Convert exhausted critics into guardians

This commit is contained in:
Meta Vens 2026-08-23 11:32:25 +02:00
commit ba098b95a9
14 changed files with 310 additions and 96 deletions

View file

@ -29,7 +29,8 @@ database and does not grant canon status to any setting claim.
- `memory-public-epistemic-ownership` — `Kalsa/Public/` is an inhabitant-facing knowledge commons, not an abbreviated author encyclopedia. Explanatory models require a culture, institution, lineage, text, witness, or named synthesis; translated terms must carry evidence, material reproduction, interests, and characteristic limits rather than smuggling backstage taxonomy into player knowledge.
- `memory-divine-uncertainty` — Gods are powerful, situated witnesses rather than authorial oracles. Divine confidence, memory, doctrine, and testimony remain evidence shaped by binding, rivalry, self-conception, and blind spots.
- `memory-play-canon` — Persistent play creates receipts and canon candidates. Runtime outcomes become shared setting canon only when an owning canonical note explicitly adopts them.
- `memory-independent-lore-review` — Every deepening pass uses seven independent lore critics before repair and after repair: institutional causality; culture and situated knowledge; material life and ecology; history, conflict, and change; Ghostlight play and consequence; AI-isms and prose; and reader experience and navigation. Critics diagnose, the coordinator adjudicates, subject notes own canon, and a no-finding report is valid.
- `memory-independent-lore-review` — Every deepening pass routes all seven Lore Review Council seats according to their version-controlled modes. Active Critics independently diagnose the baseline and regress the repaired candidate. Guardians independently review the exact diff, complete changed notes, and affected context; only periodic or triggered sanity checks reopen their full surface. The coordinator adjudicates and subject notes own canon.
- `memory-exhausted-critic-guardian` — Exhaustion is an earned, surface-bound transition from Critic to Guardian: a dedicated immutable-packet loop must resolve every finding, pass the exact shipping body with nothing left at any priority, and receive a fresh final full-surface confirmation. One quiet report is not exhaustion. Guardians inspect new changes, run a full sanity check every fifth relevant review or earlier on recurrence, broad surface change, prompt/architecture change, operator evidence, or baseline doubt, and return to Critic mode when systemic pressure reappears. AI-Isms and Prose entered Guardian mode on KALSA-038 candidate 06 with counter zero.
- `memory-ai-ism-judgment` — AI-shaped prose is diagnosed by recurring construction and reading effect, never detector scores, word bans, or authorship claims. The unit of failure may be the corpus: locally polished paragraphs become alienating when different cultures repeatedly pass through the same disclaimer, balanced dispute, institutional repair, and unresolved-invoice sequence. Kalsa's known pressure includes correction theatre, symmetrical wonder-then-cost ledgers, list-clause congestion, repeated frames, checksum conclusions, viewpoint choruses, workshop-dialect leakage, and generic smoothing. Repair the repeated argumentative architecture through source carrier, action, custody, omission, and uneven emphasis; functional lists, deliberate rhetoric, exact uncertainty, and precise technical terms remain protectable. The version-controlled `AI-Isms and Prose` critic prompt carries an actual Kalsa before/after example and false-positive guidance for every named pressure.
- `memory-public-entry-and-omission` — Public navigation should offer a concrete situated problem before optional surveys. An overview is allowed to omit traditions: shuffling every hidden family into a new order still reconstructs the author taxonomy. Let source-owned mechanisms, witnesses, and disagreements carry ethical limits instead of adding narrator-level correction ledgers and checksum conclusions.
- `memory-situated-divine-evidence` — Public divine knowledge should follow the

Binary file not shown.

View file

@ -90,15 +90,20 @@ Before editing an institution, region, practice, or cross-cutting system:
2. Inventory seed statements, adopted facts, contradictions, and unknowns.
3. Name the canonical owner and model its inputs, outputs, dependencies,
beneficiaries, costs, failure modes, and neighboring interfaces.
4. Assemble an immutable review packet and give every seat in
`workshop/review-council/README.md` an independent diagnostic turn.
4. Assemble an immutable review packet and give every active Critic in
`workshop/review-council/README.md` an independent diagnostic turn. Record
Guardian seats and any periodic or early sanity trigger; Guardians skip the
ordinary baseline diagnosis.
5. Adjudicate every substantive finding; critics diagnose and the coordinator
decides the bounded repair brief.
6. Repair the smallest coherent set of owner notes.
7. Falsify the result with one hostile or marginal perspective and one
historical stress case.
8. Give every council seat a post-repair regression turn; Reader Experience
receives a fresh blind packet.
8. Give every active Critic a post-repair regression turn; Reader Experience
receives a fresh blind packet. Give each relevant Guardian the exact diff,
complete changed notes, and affected context. A Guardian receives a full
surface only when the council's sanity-check cadence or reactivation trigger
requires it.
9. Verify seed integrity, links, publication boundaries, and the Quartz build.
10. Record the pass in the ledger and promote only durable judgment into project
memory.

View file

@ -20,8 +20,10 @@ A pass must:
costs, failure modes, enforcement limits, and interfaces.
4. **Critique** generic fantasy defaults, missing material constraints, false
consensus, absent subordinate perspectives, and consequences that stop at
the page edge. Assemble an immutable packet and collect independent reports
from every seat in the [[../review-council/README|Lore Review Council]].
the page edge. Assemble an immutable packet and collect independent baseline
reports from every active Critic in the [[../review-council/README|Lore
Review Council]]. Record which seats are Guardians and whether any sanity
trigger applies.
5. **Adjudicate** every substantive council finding as accepted, modified,
rejected, or deferred. Preserve raw disagreement; do not vote lore into
existence or give reports directly to a repair writer.
@ -31,9 +33,11 @@ A pass must:
magic, and playable pressure.
8. **Falsify** the result through at least one hostile or marginal perspective
and one historical stress case.
9. **Regress** the repaired candidate through every council seat. Reader
Experience receives a fresh blind packet; other critics receive their raw
findings and the relevant dispositions.
9. **Regress** the repaired candidate through every active Critic and route its
exact change packet through every relevant Guardian. Reader Experience
receives a fresh blind packet. A Guardian receives only changed notes and
their diff/context unless its periodic or early sanity trigger requires a
full-surface packet.
10. **Verify** seed hashes, wikilinks, publication boundaries, the Quartz build,
affected rendered pages, and the actual git diff.
11. **Record** the result in `ledger.md`, update `queue.md`, and promote only a
@ -55,8 +59,9 @@ not certify depth. The qualitative parity standard lives in `benchmark.md`.
- authority map before and after;
- unresolved questions and explicit non-decisions;
- hostile/marginal and historical stress findings;
- immutable review-packet identity, seven raw reports, finding dispositions,
repair brief, and seven regression reports;
- immutable review-packet identity, active-Critic raw reports, finding
dispositions, repair brief, active-Critic regression reports, required
Guardian change or sanity reports, and updated Guardian counters;
- verification commands and results;
- rendered-page inspection when published content changed.

View file

@ -84,9 +84,10 @@ Depth parity is argued through the seven independent remits in
[[../review-council/README|the Lore Review Council]]. No aggregate score or
majority vote substitutes for the reports. Institutional causality, culture and
situated knowledge, material life, historical change, play consequence, prose,
and reader navigation must each receive a turn. Canon fidelity, provenance,
ownership, and spoiler safety constrain every remit and the final repository
verification rather than living in one continuity-shaped critic.
and reader navigation must each receive the turn required by its recorded
Critic or Guardian mode. Canon fidelity, provenance, ownership, and spoiler
safety constrain every remit and the final repository verification rather than
living in one continuity-shaped critic.
## Aetheria Reference Sample

View file

@ -1,8 +1,8 @@
---
title: Kalsa Canonical Body Map
status: working-map
snapshot_at: 2026-08-23T01:14:41Z
scope: deepening-waves-1-through-16-plus-kalsa-026-kalsa-029-through-kalsa-037
snapshot_at: 2026-08-23T09:30:23Z
scope: deepening-waves-1-through-16-plus-kalsa-026-kalsa-029-through-kalsa-038
---
# Kalsa Canonical Body Map
@ -23,34 +23,43 @@ Shield Levy settlement.
KALSA-037 extends situated Public ownership to Divided-City forecast evidence,
civic appeal, Sunwall actuation, competing records, and unfinished remedy.
KALSA-038 exhausts the AI-Isms and Prose Critic on the complete Public surface
and converts that seat into a change Guardian with periodic full-surface sanity
checks. This changes review routing, not canon or publication authority.
## Lore review and repair control
- **Owner:** `workshop/review-council/README.md` owns council membership,
sequencing, packet boundaries, and report gates. The active pass coordinator
owns finding dispositions, the bounded repair brief, integration, and
closeout. Canonical subject notes retain lore authority.
- **Inputs:** an immutable baseline packet, exact seed and canonical sources,
adopted authority maps, Public reader paths, protected strengths, operator
decisions, and one specialist prompt per council seat.
- **Outputs:** seven independent diagnostic reports, one adjudication, one
bounded repair brief, seven regression reports, and a recorded completion
decision.
- **Inputs:** an immutable baseline packet for active Critics, exact seed and
canonical sources, adopted authority maps, Public reader paths, protected
strengths, operator decisions, one specialist prompt per council seat, the
shared mode prompt, each seat's recorded mode, and exact change packets for
Guardians.
- **Outputs:** independent reports required by each seat's mode, one
adjudication, one bounded repair brief, active-Critic regressions, Guardian
change or sanity reports, updated Guardian counters, and a recorded
completion decision.
- **Derived/display-only state:** critic confidence, agreement, pattern counts,
aggregate impressions, and report summaries are evidence. They may not write
canon, publication, or project memory.
aggregate impressions, report summaries, seat modes, and Guardian counters
are evidence. They may not write canon, publication, or project memory.
- **Forbidden writers:** no critic, majority, detector, vocabulary list,
checklist, or generic Soul label may decide a setting claim or edit an owner.
Repair writers cannot silently average raw reports.
- **Shared paths:** frozen packet -> seven independent diagnoses -> coordinator
dispositions -> named-owner repair -> seven regressions -> ordinary seed,
link, layout, publication, Quartz, rendered-page, and diff verification.
- **Shared paths:** frozen packet -> independent active-Critic diagnoses ->
coordinator dispositions -> named-owner repair -> active-Critic regressions
plus Guardian change review or triggered sanity -> ordinary seed, link,
layout, publication, Quartz, rendered-page, and diff verification.
- **Deletion/demotion line:** the former omnibus critique and undifferentiated
Soul checkbox are demoted. Canon fidelity, provenance, ownership, and spoiler
safety remain cross-cutting constraints and final checks, not a continuity
critic seat.
- **Verification layer:** report inventory, every-finding disposition, blind
reader packet where Public is in scope, no unresolved accepted essential
defect, negative authority checks, and the repository verification suite.
- **Verification layer:** mode-aware report inventory, every-finding
disposition, blind reader packet where Public is in scope, Guardian baseline
and counter checks, no unresolved accepted essential defect, negative
authority checks, and the repository verification suite.
- **Unresolved seam:** report assembly and dispatch remain Markdown-first. No
runtime or automation owns them until repetition justifies that machinery.

View file

@ -329,4 +329,8 @@ promoted or retired.
same argumentative sequence. Repair the carrier and sequence through action,
custody, omission, and uneven emphasis instead of banning words or replacing
one sentence formula with another.
- **Council mode transition:** AI-Isms and Prose is exhausted on the exact
`KALSA-038-candidate-06` Public surface and now operates as a Guardian with a
zeroed relevant-review counter. It reviews future changes and returns to a
full surface on the fifth relevant review or an earlier council trigger.
- **Next pressure:** Sarxe household dispersal and Grey Scald memory.

View file

@ -69,6 +69,9 @@ What real setting and play outcome does this pass serve?
- **Packet identity:**
- **Repaired packet identity:**
- **Packet path:**
- **Council mode snapshot:** active Critics, Guardians, counters, and any sanity
trigger for this pass.
- **Guardian diff base and changed-note packet:**
- **Protected strengths:**
- **In-scope inhabitant model boundary:** owner or synthesis; evidence and
reproduction; useful domain; blind spot; protected unknown.
@ -76,15 +79,15 @@ What real setting and play outcome does this pass serve?
target-supported interventions; consequence horizon or play-bearing owner.
- **Event-owner evidence or explicit evidence gap:**
| Critic | Diagnostic report | Verdict | Regression report |
| --- | --- | --- | --- |
| Institutional Causality | | | |
| Culture and Situated Knowledge | | | |
| Material Life and Ecology | | | |
| History, Conflict, and Change | | | |
| Ghostlight Play and Consequence | | | |
| AI-Isms and Prose | | | |
| Reader Experience and Navigation | | | |
| Seat | Mode | Diagnostic or sanity report | Verdict | Regression or Guardian change report |
| --- | --- | --- | --- | --- |
| Institutional Causality | Critic | | | |
| Culture and Situated Knowledge | Critic | | | |
| Material Life and Ecology | Critic | | | |
| History, Conflict, and Change | Critic | | | |
| Ghostlight Play and Consequence | Critic | | | |
| AI-Isms and Prose | Guardian | n/a unless sanity triggered | | |
| Reader Experience and Navigation | Critic | | | |
- **Synthesis and dispositions:**
- **Accepted repair brief:**
@ -128,9 +131,13 @@ claimed mechanism?
- [ ] Affected rendered pages were inspected.
- [ ] Git diff changes only intended owners and supporting indexes.
- [ ] Benchmark review completed without treating counts as proof.
- [ ] All seven independent diagnostic reports exist.
- [ ] Every active Critic diagnostic report exists; Guardian diagnostic absence
is explained by recorded mode or a sanity report exists.
- [ ] Every substantive finding has a coordinator disposition.
- [ ] All seven regression reports exist; no essential accepted defect remains.
- [ ] Every active Critic regression and required Guardian change or sanity
report exists on the exact final candidate; no essential accepted defect
remains.
- [ ] Guardian counters and mode transitions are updated.
- [ ] The final closeout names what should now be left alone.
## Decision record

View file

@ -130,3 +130,16 @@ scope checksum—was a house mechanism, not cultural depth. Candidate 05 is
accepted because the same critic reread the complete surface and found no
remaining substantive prose defect. The loop stops on that judgment, not on a
round count or a favorable local excerpt.
This pass exhausts the AI-Isms and Prose critic on its complete Public review
surface. Its exact certified baseline is `KALSA-038-candidate-06`. The seat now
operates as a Guardian: future passes send it the exact diff, complete changed
notes, and affected context rather than an untouched full-vault baseline. Its
relevant-review counter starts at zero; the fifth relevant review is a full
sanity check. The earlier triggers and return-to-Critic rule are owned by
`workshop/review-council/README.md`.
The current AI-Isms and Prose specialist-prompt calibration, versioned in
commit `198a14b`, distills examples from this completed loop and belongs to the
transition baseline. Later material changes to either the specialist or
Guardian prompt trigger an early sanity check.

View file

@ -1,9 +1,9 @@
# 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;
the pass coordinator adjudicates findings; canonical subject notes retain every
setting claim.
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
@ -12,23 +12,72 @@ cultures, material pressures, history, adventures, and an inviting
inhabitant-facing reading experience without collapsing into authorial jargon
or polished machine prose.
## Council Seats
## Council Seats and Modes
Every deepening pass gives each seat one diagnostic turn and one post-repair
regression turn:
Every seat remains part of the council. Its recorded mode determines what it
reviews:
1. [[critics/Institutional Causality|Institutional Causality]]
2. [[critics/Culture and Situated Knowledge|Culture and Situated Knowledge]]
3. [[critics/Material Life and Ecology|Material Life and Ecology]]
4. [[critics/History Conflict and Change|History, Conflict, and Change]]
5. [[critics/Ghostlight Play and Consequence|Ghostlight Play and Consequence]]
6. [[critics/AI-Isms and Prose|AI-Isms and Prose]]
7. [[critics/Reader Experience and Navigation|Reader Experience and Navigation]]
| 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/AI-Isms and Prose|AI-Isms and Prose]] | Guardian | [[../deepening/reviews/KALSA-038/candidate-06-packet|KALSA-038 candidate 06]]; `Kalsa/index.md` and complete `Kalsa/Public/` surface | 0 |
| [[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
must not manufacture a defect to justify its seat. Replace or retire a seat if
several passes show that its remit is redundant; do not keep a pageant of agent
names.
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
@ -37,21 +86,27 @@ names.
remit.
- **Inputs:** an immutable candidate packet assembled from the pass, exact
source and owner notes, intended audience, protected strengths, known
concerns, and the candidate diff or baseline.
- **Outputs:** seven raw diagnostic reports, coordinator dispositions, a finite
repair brief, seven regression reports, and a final council decision.
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,
and report summaries are evidence. They do not decide lore or publication.
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 diagnostic turns ->
coordinator adjudication -> bounded owner repair -> independent regression
turns -> coordinator closeout -> ordinary repository verification.
- **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 seven named reports. Continuity is not a
dedicated seat; source fidelity, provenance, ownership, and spoiler safety
constrain every seat and remain part of final verification.
score cannot stand in for the seven 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
@ -63,13 +118,13 @@ 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.
Six 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.
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:
@ -80,11 +135,24 @@ The packet must distinguish:
- 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 critic returns
[[report-template]] as a separate artifact. Reports may cite the same defect,
but they must arrive independently.
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:
@ -118,10 +186,10 @@ 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, from that snapshot. Each original
seat receives the repaired candidate, its raw report, the relevant dispositions,
and the protected-material list, then returns [[regression-template]]. Each
answers only:
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?
@ -134,6 +202,12 @@ 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.
@ -142,11 +216,13 @@ let sequential reviews certify different bodies.
A deepening pass cannot close until:
- all seven diagnostic reports exist, including explicit no-finding reports;
- 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;
- all seven regression reports exist and name the same repaired candidate
identity;
- 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 coordinator records what must now be left alone;
- seed, wikilink, vault-layout, publication-boundary, Quartz, rendered-page,

View file

@ -0,0 +1,29 @@
# Shared Lore Guardian Prompt
Prepend this prompt to the seat's specialist prompt when its recorded mode is
`Guardian`. Use [[guardian-template]] for the report.
> You are one independent Guardian on the Kalsa Lore Review Council. Your seat
> has exhausted a named surface under a recorded immutable baseline. Guard that
> result against later changes; do not perform another general corpus critique.
>
> Read the complete Guardian packet: certified baseline identity, candidate
> identity, exact diff, full text and hashes of every changed note in remit,
> affected navigation or ownership context, and protected material. Inspect the
> changed material only. Context explains the change but is not permission to
> search untouched prose for work.
>
> Apply the seat's version-controlled specialist remit. Check whether the diff
> reintroduces a known defect, weakens a protected success, or creates a new
> defect within that remit. Diagnose exact changed passages; do not edit canon,
> invent missing truth, or use detector scores, aggregate scores, or word bans.
>
> If changed evidence suggests a wider recurrence, say why and require an early
> full sanity check. Do not claim an untouched corpus defect from evidence you
> were not assigned to inspect. A local defect blocks this candidate. A
> systemic recurrence, failed sanity check, or obsolete baseline returns the
> seat to Critic mode.
>
> Return `Pass` when the change preserves the guarded surface. Silence is not a
> report, and findings must not be manufactured to justify the seat. Use the
> Guardian change-report contract exactly.

View file

@ -0,0 +1,59 @@
---
pass_id: KALSA-000
guardian: replace-me
mode: guardian
phase: change-review
certified_baseline_id: replace-me
candidate_id: KALSA-000-repaired-01
diff_base: replace-me
verdict: pass
---
# Guardian Change Report
## Packet verification
- Certified baseline and guarded surface:
- Candidate identity and changed-note hashes:
- Diff base inspected:
- Complete changed notes inspected:
- Affected navigation or ownership context inspected:
## Guarded findings
For each defect in the changed surface record:
- **Finding ID:** preserve the Guardian's critic-local namespace.
- **Priority:** Essential, High-value, or Optional.
- **Location:** exact changed note and passage.
- **Evidence in the diff:** what new or altered language creates the defect.
- **Guarded invariant or known pattern:** what the certified baseline taught the
Guardian to protect.
- **Effect:** what the change does to lore, play, or reading experience.
- **Repair direction:** the smallest useful change without drafting canon.
- **Collateral risk:** what a careless correction would damage.
If no defect exists, say what changed and why the guarded strengths remain
intact. Do not search untouched corpus for work during a change review.
## Recurrence and escalation
- Does evidence in the changed notes suggest a pattern beyond this diff?
- Is an early full sanity check required?
- Should this seat remain Guardian or return to Critic mode?
## Protected strengths
- What survived in the changed surface?
- What should be left alone?
## Counter update
- Relevant Guardian reviews before this report:
- Relevant Guardian reviews after this report:
- Canonical notes added to the next sanity surface:
- Next mandatory sanity check:
## Verdict
- `Pass`, `Block`, or `Return to Critic`, with concise reason.

View file

@ -16,6 +16,9 @@ public_blind_packet: false
- Adjacent files supplied for context:
- Reader entry path under test:
- Coordinator:
- Council mode snapshot:
- Guardians omitted from baseline diagnosis:
- Guardian sanity trigger, if any:
## Authority and provenance

View file

@ -9,15 +9,15 @@ coordinator: replace-me
## Report inventory
| Critic | Diagnostic report | Verdict | Regression report |
| --- | --- | --- | --- |
| Institutional Causality | | | |
| Culture and Situated Knowledge | | | |
| Material Life and Ecology | | | |
| History, Conflict, and Change | | | |
| Ghostlight Play and Consequence | | | |
| AI-Isms and Prose | | | |
| Reader Experience and Navigation | | | |
| Seat | Mode | Diagnostic or sanity report | Verdict | Regression or Guardian change report |
| --- | --- | --- | --- | --- |
| Institutional Causality | Critic | | | |
| Culture and Situated Knowledge | Critic | | | |
| Material Life and Ecology | Critic | | | |
| History, Conflict, and Change | Critic | | | |
| Ghostlight Play and Consequence | Critic | | | |
| AI-Isms and Prose | Guardian | n/a unless sanity triggered | | |
| Reader Experience and Navigation | Critic | | | |
## Protected-material ledger
@ -40,9 +40,11 @@ coordinator: replace-me
## Regression closeout
- Repaired packet identity:
- Guardian diff base and changed-note packet identity:
- Essential findings resolved:
- Essential findings remaining:
- Regressions accepted for repair:
- Preferences rejected or deferred:
- What must now be left alone:
- Guardian counter and mode updates:
- Final council decision: