Observation
All propositions in PropositionRepository are effectively peers. confidence and importance exist, but there's no framework-level concept of which propositions matter most — which should be retained vs. evicted under pressure, which should win in conflicts, or which an LLM should comply with more strongly.
This matters when:
- Token budgets are limited — not all propositions fit in a prompt; which get included?
- Knowledge conflicts arise — two conflicting propositions exist; which wins?
- Memory pressure — too many propositions; which get evicted?
- Compliance directives — some facts must be respected ("patient is allergic to penicillin") while others are suggestions
What DICE already has
confidence: ZeroToOne — LLM's extraction-time certainty. Doesn't evolve post-extraction (except through effectiveConfidence() decay).importance: ZeroToOne — LLM's assessment of how much a fact matters. Set at extraction time, doesn't change.decay: ZeroToOne — staleness rate. Feeds effectiveConfidence() but doesn't drive behavior beyond that.reinforceCount: Int — tracks merge/reinforcement frequency. Stored but not acted on.status: PropositionStatus — ACTIVE, SUPERSEDED, CONTRADICTED, PROMOTED. Lifecycle state, not priority.
The gap: confidence and importance are extraction-time snapshots. There's no lifecycle property that evolves based on reinforcement, decay, conflict resolution, or operator action.
The question
Is confidence + importance sufficient, or do propositions need a priority model that evolves over their lifecycle?
Some possibilities:
-
Make existing fields dynamic — instead of new fields, let confidence and importance be updated by reinforcement, decay, and revision. reinforceCount already tracks frequency; a policy could boost confidence based on it.
-
Add a rank field — an integer priority score (e.g., 0 = unranked, higher = more important). Influences retrieval ordering, eviction priority, conflict resolution. Evolves over time based on reinforcement and decay.
-
Add an authority level — an enum indicating compliance strength, e.g.:
| Level |
Semantics |
| PROVISIONAL |
May consider — unverified, low confidence |
| RELIABLE |
Should respect — well-established |
| CANON |
Must respect — ground truth, never auto-demoted |
Authority transitions: reinforcement promotes (PROVISIONAL → RELIABLE), decay demotes (RELIABLE → PROVISIONAL), CANON requires explicit operator action.
-
Rank + authority together — rank = relative importance within an authority tier. Authority = how strongly the LLM should comply. Orthogonal dimensions.
Why not just use importance?
importance is a 0-1 float set at extraction time reflecting the LLM's assessment. Rank and authority would be lifecycle properties that evolve based on what happens after extraction:
importance: "How important did the extractor think this was?"rank: "How important is this proposition now, given everything since extraction?"authority: "How strongly should an LLM comply with this?"
Observation
All propositions in
PropositionRepositoryare effectively peers.confidenceandimportanceexist, but there's no framework-level concept of which propositions matter most — which should be retained vs. evicted under pressure, which should win in conflicts, or which an LLM should comply with more strongly.This matters when:
What DICE already has
confidence: ZeroToOne— LLM's extraction-time certainty. Doesn't evolve post-extraction (except througheffectiveConfidence()decay).importance: ZeroToOne— LLM's assessment of how much a fact matters. Set at extraction time, doesn't change.decay: ZeroToOne— staleness rate. FeedseffectiveConfidence()but doesn't drive behavior beyond that.reinforceCount: Int— tracks merge/reinforcement frequency. Stored but not acted on.status: PropositionStatus— ACTIVE, SUPERSEDED, CONTRADICTED, PROMOTED. Lifecycle state, not priority.The gap:
confidenceandimportanceare extraction-time snapshots. There's no lifecycle property that evolves based on reinforcement, decay, conflict resolution, or operator action.The question
Is
confidence+importancesufficient, or do propositions need a priority model that evolves over their lifecycle?Some possibilities:
Make existing fields dynamic — instead of new fields, let
confidenceandimportancebe updated by reinforcement, decay, and revision.reinforceCountalready tracks frequency; a policy could boostconfidencebased on it.Add a
rankfield — an integer priority score (e.g., 0 = unranked, higher = more important). Influences retrieval ordering, eviction priority, conflict resolution. Evolves over time based on reinforcement and decay.Add an
authoritylevel — an enum indicating compliance strength, e.g.:Authority transitions: reinforcement promotes (PROVISIONAL → RELIABLE), decay demotes (RELIABLE → PROVISIONAL), CANON requires explicit operator action.
Rank + authority together — rank = relative importance within an authority tier. Authority = how strongly the LLM should comply. Orthogonal dimensions.
Why not just use
importance?importanceis a 0-1 float set at extraction time reflecting the LLM's assessment. Rank and authority would be lifecycle properties that evolve based on what happens after extraction:importance: "How important did the extractor think this was?"rank: "How important is this proposition now, given everything since extraction?"authority: "How strongly should an LLM comply with this?"