fix: bescherm excluded en locked topics tegen AI-deletion/merge
All checks were successful
On Pull Request to Main / test (pull_request) Successful in 46s
On Pull Request to Main / publish (pull_request) Successful in 1m26s
On Pull Request to Main / deploy-dev (pull_request) Successful in 1m49s

Topics met learning_relevance="exclude" of relevance_locked=true werden
door de Full Analysis stilletjes verwijderd: het model zag exclude als
"irrelevant" en stelde ze voor in actions.deletions / actions.merges,
en analyzeGraph paste die zonder filter toe. Het locked-vlaggetje werd
in full-scope payload niet eens meegestuurd, dus zelfs een nieuwe prompt
kon niet helpen.

Defense-in-depth fixes:

1. Pure graphGuard.filterAiActions strips elke deletion/merge waarvan de
   target excluded of locked is, vóór bulkSave. Merges waarin een
   protected topic juist de keepId is (canonical survivor) blijven door.

2. SYSTEM_PROMPTS.full krijgt een expliciete "PROTECTED TOPICS" sectie
   die exclude en locked topics als nooit-te-verwijderen markeert.

3. relevance_locked wordt nu ook in de full-scope compactTopics payload
   meegestuurd, zodat het model de vlag überhaupt ziet.

4. UI-feedback: GraphControls toont een ShieldCheck banner met aantal
   geblokkeerde acties na de analyze, en het excluded-aantal naast de
   "Show Excluded Nodes" toggle (visible/hidden).

5. 13 nieuwe Vitest cases dekken protected detection en alle drop/keep
   paden van filterAiActions.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
RaymondVerhoef
2026-06-04 08:22:12 +02:00
parent e310b6d85b
commit 9395ea11fe
4 changed files with 267 additions and 6 deletions

View File

@@ -0,0 +1,125 @@
import { describe, expect, it } from 'vitest';
import { isProtectedTopic, filterAiActions } from '../graphGuard';
const topic = (id, extras = {}) => ({
id,
label: `Topic ${id}`,
type: 'concept',
learning_relevance: 'standard',
relevance_locked: false,
...extras,
});
describe('isProtectedTopic', () => {
it('flags excluded topics', () => {
expect(isProtectedTopic(topic('a', { learning_relevance: 'exclude' }))).toBe(true);
});
it('flags locked topics', () => {
expect(isProtectedTopic(topic('a', { relevance_locked: true }))).toBe(true);
});
it('flags topics that are both excluded and locked', () => {
expect(
isProtectedTopic(topic('a', { learning_relevance: 'exclude', relevance_locked: true })),
).toBe(true);
});
it('does not flag standard / core / peripheral topics', () => {
expect(isProtectedTopic(topic('a', { learning_relevance: 'core' }))).toBe(false);
expect(isProtectedTopic(topic('a', { learning_relevance: 'standard' }))).toBe(false);
expect(isProtectedTopic(topic('a', { learning_relevance: 'peripheral' }))).toBe(false);
});
it('treats null/undefined as not protected', () => {
expect(isProtectedTopic(null)).toBe(false);
expect(isProtectedTopic(undefined)).toBe(false);
});
});
describe('filterAiActions', () => {
const topics = [
topic('safe'),
topic('excluded', { learning_relevance: 'exclude' }),
topic('locked', { relevance_locked: true }),
topic('both', { learning_relevance: 'exclude', relevance_locked: true }),
topic('other'),
];
it('drops deletions that target excluded topics', () => {
const { filtered, dropped } = filterAiActions(topics, {
deletions: ['safe', 'excluded', 'other'],
merges: [],
});
expect(filtered.deletions).toEqual(['safe', 'other']);
expect(dropped.deletions).toEqual(['excluded']);
});
it('drops deletions that target locked topics', () => {
const { filtered, dropped } = filterAiActions(topics, {
deletions: ['locked'],
merges: [],
});
expect(filtered.deletions).toEqual([]);
expect(dropped.deletions).toEqual(['locked']);
});
it('drops merges whose deleteId is protected', () => {
const merges = [
{ keepId: 'safe', deleteId: 'other' }, // ok
{ keepId: 'safe', deleteId: 'excluded' }, // drop — excluded
{ keepId: 'other', deleteId: 'locked' }, // drop — locked
{ keepId: 'safe', deleteId: 'both' }, // drop — both flags
];
const { filtered, dropped } = filterAiActions(topics, { merges, deletions: [] });
expect(filtered.merges).toEqual([{ keepId: 'safe', deleteId: 'other' }]);
expect(dropped.merges).toHaveLength(3);
expect(dropped.merges.map(m => m.deleteId)).toEqual(['excluded', 'locked', 'both']);
});
it('keeps merges where a protected topic is the keepId (canonical survivor)', () => {
const merges = [
{ keepId: 'excluded', deleteId: 'safe' }, // ok — protected is survivor
{ keepId: 'locked', deleteId: 'other' }, // ok — protected is survivor
];
const { filtered, dropped } = filterAiActions(topics, { merges, deletions: [] });
expect(filtered.merges).toEqual(merges);
expect(dropped.merges).toEqual([]);
});
it('passes through other action keys untouched', () => {
const actions = {
deletions: [],
merges: [],
newRelations: [{ source: 'a', target: 'b', type: 'related_to' }],
relevanceUpdates: [{ id: 'safe', learning_relevance: 'peripheral' }],
};
const { filtered } = filterAiActions(topics, actions);
expect(filtered.newRelations).toEqual(actions.newRelations);
expect(filtered.relevanceUpdates).toEqual(actions.relevanceUpdates);
});
it('tolerates missing actions keys', () => {
const { filtered, dropped } = filterAiActions(topics, {});
expect(filtered.deletions).toEqual([]);
expect(filtered.merges).toEqual([]);
expect(dropped.deletions).toEqual([]);
expect(dropped.merges).toEqual([]);
});
it('tolerates null actions', () => {
const { filtered, dropped } = filterAiActions(topics, null);
expect(filtered.deletions).toEqual([]);
expect(filtered.merges).toEqual([]);
expect(dropped.deletions).toEqual([]);
expect(dropped.merges).toEqual([]);
});
it('drops references to unknown ids by treating them as not protected (delete pass-through)', () => {
// If the AI hallucinates an id, the deletion is harmless downstream (no
// topic by that id exists). We do NOT block on unknown ids — that would
// mask real bugs. Verify pass-through behavior.
const { filtered } = filterAiActions(topics, { deletions: ['ghost-id'], merges: [] });
expect(filtered.deletions).toEqual(['ghost-id']);
});
});

81
src/lib/graphGuard.js Normal file
View File

@@ -0,0 +1,81 @@
/**
* Knowledge-graph safety guards.
*
* The "Full Analysis" pass lets the LLM propose `merges` (collapse two topics
* into one) and `deletions` (drop a topic entirely). Both are destructive and
* persisted via bulkSave → db.saveTopics, which wipes-and-rewrites the
* `topics` collection.
*
* Two classes of topic must NEVER be removed by an AI suggestion:
*
* 1. `learning_relevance === 'exclude'`
* Reference material — kept on purpose, surfaced only to R42 search,
* intentionally excluded from theme topics. An admin made this choice;
* the AI's "irrelevant" judgement must not override it.
*
* 2. `relevance_locked === true`
* Admin pinned this row's relevance. By extension the row itself is
* also pinned — deleting it would be a louder change than re-scoring it.
*
* `filterAiActions` strips any merge/deletion that would touch a protected id
* BEFORE it reaches bulkSave. The model still gets to suggest them (we can't
* stop it generating), but those suggestions are dropped client-side.
*
* The function is intentionally pure so it is trivially unit-testable.
*/
/**
* @param {{ learning_relevance?: string, relevance_locked?: boolean }} topic
*/
export function isProtectedTopic(topic) {
if (!topic) return false;
if (topic.learning_relevance === 'exclude') return true;
if (topic.relevance_locked === true) return true;
return false;
}
/**
* Strip protected topics out of `actions.deletions` and `actions.merges`.
*
* Merges have two ids:
* - `keepId` — the survivor
* - `deleteId` — gets removed and its relations re-pointed
*
* If `deleteId` is protected we drop the entire merge (we will not delete a
* protected topic, even to consolidate it). If `keepId` is protected we keep
* the merge — folding others into a protected canonical topic is fine.
*
* @param {object[]} currentTopics
* @param {{ merges?: {keepId: string, deleteId: string}[], deletions?: string[], newRelations?: object[], relevanceUpdates?: object[] }} actions
* @returns {{ filtered: object, dropped: { deletions: string[], merges: object[] } }}
*/
export function filterAiActions(currentTopics, actions) {
const byId = new Map(currentTopics.map(t => [t.id, t]));
const isProtected = (id) => isProtectedTopic(byId.get(id));
const droppedDeletions = [];
const keptDeletions = [];
for (const id of actions?.deletions ?? []) {
if (isProtected(id)) droppedDeletions.push(id);
else keptDeletions.push(id);
}
const droppedMerges = [];
const keptMerges = [];
for (const m of actions?.merges ?? []) {
if (isProtected(m?.deleteId)) droppedMerges.push(m);
else keptMerges.push(m);
}
return {
filtered: {
...(actions || {}),
deletions: keptDeletions,
merges: keptMerges,
},
dropped: {
deletions: droppedDeletions,
merges: droppedMerges,
},
};
}