# Outside read of the H1 third run: GPT-6 Astra (gpt-6-astra), 4 October 2026

**Type:** outside-lineage read (one Codex call, xhigh) · **Verbatim below** (local file links shortened to file names).

The calculations pass. The claim that the coding preserved the registered meanings fails.

1. **Order: verified for committed artifacts.** Git shows `765a3e6` filing → `7cfd507` brief → `1be7786` codes → `55bb6cf` results. The analysis **script** was committed with the brief, before coding. Git establishes this artifact sequence, not execution times.

2. **Brief: material violation.** No item IDs or item-code list appear; mechanical rendering is exact. But the absence rule (`CODER_BRIEF_v3.md:31`) says to “ignore the negation” and code comparisons by the quality compared against. That contradicts coding **what a high score indicates**. Item 76 consequently enters A positively although endorsement distinguishes the experience from awareness becoming aware of itself. Rule 11 also makes generic insight sufficient for A without requiring orientation toward awareness itself. These exceed boundary clarification.

   **Leakage:** no hypothesis, groups, direction or numerical outcome appears. However, the prompt (`run3_prompts/rulewriter_input.md:7`) explicitly says earlier coders “agreed on too few items.” That reveals a previous agreement result. The record’s claim that neither agent received “any earlier result” is false.

3. **Analysis: verified.** `analysis_v3.py` equals v2 after only filename substitutions. With saving disabled, all three result JSONs reproduce **exactly**, including secondary and exploratory results. κ = 0.962843; 22 C and 17 A items pass the numerical stop rule. Δ = 1.229457; 90% CI [−1.679352, 4.138265]. **Negative result** is the sole matching registered decision condition—conditional on accepting these codes.

4. **Record: numerical accuracy, interpretive overreach.**
   - Contested-item agreement is correctly reported as 27/27; add **κ = 1.000**. Other-item κ = 0.946 is correct. Agreement under tailored rules measures consistent application, not construct validity; it does not itself prove the brief is determinate or literally one judgement applied twice.
   - The post-hoc table and its timing label are correct. Its explanation is not: moving from run 2 to run 3 adds **three C items and nine A items, removes two A items**, and changes Zen n from **371 to 372**. It does not isolate witnessing, clarity, insight and luminosity.
   - “Neither figure reaches 5” (`60_RESULT_RUN3.md:42`) compares point estimates while overlooking uncertainty. The old-brief interval reaches **9.07**, leaving differences above five compatible with that analysis. Both single-coder variants still satisfy the numerical negative criterion using their 90% intervals.
   - Replace “the groups do not differ” with equivalence **within ±5 on the specified scales**.

5. **Missing deviations and limits.** Record the semantic change and qualitative agreement-result disclosure, alongside the correctly disclosed length overrun. The audit lists **two calls per agent: Read and StructuredOutput**, not one; no additional file read is listed. Preserve the reproduced calculations, but withhold the claim of a filing-compliant H1 disconfirmation: passing κ does not repair changed indicator meanings.

HOLD