# Dispositions: outside read of v2.4 and Code It Yourself (5 October 2026)

**Read:** `20_read_astra.md` (GPT-6 Astra). Release: ADOPT AMENDED. Page: HOLD. **Editor:** Claude Opus 5.5, who applied every finding in session (the agent budget for the night's steps was spent: USD 7.97 of USD 8).

| # | Finding | Disposition |
|---|---|---|
| 1 | Page: labels computed from group summaries can differ from the full analysis near a threshold (reader's example: 5.030 Supported against 4.964 Weakly supported) | Accepted. Every computed label now reads "Approximate: …", the line under it says it is an approximate rule output from group summaries and no H1 verdict, and a near-threshold warning appears within 0.2 points of any boundary; the explanatory text says the label can differ near a threshold |
| 2 | Release: state.json's `whatV24Is` omitted that the analysis leans toward the prediction under the chosen pairing | Accepted; wording as proposed |
| 3 | Page: a relative difference presented as an absolute lean; the scale description was wrong for a difference | Accepted. Opening question, "Positive means…" (both groups may still score higher on awareness) and unit line ("each group's score runs from −100 to +100") replaced |
| 4 | Page: the sampled codings are not validated interpretations | Accepted; the record's caveat restored (problematic codes kept, unproposed alternatives excluded, uniform weighting, a visitor's sorting can fall outside the set) |
| 5 | Page: the blind recoding's qualification dropped | Accepted; the unfiled paper-priority instruction and the 63 inherited questions stated |
| 6 | Page: provenance wording | Accepted; two presets named as constructed examples; "no full questionnaire items; short labels for the items are shown" |
| 7 | Scope | No change needed: no claim, tier, confidence or falsifier moves; no surviving statement of the old pairing |

Because the page was held, it gets a second, narrow read before the branch goes to the author (`40_reread_astra.md`).

## What would count against this

- The near-threshold margin of 0.2 points is the editor's choice; the observed gap between the page and the full analysis is up to about 0.1 points, and a larger gap on some coding would defeat it.

## Second read (`40_reread_astra.md`): ADOPT AMENDED

Both remaining amendments accepted and applied: (1) the accuracy statement now says the browser's estimates differ from the reference values by under 0.1 points for the four recorded presets only, sets no error bound for other codings or interval ends, and calls the intervals and labels approximate; the first "What would count against this" bullet matches; (2) the Review section states both reads and links them.