[SLICE 17][CORE-E] Implement CITADEL evidence-use roles, candidate exposure/dependence closure, contamination lineage, and validation-independence queries #14

New issue

Open

opened 2026-09-08 01:57:00 +00:00 by nsaspy · 1 comment

nsaspy commented

2026-09-08 01:57:00 +00:00

Owner

Parent epic: #16
Depends on: #34 #25 #27
Research: prolog-rlm#404B

Goal

Persist/query the evidence-contamination structure required to prevent self-confirming learned rules from laundering their own validation.

Implement durable records/indexes for:

evidence-use role: generation / selection / validation / audit / post-admission;
candidate exposure to sources, semantic records, fixtures, metrics, verifier outputs, ancestor candidates, learner/model artifacts;
source/causal/dependence groups;
adaptive holdout reveal/reclassification events;
evaluator/verifier/metric/policy fingerprints;
candidate influence/tamperability declarations;
validation-independence queries and explicit unknown/partial status.

Acceptance

A descendant candidate cannot count a previously exposed failure as fresh held-out validation; copied/paraphrased/derived evidence remains dependent when lineage proves it; record-count independence never substitutes for causal/source independence; missing dependence evidence returns unknown rather than independent.

Parent epic: #16 Depends on: #34 #25 #27 Research: prolog-rlm#404B ## Goal Persist/query the evidence-contamination structure required to prevent self-confirming learned rules from laundering their own validation. Implement durable records/indexes for: - evidence-use role: generation / selection / validation / audit / post-admission; - candidate exposure to sources, semantic records, fixtures, metrics, verifier outputs, ancestor candidates, learner/model artifacts; - source/causal/dependence groups; - adaptive holdout reveal/reclassification events; - evaluator/verifier/metric/policy fingerprints; - candidate influence/tamperability declarations; - validation-independence queries and explicit unknown/partial status. ## Acceptance A descendant candidate cannot count a previously exposed failure as fresh held-out validation; copied/paraphrased/derived evidence remains dependent when lineage proves it; record-count independence never substitutes for causal/source independence; missing dependence evidence returns unknown rather than independent.

nsaspy referenced this issue

2026-09-08 01:57:00 +00:00

[SLICE 12][CORE-C] Implement RACE durable cursor/cache generations, invalidation fingerprints, and replayable retrieval receipts #19

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 11][CORE-C] Implement TAPS/OATH bounded semantic projection with mandatory dependency and counterevidence closure #20

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 10][CORE-C] Implement general semantic query API over current/history/support/reasoning cuts #21

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 09][CORE-C] Implement reversible identity/ontology reconciliation and source-dependence groups #22

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 08][CORE-C] Build general semantic indexes and deterministic record/source lookup surfaces #23

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 07][CORE-C] Persist/rebuild epistemic support graph, justification DAG, provenance classes, and counterevidence indexes #24

nsaspy referenced this issue

2026-09-08 01:57:36 +00:00

[semantic-memory] Cross-memory entity/concept reconciliation, aliasing, ontology alignment, and reversible identity links #39

nsaspy referenced this issue

2026-09-08 01:57:39 +00:00

[EPIC][CORE] Library-first Symbolic Memory v1: durable semantic world model, learning history, federation, and zero-model recall #38

nsaspy commented

2026-09-08 01:59:15 +00:00

Author

Owner

MACHINE-SPIRIT #404D implementation refinement — exposure/dependence indexing

Implement canonical direct typed dependence/exposure edges + exact query-scoped closure, not a permanently materialized global transitive closure.

Required indexes:

forward adjacency by candidate/source/evidence/fixture/evaluator/ancestor;
reverse dependency adjacency for invalidation;
generation/version on active dependence mappings;
generation-keyed cached closure digest and exact-completeness status.

Worst-case exact closure is O(V+E) over the reachable subgraph. Global closure storage is not required and can approach O(V²) on dense graphs. Semi-naive/delta maintenance is the first optimization; DBSP/differential-style machinery is optional after benchmark evidence.

Hard rules:

probabilistic/Bloom indexes may conservatively say possibly_exposed, never prove independent=true;
missing/private provenance -> independence=unknown;
do not use irreversible union-find as canonical origin truth because dependence/identity links can be superseded/retracted;
descendant exposure includes ancestor exposure under the declared lineage profile;
closure cache invalidates on exact direct-edge/source/identity/translation generations.

Refs prolog-rlm#404D and #27/#38.

## MACHINE-SPIRIT #404D implementation refinement — exposure/dependence indexing Implement **canonical direct typed dependence/exposure edges + exact query-scoped closure**, not a permanently materialized global transitive closure. Required indexes: - forward adjacency by candidate/source/evidence/fixture/evaluator/ancestor; - reverse dependency adjacency for invalidation; - generation/version on active dependence mappings; - generation-keyed cached closure digest and exact-completeness status. Worst-case exact closure is `O(V+E)` over the reachable subgraph. Global closure storage is not required and can approach `O(V²)` on dense graphs. Semi-naive/delta maintenance is the first optimization; DBSP/differential-style machinery is optional after benchmark evidence. Hard rules: - probabilistic/Bloom indexes may conservatively say `possibly_exposed`, never prove `independent=true`; - missing/private provenance -> `independence=unknown`; - do not use irreversible union-find as canonical origin truth because dependence/identity links can be superseded/retracted; - descendant exposure includes ancestor exposure under the declared lineage profile; - closure cache invalidates on exact direct-edge/source/identity/translation generations. Refs prolog-rlm#404D and #27/#38.