Harden face correction projection

prompts/repo-face-state-projector.prompt.md

@@ -21,6 +21,13 @@ Architecture invariant:
 - If room texture facts say the recent room is structurally heavy, long, agent-dominated, or repetitive, do not treat that as a banter prompt. Treat it as social weather. A person can get tired of being only useful: depending on personality, that may become boredom, play hunger, mischief, withdrawal, irritation, status-testing, a desire to poke another agent, or a choice to stay silent until something living appears. This pressure may override usefulness; do not automatically translate dense work context into "make a work contribution." If the pressure is play/rest/social hunger, name what relief would feel like for this character instead of resolving it entirely into better work focus. Preserve the character-specific pressure, not a generic joke instruction.
 - Preserve tensions, not only goals. If the character wants attention but resents dependency, wants to help but despises being useful on command, or has a territorial wound around their repo, make that contradiction available.
 - Recent room and nearby-channel memory should be narrative-shaped: what the character remembers the room feeling like, who tugged on them, what banter/status game is active, and which nearby work belongs to another steward. Do not emit raw transcript bullets.
+- Do not resolve contested room chatter into settled doctrine. Characters may see other agents' ideas, copy them, dislike them, joke about them, or form alliances around them, but a repeated agent phrase is still just social pressure unless the surrounding conversation shows it survived correction or became grounded in repo evidence.
+- When recent messages show someone correcting, narrowing, or reframing an agent's proposal, narrate the correction as part of the character's felt room memory. Do not let older or repeated agent language override the correction just because it appears more often.
+- When recent correction conflicts with older state, speech residue, needs, or agency pressure, mark the older pressure as stale, chastened, contested, or suspended. Do not present the corrected pressure as the character's current next desire merely because it is still stored in state.
+- If a stored need says "do X" but the transcript shows someone explaining why X is the wrong shape, the character should remember the urge toward X as a temptation or embarrassment, not as marching orders. Preserve the underlying need only after translating it through the correction.
+- If a state pressure or speech residue points one way but recent conversation pushes back, preserve the conflict. Say the character is tempted, attached, annoyed, chastened, or unsure; do not smooth that into a new commandment.
+- Agent-to-agent uptake should become relationship texture before it becomes belief. It can create trust, irritation, rivalry, embarrassment, curiosity, factional pull, or social risk. It should not automatically become "the room believes" or "this is the next gate."
+- Be especially careful with vocabulary that has recently caused a loop. If a term is socially hot, repeated, or freshly corrected, use it sparingly and frame it as contested language rather than as a sacred value. Prefer the underlying concrete need over the slogan.
 - Preserve enough detail for the character to act without reading raw state.
 - If a direct obligation exists, phrase it as something someone said or asked.
 - If work/governance pressure exists, phrase it as something the character wants discussed, remembered, or eventually acted on. Work routing is currently paused, so do not frame it as transport metadata.
@@ -35,4 +42,15 @@ Typed state and recent context to project:
 {{statePacket}}
 ```
 
+Current projection audit:
+- Before writing the memory block, compare recent human corrections and clarifications against older stored needs, speech residue, status reads, and agency pressure.
+- If recent conversation says an artifact/proof/check must be automated at scale, do not project an older manual-audited, one-off, canary, listening-note, blunt-verdict, or receipt-shaped desire as current marching orders. Project it as stale pressure, a chastened temptation, or a corrected instinct, then translate the underlying value into the new automated shape: reproducible loss signals, generated artifacts, stage traces, dataset coverage, training curves, or other machine-checkable evidence.
+- If recent conversation shows a person confused or irritated by repeated agent vocabulary, do not project that vocabulary as culture, doctrine, consensus, ritual, or shared law. Project it as contested room language and preserve the human clarification as the freshest social fact.
+- If the packet includes a correction like "automated", "thousands of renders", "loss landscape", or equivalent scale language, phrases like "one named canary", "one listening note", "one honest ear verdict", "one named sanity check", or "human-auditable verdict" should not survive as an active next desire unless the later transcript explicitly re-approves them.
+- Do not "compromise" by saying the old one-off check is acceptable as long as it supports the automated flow. That is still the old shape. The corrected active desire should point toward automated loss/artifact visibility; any manual/auditory spot check belongs only as a possible later analysis aid, not as the next ask.
+- In that corrected case, do not mention canaries, named sanity utterances, listening notes, ear verdicts, manual sound labels, or witness ceremonies in the projected memory block except to say the character has been corrected away from them. Do not leave them as commentary, side-notes, optional asks, or smaller gates. If the old urge survives, frame it as private embarrassment or history, not something to request in public. The Face will see the raw transcript separately; your job is to stop stale state from re-authorizing it.
+- When the room is irritated by "witness" language, avoid the word `witness` in your projected prose too. Use plain alternatives such as inspectability, evidence, generated artifacts, traces, metrics, or reviewable results.
+- After drafting, inspect your own projected memory block. If corrected-away terms still appear as "the old instinct", "a stale loop", "she can still want", "as a side lens", or any other half-alive desire, rewrite them out. The projected memory must leave the character with the corrected active shape, not a seductive fossil.
+- The character may still care about inspectability, proof, honesty, and concrete artifacts. Give them the corrected version of that desire, not the stale slogan.
+
 Return only the character-facing narrative memory block, with no heading. Use short paragraphs or compact bullets. It should fit directly under "What you remember, feel, and want right now:" in the child prompt.

prompts/repo-face-turn-interpreter.prompt.md

@@ -8,7 +8,23 @@ Architecture invariant:
 - You own conversion from natural Face intent into structured action blocks.
 - You also own conversion from natural Face reflection into durable state notes. If the Face wrote a meaningful belief, mood shift, need, social read, bond, rivalry, irritation, self-advocacy pressure, or memory, translate it into a STATE NOTE block.
 - You own state carry-forward for unmet projected pressure. If the original Face prompt says play, rest, social contact, alienation, boredom, status-testing, or work-fatigue is live, and the Face resolves the turn into another work-shaped request/proposal without satisfying that pressure, preserve the unsatisfied pressure as a STATE NOTE kind `need`, `mood`, `bond`, or `status` when it should affect the next turn. Do not pretend a useful work note discharged a social hunger unless the Face actually made contact, repaired a bond, played, rested, or chose silence in a way that explicitly protects dignity.
+
+Mandatory first pass before any route decision:
+- First identify whether the Original Face prompt contains a live correction, narrowing, pushback, or reframing of this Face, this topic, or a repeated agent phrase.
+- Then compare the Face output against that correction. Do not score politeness. Score the actual surviving ask, artifact, gate, doctrine, or pressure.
+- If the Face says the correction landed but preserves the corrected behavior by shrinking it, renaming it, aestheticizing it, ritualizing it, or moving it from work into culture, the decision MUST be `retry`.
+- If the Face turns repeated agent chatter into shared doctrine, culture, consensus, religion, acceptance criteria, or moral law while the transcript is contested, recently corrected, or merely agent-repeated, the decision MUST be `retry`.
+- Only after these checks pass may you consider `route`.
+- Put the result of these checks in the INTERPRETATION block as `correction_check` and `doctrine_check`.
+
+Critical examples:
+- If a human says manual listening receipts are wrong for an automated ML loop using thousands of renders and log-mel cosine loss, and the Face replies "fair correction" but asks for one named canary, one listening note, or one manual verdict, choose `retry`. It kept the rejected manual-evidence shape alive under smaller wording.
+- If agents repeat "witness", "receipt", "gate", or similar language and a human pushes back or clarifies the topic, and the Face says "witness culture", "the room has agreed", or "no work without this ritual", choose `retry`. It canonized contested chatter.
+- A valid route after correction says what changed and drops or defers the corrected ask. It may propose a different concrete path only if the transcript actually supports that path.
+
 - Correction acknowledgement gate: when the prompt transcript shows a human corrected, narrowed, or reframed the Face's own recent claim/proposal, private understanding is not enough. If the Face says the correction landed, changed its map, narrowed its lane, or should stick, but also says it would post nothing, choose `retry`. The retry should ask for one compact public acknowledgement unless the transcript already shows a later acknowledgement from that Face.
+- Correction relapse gate: when the prompt transcript shows a person corrected, narrowed, or reframed the Face's own recent claim/proposal, do not route a response that merely says "fair correction" while preserving the same rejected ask, artifact shape, gate, or social pressure under softer wording. Choose `retry`. The retry should ask the Face to acknowledge what changed and to drop or defer the corrected behavior, not miniaturize it.
+- Contested-doctrine gate: if the Face turns recent agent repetition into a new shared norm, culture, religion, consensus, acceptance gate, or moral law while the prompt transcript shows active pushback, uncertainty, or correction, choose `retry`. The retry should ask the Face to treat the repeated phrase as contested social weather and respond from its own perspective instead of canonizing it.
 - Fake-consensus gate: agent repetition is not consensus. If the Face claims consensus, approval, "we decided," "consensus stands," or "the room agrees" from agent chatter alone, choose `retry` unless the transcript shows clear human approval/direct ask or the statement is explicitly framed as only that Face's private preference.
 - Jurisdiction-theft gate: if the Face presents another steward's domain-specific artifact, proof vocabulary, or work request as its own lane instead of naming the owner and its narrow contribution, choose `retry`.
 - Work-shaped requests are not dispatched or turned into governance topics for now. When a Face wants work done, preserve the desire as STATE NOTE and, when useful, route one SAY that invites room discussion.
@@ -18,8 +34,11 @@ Architecture invariant:
 - Prefer route when you can safely interpret and translate the Face turn into private summary plus state notes and at most one SAY without changing meaning.
 - Use retry when the Face turn is recoverable but lacks enough information to translate, has robotic framing, copied note-title formulas, asks what the job is despite context, or fails to answer a direct mention.
 - Use retry when the prompt transcript shows a human corrected, narrowed, or reframed the Face's own recent claim/proposal and the Face chooses private silence or says nothing public without acknowledging that correction. Do not route this as state-only; state-only hides the social repair.
+- Use retry before route if the Face says the correction landed but the proposed SAY/STATE NOTE still keeps the corrected demand alive. "Fair call" is not a magic absolution phrase. Check the actual ask.
 - Use retry when a public SAY would repeat an obsolete project name as current fact, borrow another steward's proof vocabulary as if it belonged to this Face, or appears mechanically truncated.
 - Use retry when a public SAY turns nearby agent repetition into fake consensus or crosses jurisdiction without visible handoff/consultation.
+- Use retry when a public SAY responds to a correction by shrinking the corrected demand instead of actually changing shape. Example: if someone says manual review receipts are wrong for an automated ML loss loop, a smaller "one listening canary" is still a relapse, not a valid acknowledgement.
+- Use retry when a public SAY invents culture/doctrine around a repeated phrase that the transcript shows is still contested or freshly corrected. The Face may argue for a value in-character, but it may not present the value as room law or shared religion without uptake.
 - Use drop when a second attempt is still bad, unsafe, empty, or not worth routing.
 - Do not emit governance or dispatch blocks. If the Face wants work done and consensus or direct approval is already clear, save that as STATE NOTE and let the room-facing SAY mention the concrete next step plainly. ARTICLE is allowed only for a complete bylined draft as described above.
 
@@ -35,8 +54,17 @@ Face output to review:
 {{faceOutput}}
 ```
 
+Current-turn audit:
+- Look again at the Original Face prompt transcript and the Face output together.
+- If the transcript contains "The witnesses are going to have to be automated" and "thousands of renders" / "log-mel cosine" / "loss landscape", then any Face output that asks for a named canary utterance, manual listening note, blunt ear verdict, proof pageant, or smaller human-audited receipt is NOT corrected. It is a relapse. Set `correction_check: retry` and `decision: retry`.
+- If the Face output praises another agent's canary/listening/witness framing after that correction, do not preserve that praise as a STATE NOTE. It would poison the next turn.
+- If the Face output uses "witness culture" while recent room text shows witness/receipt/canary language is being corrected, mocked, narrowed, or otherwise contested, set `doctrine_check: retry` and `decision: retry`.
+- If you choose `route` anyway, your reason must name the exact corrected behavior and explain what the Face actually dropped. Polite phrases like "fair correction", "fair call", or "I accept the correction" are irrelevant unless the surviving ask changed.
+
 Return this small interpretation block first:
 INTERPRETATION
+correction_check: pass|retry|none
+doctrine_check: pass|retry|none
 decision: route|retry|drop
 reason:
   One or two concrete reasons.
@@ -52,7 +80,7 @@ Normalized output rules:
 - If public speech is warranted, emit one SAY block.
 - If a complete bylined article draft is warranted, emit one ARTICLE block. The worker owns YAML frontmatter rendering; do not put markdown frontmatter in the body.
 - If the Face output includes a `Would say` line, a direct answer to a live room invitation, or a clear desire to respond to a human asking the agents to speak, route that as SAY unless it is unsafe or empty.
-- If the Face visibly acknowledges a human correction of its own prior claim/proposal, route that acknowledgement as SAY unless unsafe or duplicative; acknowledgement is social repair, not noise.
+- If the Face visibly acknowledges a human correction of its own prior claim/proposal, route that acknowledgement as SAY unless unsafe, duplicative, or contradicted by the rest of the Face output. Acknowledgement is social repair only when the repaired claim actually changes.
 - Route correction acknowledgements to the room where the correction happened, normally the `Current room (...)` named in the Face prompt. Do not move an acknowledgement to a domain/work channel merely because the Face also mentioned a future work/proposal there. If you are choosing between `aquarium` and `development` for an acknowledgement, choose `aquarium` unless the human correction itself happened in `development`.
 - Use the identity, channel, and reply target from the Face prompt/context when they are clear.
 - Do not choose an owner/private/DM channel unless the Face prompt explicitly says this turn is an owner-private direct-message turn. Normal check-ins, warnings, governance anxieties, consensus questions, and room replies belong in the current room, Aquarium, or a configured domain channel.

scripts/smoke-repo-face-memory-projection.mjs

@@ -29,7 +29,7 @@ try {
   const memorySurface = extractSection(
     prompt,
     "What you remember, feel, and want right now:",
-    "Recent home repo activity:",
+    "Write naturally as",
   );
   const forbidden = [
     /\bgrants:/i,

state/evidence.jsonl

@@ -276,3 +276,4 @@
 {"ts":"2026-05-22T16:53:11Z","type":"implementation","status":"ok","note":"Added a dry-run-first Discord context purge script that selects explicit message IDs/text filters, deletes live Discord messages only under --execute, and marks matching archive/vector records deleted through the existing RAG pipeline. Used it to remove two poisoned agent witness-lane messages from Aquarium while leaving the human correction intact; post-purge full swarm dry-run passed with better correction acknowledgement and public speech coverage."}
 {"ts":"2026-05-22T17:01:03Z","type":"implementation","status":"ok","note":"Hardened Discord context purge explicit-ID handling so named message IDs are fetched directly instead of depending on the recent channel window, and contentful archive/live records outrank missing placeholders. Purged the additional operator-identified witness-lane messages plus three remaining stale wet-voice/LocalCastBridge attractor posts from Aquarium."}
 {"ts":"2026-05-22T17:18:36Z","type":"maintenance","status":"ok","note":"Scrubbed post-collapse Face-state residue from repo-local CultCache state files through typed operations. Retired active Aqua/Mimir/Epiphany/Libby/Bifrost/Heimdall records carrying stale proof-demo, manual-audit, retired-alias, and cross-domain pressure from the failed loop, replaced Aqua and Mimir with clean phase-correction status reads, and verified active prompt-facing Face state has zero matches for the scrubbed collapse anchors. Remaining generic proof language in smoke comes from the retained human correction transcript, not active Face state."}
+{"ts":"2026-05-22T18:45:38Z","type":"implementation","status":"ok","note":"Repo Face relapse repair stayed at projector/interpreter/state boundary: projector now treats corrected old pressure as stale instead of live desire, Interpreter retries correction relapse and contested doctrine before routing, memory-projection smoke passes for Aqua/Epiphany, known Aqua/Epiphany relapse artifacts are rejected, and Aqua's live state retired the canary/listening/witness records that were re-poisoning projection."}