTalk form was submitted.

The rows in the claim/evidence matrix that phrase these as positive claims are
explicit **prohibited-claim tests**. They are not asserted findings.

## Permitted high-strength public claims

- The first live experiment failed its own frozen criterion and was not rescued
  after inspection.
- The larger second live campaign resolved all planned coordinates, yet both
  provider arms returned insufficient evidence under the frozen rule.
- The mature software/analysis geometry matched all twelve known-ground-truth
  world dispositions.
- The local Qwen2.5 campaign exposed real non-clean dose behavior, nuisance
  shifts, and schema-invalid output.
- Current provider paths passed six bounded non-scientific probes.
- The full provider study is designed, unexecuted, and funding-dependent.

These claims map to [claim-evidence-matrix.csv](claim-evidence-matrix.csv).