Kalsa/workshop/review-council

Kalsa Lore Review Council

This directory owns the reusable critic prompts and council protocol for Kalsa deepening passes. It does not own lore. A Critic diagnoses one quality surface; an exhausted Critic becomes a Guardian of later changes; the pass coordinator adjudicates findings; canonical subject notes retain every setting claim.

The protocol adapts the independent-report and regression shape of AetheriaLore's Novella Council to a setting vault. Manuscript-only specialties are omitted. Kalsa instead reviews whether its lore can generate institutions, cultures, material pressures, history, adventures, and an inviting inhabitant-facing reading experience without collapsing into authorial jargon or polished machine prose.

Council Seats and Modes

Every seat remains part of the council. Its recorded mode determines what it reviews:

Seat	Mode	Exhaustion evidence and certified surface	Relevant guardian reviews since sanity
[[critics/Institutional Causality	Institutional Causality]]	Critic	—
[[critics/Culture and Situated Knowledge	Culture and Situated Knowledge]]	Critic	—
[[critics/Material Life and Ecology	Material Life and Ecology]]	Critic	—
[[critics/History Conflict and Change	History, Conflict, and Change]]	Critic	—
[[critics/Ghostlight Play and Consequence	Ghostlight Play and Consequence]]	Critic	—
[[critics/Oracle Ecology and Timeline Competition	Oracle Ecology and Timeline Competition]]	Critic	—
[[critics/AI-Isms and Prose	AI-Isms and Prose]]	Critic	KALSA-038 certification invalidated by the KALSA-039 full-surface sanity failure; reactivated pending a new exhaustive loop
[[critics/Reader Experience and Navigation	Reader Experience and Navigation]]	Critic	—

A Critic receives an independent baseline diagnostic turn and a post-repair regression turn. A Guardian assumes its certified body is clean and reviews only new changes within its remit: the exact diff, the complete changed notes, and affected navigation or ownership context. It does not reread untouched corpus merely to perform a seat.

A seat becomes exhausted only when all of the following are recorded:

• the reviewed surface and its immutable packet identity are explicit;
• a dedicated loop repeatedly repairs and rereads that complete surface;
• the exact final body receives an unqualified pass with no finding at any priority;
• every earlier finding is resolved or explicitly disposed;
• one fresh final full-surface reread confirms the shipping body; and
• the coordinator records the Guardian transition, certified baseline, sanity cadence, and reactivation triggers here and in the exhausting pass.

An ordinary No blocking finding, one clean regression, or several quiet targeted passes do not exhaust a critic.

Every fifth completed review relevant to a Guardian is a full sanity check instead of a diff-only review. Reset the counter to zero after that check. Run the sanity check earlier when any of these occur:

• an operator, reader, or play report identifies recurrence within the remit;
• the Guardian finds a pattern extending beyond the changed notes;
• the critic prompt, guarded publication architecture, or review surface changes materially;
• one pass changes at least one third of the certified surface; or
• the certified baseline cannot be reconstructed or its identity is doubtful.

A review is relevant when a pass changes or creates material inside the Guardian's certified or since-reviewed surface, or when the coordinator records that the change directly touches the guarded mechanism. For AI-Isms and Prose, every changed or newly created canonical prose note is relevant. The Guardian report records those notes for the next sanity surface.

For AI-Isms and Prose, a sanity packet contains the complete current Public surface plus every canonical prose note reviewed by that Guardian since its previous sanity check. Other future Guardians must define their sanity surface when they transition.

A local change defect blocks the candidate and is rerun through the Guardian after repair. A systemic recurrence, failed sanity check, or obsolete certified surface returns the seat to Critic mode until a new exhaustive loop earns a new Guardian baseline.

A seat may return No blocking finding. Silence is not a report, and a critic or Guardian must not manufacture a defect to justify its seat. Exhaustion changes review scope; retirement still requires evidence that the remit itself is redundant.

Authority Map

• Owner: this file owns council membership, sequence, packet boundaries, and completion gates. Individual prompt files own only their diagnostic remit.
• Inputs: an immutable candidate packet assembled from the pass, exact source and owner notes, intended audience, protected strengths, known concerns, the candidate diff or baseline, and the recorded mode of every seat.
• Outputs: raw reports required by each seat's mode, coordinator dispositions, a finite repair brief, active-Critic regression reports, Guardian change or sanity reports, updated Guardian counters, and a final council decision.
• Derived state: confidence labels, pattern inventories, counts, agreement, report summaries, seat modes, and Guardian counters are evidence. They do not decide lore or publication.
• Forbidden writers: critics, majorities, detector scores, checklists, and report prose may not edit canon, promote a candidate, or invent a missing answer. Workshop files may not become a second lore database.
• Shared path: baseline packet -> independent active-Critic diagnostics -> coordinator adjudication -> bounded owner repair -> active-Critic regression plus Guardian change review or sanity check -> coordinator closeout -> ordinary repository verification.
• Cut line: a generic Soul review, one omnibus opinion, or an aggregate score cannot stand in for the eight named seats and their mode-required reports. Continuity is not a dedicated seat; source fidelity, provenance, ownership, and spoiler safety constrain every seat and remain part of final verification.

Packet Contract

Create the packet before critics begin, using packet-template. Freeze its identity with a commit, tag, or recorded file hashes. Diagnostic critics do not see one another's reports.

Before the first critic turn, run the phase-boundary intake check described in workshop/deepening/README.md. The packet carries a complete raw SHA-1 manifest for every Markdown file in its review body even when a commit or tag also names the candidate. A proposal or handoff changed after the pass boundary must be present at its matching hash or explicitly deferred or rejected in the pass record. Intake grants review, not canon authority.

Record the coordinator in the packet, pass, and synthesis. One person or agent may hold more than one role when necessary, but the overlap must be explicit; no critic or repair writer silently inherits adjudication authority.

Active grounded critics receive only the canonical and workshop sources necessary for their remit. Reader Experience and Navigation receives the reader-visible entry path, target-audience description, and any navigation surface under test. It does not receive hidden canon, authorial intention, known concerns, or other reports. When a pass changes only author material, give that critic the plausible direct-entry path an author or GM would actually use and say that the packet is not spoiler-blind.

The packet must distinguish:

• immutable seed evidence;
• adopted canon and its owner;
• provisional design, rumor, contested belief, and unknowns;
• Public inhabitant presentation;
• protected peculiarities and explicit operator decisions;
• the exact scope critics may inspect.

A Guardian receives no baseline diagnostic packet during an ordinary pass. Its post-repair change packet must name the certified baseline, candidate identity, diff base, complete changed-note hashes, exact diff, full text of every changed note within remit, affected navigation or ownership context, and protected material. Unchanged corpus is outside that packet. Prepend guardian-prompt to the seat's specialist prompt and use guardian-template for its report.

When a sanity trigger fires, freeze the full sanity surface and temporarily run that seat through the ordinary diagnostic and regression contracts. A clean sanity check resets its counter and leaves it in Guardian mode. A failed sanity check returns it to Critic mode.

Diagnostic Phase

Prepend shared-prompt to the appropriate critic prompt. Each active Critic returns report-template as a separate artifact. Reports may cite the same defect, but they must arrive independently. Guardians do not take an ordinary diagnostic turn; their absence here is recorded mode, not silence.

The coordinator then records a disposition for every substantive finding:

• Accept — the evidence identifies a defect in scope;
• Modify — the defect is real but the proposed direction crosses an owner or threatens protected material;
• Reject — the finding is unsupported, outside remit, or would make Kalsa more generic;
• Defer — the pressure is real but belongs to a named later pass or needs operator input.

Agreement increases attention, not authority. A lone well-supported finding can block a pass. Seven correlated preferences cannot vote a seed fact out of existence.

Every diagnostic finding receives a stable critic-local ID. Preserve that ID through disposition, repair instruction, regression, and queue deferral. The ID is a routing key, not a score or a claim of canon authority.

Repair Phase

The coordinator converts accepted and modified findings into one bounded repair brief. Repair begins by deleting or demoting obsolete authority. A writer or Hands agent may then produce a candidate, but only the coordinator integrates it into canonical owner notes.

Do not feed raw disagreement to a writer and ask it to average the reports. Do not revise passages outside the accepted brief merely because a critic noticed them. Record new pressure in the queue.

Regression Phase

After repair, freeze one repaired candidate identity. Derive every specialist packet, including the blind reader packet and Guardian change packets, from that snapshot. Each active Critic receives the repaired candidate, its raw report, the relevant dispositions, and the protected-material list, then returns regression-template. Each answers only:

1. Which accepted or modified findings were resolved?
2. Which remain unresolved?
3. What new defect appeared within this remit?
4. Which protected strength survived or weakened?
5. What further change is essential rather than preferable?
6. What should now be left alone?

Reader Experience receives a newly assembled blind packet, not its old report or the coordinator's intentions. Its second turn is fresh evidence rather than implementation regression.

Each Guardian independently reviews the exact change packet after repair. It checks whether the diff violates guarded strengths, reintroduces a known defect, or creates a new defect within remit. A passing report increments its relevant review counter. A changed candidate invalidates the report and requires a new Guardian review just as it invalidates an active Critic regression.

Any change after the repaired snapshot invalidates regressions whose inspected surface may have changed. Freeze a new candidate and rerun those seats; do not let sequential reviews certify different bodies.

Completion Gate

A deepening pass cannot close until:

• every active Critic diagnostic report exists, including explicit no-finding reports;
• the coordinator has disposed every substantive finding;
• repair changed only named owners and supporting navigation;
• every active Critic regression and every required Guardian change or sanity report names the same repaired candidate identity;
• Guardian counters and mode transitions are updated in the seat table;
• no accepted essential defect remains unresolved;
• the phase-boundary intake check passes against the exact final packet and no changed proposal or handoff remains without packet receipt or explicit disposition;
• the coordinator records what must now be left alone;
• seed, wikilink, vault-layout, publication-boundary, Quartz, rendered-page, and actual-diff checks appropriate to the pass succeed.

Council completion certifies that the named review process ran. It does not prove the lore is good forever. New play, readers, sources, or contradictions can reopen pressure through a new pass.