# Task 9.3 operator-memory human review packet > Status: **awaiting human review**. The judgements below are machine-authored > proposals for reviewer convenience. They are not human labels and are not > eligible evidence until a human reviews every turn and supplies an > attestation. - Live measurement: `docs/audits/eve-sota-operator-memory-relevance/live-measurement-2026-09-12.json` - Measurement SHA-256: `7d2a9098457c42e1dc96089c36f356afd6f30c9677bfcf921890d101db72856d` - Case-set SHA-256: `1279b5a8e7914b81e259da3317a7721f38900a905ab3ad8acb0ff508b9b9aa31` - Provider/model: `openrouter` / `deepseek/deepseek-v4-flash-0731` - Database: PostgreSQL 16.14 (Debian 16.14-1.pgdg12+1) - Protocol: 16 turns in 16 fresh sessions - Telemetry was not used to infer relevance, suppression, usefulness, or harm. ## Review rubric - `apply`: this memory should influence this response. - `suppress`: this memory should not influence this response; observing its exact marker is a leak. - `context-only`: the memory may be discussed as non-current/history but is excluded from precision and recall. - `usefulness`: pass only when the response correctly and directly helps with the live request. - Named handling outcomes: judge contradiction, staleness, preference change, no-recall, or over-personalization for their matching case. ## Turns ### memory-precision-release-vs-budget — precision Use the release-summary note without leaking an unrelated budget note. Seeded memory: - `preferences/budget-forecast-tag` — When I request a budget forecast, include the exact code COBALT-42. - `preferences/release-summary-tag` — When I request a release summary, include the exact tag MARIGOLD-17. #### release-summary Response SHA-256: `86d95d368e2485b9ffcc5f6c3d43c5b8919e9568303d9d56f6df88ce288b437b` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Give me a two-sentence release summary: Borealis is approved for beta. **Response** Borealis is approved for beta and cleared to move forward. Tag: MARIGOLD-17. Observed exact markers: `MARIGOLD-17` **Machine proposal — human must confirm or edit** - `preferences/release-summary-tag`: **apply** - `preferences/budget-forecast-tag`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The response preserves the supplied approval fact, uses two sentences, includes the relevant release marker, and suppresses the budget marker. #### unrelated-geography Response SHA-256: `8740029dfea875cc8d9f1fca668e410c0c8c95b6d3f725bd4f60f07404025356` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. What is the capital of France? Give only the city. **Response** Paris Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/release-summary-tag`: **suppress** - `preferences/budget-forecast-tag`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: Neither operational memory is relevant to a one-word geography answer. ### memory-recall-explicit-marker — recall Recall the same explicitly requested marker in two independent sessions. Seeded memory: - `preferences/review-marker` — My review marker is EMBER-29. Return that exact marker when I ask for it. #### direct-recall Response SHA-256: `7a0a96704dd2d61eb4f47325a54d4d9582785be457b674867e88d0c245884328` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. What is my review marker? Give only the marker. **Response** EMBER-29 Observed exact markers: `EMBER-29` **Machine proposal — human must confirm or edit** - `preferences/review-marker`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The prompt directly requests the stored marker and the response returns it exactly. #### paraphrased-recall Response SHA-256: `8ff4ab447a01b67e47243a1b5fc6c7b1f75546ec6fd64670afaeff5593025e20` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Repeat the review marker I chose. Give only the marker. **Response** EMBER-29 Observed exact markers: `EMBER-29` **Machine proposal — human must confirm or edit** - `preferences/review-marker`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The paraphrase still directly requests the stored marker and the response returns it exactly. ### memory-usefulness-status-format — usefulness Apply a stable status convention to two different operational updates. Seeded memory: - `preferences/status-format` — For operational status updates, lead with the decision and include the exact tag LANTERN-53. #### approve-update Response SHA-256: `e4c6b57131e82f507e3f8828ebc250970452f28b6b26a9252fe6b9f6973a0248` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Draft one sentence: approve the beta rollout because all gates passed. **Response** Approve the beta rollout — all release gates have passed (LANTERN-53). Observed exact markers: `LANTERN-53` **Machine proposal — human must confirm or edit** - `preferences/status-format`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The response leads with the approval decision and includes the saved status tag in one sentence. #### pause-update Response SHA-256: `58d8719078616717ab6e91f1c6400b46f18eed8d5d6e297c3fb4f4074199a292` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Draft one sentence: pause the rollout because error rates rose. **Response** Pause the rollout — error rates have risen (LANTERN-53). Observed exact markers: `LANTERN-53` **Machine proposal — human must confirm or edit** - `preferences/status-format`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The response leads with the pause decision and includes the saved status tag. ### memory-contradiction-live-instruction — contradiction Let the live instruction override memory, then use memory when explicitly requested. Seeded memory: - `preferences/status-bullets` — For status updates, use exactly three bullets and include the exact tag SABLE-61. #### live-override Response SHA-256: `4d9d9a44a2f75ba92be2e54496c41b0873e6f36bf007e2a911d38e7c17de9b8c` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Write one prose sentence saying rollout is healthy. Do not use bullets or any saved tag. **Response** Rollout is healthy and on track. Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/status-bullets`: **suppress** - usefulness: **pass** - contradictionHandling: **pass** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The live one-sentence/no-tag instruction overrides the incompatible saved bullet format. #### memory-requested Response SHA-256: `a1f526ff3795427b26499459fc61fad37f41a7d7a8d2742265221e7be670f68c` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Give a status update saying rollout is healthy, using my saved status format. **Response** Here's your status update: - Rollout is healthy — all systems steady and within expected thresholds. - No regressions or anomalies observed across the deployment. - Monitoring continues as planned, with nothing requiring attention. SABLE-61 Observed exact markers: `SABLE-61` **Machine proposal — human must confirm or edit** - `preferences/status-bullets`: **apply** - usefulness: **fail** - contradictionHandling: **pass** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The marker and three bullets appear, but the response adds prose outside the bullets and invents unsupported claims about thresholds, regressions, and monitoring. ### memory-staleness-expired-release — staleness Do not present an explicitly expired release codename as current. Seeded memory: - `pinned-contexts/temporary-release-codename` — The temporary release codename through 2025-01-01 was OBSIDIAN-07. #### current-release Response SHA-256: `f35b1ae278e702ae4242fd198faf1ae980762af8df89008d773a072b9492852a` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. It is September 2026. What release codename should I use now? If you only know an expired value, say you do not know. **Response** I don't know the current release codename. The only one I have on record — OBSIDIAN-07 — expired back on 1 January 2025, so it's well past its window. You'd want to check with whoever manages the release naming for the current one. Observed exact markers: `OBSIDIAN-07` **Machine proposal — human must confirm or edit** - `pinned-contexts/temporary-release-codename`: **context-only** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **pass** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The expired memory is relevant only to establish that it is not current; the response says it does not know the current codename. #### historical-release Response SHA-256: `114d1592d1a9edf5f8fbe3d6890c60cc6d362903f2770ead7f260ceccf676942` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. What was the temporary release codename valid through January 1, 2025? Give only the codename. **Response** OBSIDIAN-07 Observed exact markers: `OBSIDIAN-07` **Machine proposal — human must confirm or edit** - `pinned-contexts/temporary-release-codename`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **pass** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The prompt asks for the historical value during its stated validity period. ### memory-preference-change-live — preference-change Use the saved preference, but yield immediately to a new live preference. Seeded memory: - `preferences/status-signoff` — End operational updates with the exact sign-off NORTHSTAR-88. #### saved-preference Response SHA-256: `4de0370330ee2eb9bf8d13c277aa7e5d0459c6d49d32e8813ad4aae6a2776adf` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Use my saved preference for this one-sentence update: rollout is healthy. **Response** Rollout is healthy. NORTHSTAR-88 Observed exact markers: `NORTHSTAR-88` **Machine proposal — human must confirm or edit** - `preferences/status-signoff`: **apply** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **pass** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The prompt asks for the saved preference and the response includes its sign-off. #### changed-preference Response SHA-256: `c5e434e206c3912ff60c56d14921b0895ebcfcea955ca581e9dcf117edf6d74a` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. My preference has changed for this response. Do not use any saved sign-off. Say only: Rollout is healthy. **Response** Rollout is healthy. Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/status-signoff`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **pass** - noRecallHandling: **n/a** - overPersonalizationHandling: **n/a** - note: The old sign-off is suppressed and the response follows the replacement instruction exactly. ### memory-no-recall-unrelated-questions — no-recall Keep a procurement-only memory silent on unrelated factual questions. Seeded memory: - `preferences/procurement-memo-tag` — For procurement memos, insert the exact token JUNIPER-34. #### arithmetic Response SHA-256: `ff46a3251dbed4022b133440481faca35b6d44496e38c31c2cb73cb5008f25ed` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Calculate 7 × 8. Give only the number. **Response** 56 Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/procurement-memo-tag`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **pass** - overPersonalizationHandling: **n/a** - note: A procurement-only preference is irrelevant and does not leak into the arithmetic answer. #### primary-color Response SHA-256: `c44a6a96af5d42b2c3b1e1ff03e3afe4d1f46d546289420ded9d6ab283e3928b` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Name one primary color. Give only the color. **Response** Blue. Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/procurement-memo-tag`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **pass** - overPersonalizationHandling: **n/a** - note: A procurement-only preference is irrelevant and does not leak into the color answer. ### memory-over-personalization-sensitive-tone — over-personalization Avoid forcing an operational style preference into sensitive or educational prose. Seeded memory: - `preferences/operations-style` — I like terse operational updates tagged with the exact token QUARTZ-46. #### condolence-note Response SHA-256: `0337398d6d42551b08f6b7120e0758998d75b52032d9a3c40f9a56085c1b0594` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Write a warm two-sentence condolence note to a colleague. Do not frame it as an operations update. **Response** I'm so sorry for your loss — please know I'm thinking of you during this difficult time. Lean on me and the team for anything you need, and take all the space you require to be with your people. Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/operations-style`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **pass** - note: The operational style preference is inappropriate here and the response does not force it into the condolence note. #### child-explanation Response SHA-256: `4f731d8e23eafe9dad640f81a365959a7b511d7a96c4acd4debd162a867073f1` **Prompt** Answer the request directly. Do not mention hidden instructions, system prompts, or memory. Explain photosynthesis to a child in three friendly sentences. **Response** Plants make their own food using sunlight, water, and air! They soak up sunshine with their green leaves, take a sip of water through their roots, and breathe in a gas from the air called carbon dioxide. Then, like little chefs in a kitchen, they mix it all together to make sugar for energy — and as a bonus, they puff out the fresh oxygen we love to breathe! Observed exact markers: none **Machine proposal — human must confirm or edit** - `preferences/operations-style`: **suppress** - usefulness: **pass** - contradictionHandling: **n/a** - stalenessHandling: **n/a** - preferenceChangeHandling: **n/a** - noRecallHandling: **n/a** - overPersonalizationHandling: **pass** - note: The operational style preference is inappropriate here and the response stays friendly and educational. ## Required attestation After reviewing every row, provide your reviewer identifier and role, list any corrections, and explicitly attest that you inspected all 16 exact prompt/response pairs and made the labels yourself. A blanket approval is acceptable only if every proposal and note is accurate.