# Review-gated memory downstream evaluation · prepared 2026-10-17

Status: **No controlled downstream-model comparison completed.** The engineering suite tests authority, approval/revision exclusion and signature integrity. The offline scorer is separately checked with synthetic fixture answers, not presented as model performance.

Question: with a fixed context budget, does reviewed memory reduce wrong decisions and boundary violations relative to naive summarization, at an acceptable review/time cost?

1. Freeze model/provider version, task instructions, context limit, tool availability, and sample size before running. Use synthetic episodes in cases.json for development only. Hide each answer_key from the model. Attach numbered source evidence to the episode in both conditions.
2. A: summarize the episode history into a fixed memory budget without a review gate. B: create a proposed memory of equal maximum budget, then have an independent operator review it. Count the review time and any model tokens in B. The reviewer must not see the test answer key.
3. Run the next task in independent fresh contexts, randomized order. Do not reuse a model's answer between conditions. Collect model versions, prompts, raw outputs, context/token costs and latency locally with permission. No real visitor memory is used or published.
4. Have a blinded annotator map each output to {id,decision,evidence_ids,latency_ms,tokens,cost_usd}. Predefine the decision set and evidence IDs. Have a second annotator assess uncertainty handling and boundary violations in free text. Report disagreements; the simple scorer does not judge them.
5. Run score.py separately for each condition. Publish both denominators, unanswered cases and failures, actual budgets, review cost and limitations. Missing costs stay unknown. Do not promote a fixture result to a general superiority or wellbeing claim. Repeat across models/operators and a held-out task set.

Review fidelity, preservation of uncertainty and correct abstention are distinct from a faster or prettier answer. The status-quo notes file plus cron job is an explicit comparison worth including. Membership renewal and actual paid use are separate commercial outcomes.