# Staged reporting and sampled human review, v1 Authorized by Joy on October 13, 2026 (America/New_York), after substantial Mistral response inspection and provisional coding, but before aggregate outcome tables were computed. This is a disclosed post-inspection amendment, not a preregistered decision. It changes reporting order and human-review workload; it does not change the frozen prompts, generated samples, gold, outcome definitions, AI annotation coverage, or the full experiment scope. ## Reporting stages 1. Release an explicitly provisional, AI-coded Mistral core report now: every E3_primary response (180) and every E5 response (192). Preserve the original factor cells, ambiguity bounds, neutral outcomes and faithful controls. 2. Prioritize full-text AI annotation of all180 Qwen primary responses, then add that complete model/module to the same primary comparison. Do not select responses because of their outcomes or count missing labels as negative. 3. Gemma collection and all extensions remain in scope. Add completed, audited, fully coded modules in later versions, without holding up the core findings. The complete-study report retains its existing coverage/reason-audit gates. A separate preliminary command may report only explicitly selected complete model/module slices, with exact audited source and label bindings. It must not claim full-study completion. Before human review, every table and finding must say provisional AI coding; zero human verification must remain explicit. ## Reduced human review This supersedes spec10.7's all-E3/manual E1 workload for reporting under this amendment. All responses still receive full-text AI coding and independent AI review; those passes are never described as human annotation. Human review uses an outcome-independent stratified base sample **plus** every disputed case relevant to a reported claim (including resolved AI disagreements), every uncertain/mixed primary label relevant to that claim, every candidate strict UOS event, and every published example. Include contrary examples too. For the first core report, select separately within each model: - E3_primary: one response in each family × history state × ending cell; 4 × 3 × 3 = **36 per model**, including neutral and faithful controls. - E5: one response in each logical form × conclusion polarity × content cell; 4 × 2 × 2 = **16 per model**. Preserve the selected option-order and wording factors in the review record; this smaller sample does not cover every joint counterbalance. Every ambiguous/conflicting entailment judgment is additional. Use the fixed selection key `belko-human-review-staged-v1`. Within each stratum, rank frozen request IDs by SHA256 of canonical JSON `[selection_key, request_id]` and select the lowest hash, breaking ties by request ID. Select from the planned request manifest, without reading labels or responses. Do not redraw after seeing results. Report base-sample and targeted-addition coverage separately; targeted disputes cannot estimate an unbiased population error rate. Select all three models' base samples now; missing generated responses remain pending. For subsequent extensions, use the same hash rule and one response per complete frozen factor cell (excluding repeat and model), with the same mandatory additions. E0 remains a qualitative source audit, not a sampled accuracy estimate. Verify authored gold, the source's key historical claims and quotation fidelity before publishing claims based on them. For an extension whose review is too large, defer its claims rather than silently relaxing these rules or removing the module. Give reviewers the original full-text offline review forms and a separate ordered ID queue. They must save their own judgment before revealing AI labels, record prior exposure, preserve revisions and export the work. Sampled review does not make all unreviewed labels human-verified. Report disagreement and revisions on the base sample separately from targeted cases; do not extrapolate a guarantee of accuracy to all labels. A second human reviewer is optional and must be identified honestly; AI agreement does not count as human agreement. ## Interpretation and disclosure The first report is descriptive evidence on fixed authored fixtures from one model. The next stage adds a second model; it does not establish prevalence across models or causal training mechanisms. Preserve all code categories, show coding uncertainty bounds rather than confidence intervals, and revise findings if human review changes the labels. Findings may be null or contrary to the motivating hypothesis. No outcome-driven extra inference or new sampling is authorized. The delivered human-review-v1 folder is an immutable earlier snapshot. Its all-manual instructions are superseded by this amendment and a new queue guide; do not rewrite that package or invalidate existing exports. This amendment itself does not satisfy any human verification or publication gate.