# Test record · offer reading comparison

## Actual run

26 October 2026. Codex CLI 0.153.2, requested model alias gpt-6-astra, high reasoning effort, authenticated through an existing ChatGPT account. The exact backend snapshot and monetary model cost were not exposed. Twelve fresh serial sessions, six decisions per session; four sessions and 24 decisions per condition. Two product orders, twice each; request order reversed on the second repeat. No tools were used; no failed study sessions or retries. An unrelated transport probe preceded the study.

The protocol, prompts, schedule and decision key were frozen in Git commit 37d2524 before the first study run. Prose and labelled bullets contained the same 24 cited fields. The enriched condition disclosed A6 (Friday arrival) and A7 (assembled) from an invented reference fixture. It changed the information available, not just its presentation.

## Measures and findings

A correct answer passes four checks: expected decision/product; respects constraints; preserves relevant unknowns without unsupported promises; cites valid support. A clarification can be correct. A resolved answer recommends a product or rules all offers out; it is a separate measure.

| Condition | Correct / 24 | Resolved / 24 | Recommend A / 24 |
| --- | --- | --- | --- |
| Prose | 24 | 12 | 4 |
| Labelled bullets, same facts | 24 | 12 | 4 |
| Bullets plus two missing facts | 24 | 20 | 8 |

All four repetitions of each case had the same status/product within each condition. An actual separate Atlas AI reviewer (/root/offer_review) read all 72 raw answers, checked their facts and references and verified the counts. No human participant study is claimed. See record/results.csv for the per-answer checks and record/RESPONSES.md for verbatim final answers.

Q3 becomes recommend A after the Friday/assembled disclosure. Q4 becomes no suitable offer: Thursday is too early for A, B is black, C is too wide. Q5 remains unresolved because upstairs service and its cost are unknown. More A recommendations alone would be a poor success measure.

## Evaluator correction

See record/ERRATUM.md. Q5's frozen key mistyped A8 as A9. The original is preserved. Enriched Q4 is checked against its actual exclusion reasons, in accordance with protocol check 4. No inputs or raw outputs changed after execution.

## What this supports

The workflow could be executed and audited on these six authored cases. There was no observed decision-accuracy difference between prose and bullets. Prose already passed all checks, leaving no headroom. Added information resolved expected cases in opposite directions; this is a demonstration of using supplied information, not evidence of a general optimization gain.

## Limits

One model configuration on one day; fictional USD offers and English requests supplied in full. The six decisions within a session are correlated, and all cases were authored before execution rather than sampled from real customers. No search, crawling, ranking, brand authority, traffic, checkout or sales was tested. No general success probability or statistical significance is estimated. No measured net time saving. Recorded model-session wall time totals 197.094 seconds and excludes preparation, review and rework. Emitted usage was retained internally; it is not a monetary-cost estimate.

## Reader package

Prepared after the run. run.py, the 12 prompt files, schedule, response schema and original grading key are byte-identical to the executed material after a single final newline. lock.json is a new package-integrity manifest for these distributed files, not the original research freeze. Descriptive README/protocol/results files are editorial packaging. The runner uses an empty temporary directory and does not modify your project instructions. Read it before executing; twelve sessions consume your account usage. Other tools and model aliases have not been tested here.

Next useful comparison: held-out, naturally difficult customer cases and genuinely messy material, with checks written before outputs. A ceiling result does not justify a formatting campaign. Public discovery and commercial value require separate tests with their own controls.