Identify
Name the requesting client and record its purpose.
Research direction / PACT
The minimum context that still earns the answer.
OBJECTIVE
answer usefulness
−token cost
−disclosure risk
subject to permission + provenanceMore context is not automatically better context. Ninai explores retrieval as a constrained disclosure problem: preserve answer usefulness while reducing irrelevant tokens and unnecessary exposure.
Honest status
PACT is a working engineering method in the MVP. It is not presented as academically novel until it is evaluated against credible baselines.
Pipeline / 01
Constrain
Resolve explicit grants before any search begins.
Retrieve
Search only rows within the allowed scope set.
Rank
Combine relevance, type-aware freshness, importance, and confidence.
Compose
Select the highest-value evidence that fits the packet budget.
Log
Record the exact memory IDs, scopes, purpose, and estimated tokens.
Current signals / 02
Inspectable weights, not hidden magic.
The MVP uses deterministic weights that can be tested and changed. They are evaluation hypotheses, not immutable product truth.
SignalWeightQuestion
Lexical relevance0.34
Does the fact address this request?
Search position0.18
How strongly did local retrieval surface it?
Freshness0.16
Is this memory current for its type?
Importance0.14
Was the outcome marked consequential?
Confidence0.12
How reliable was extraction or entry?
Reinforcement0.06
Has this memory proved useful before?
Evaluation plan / 03
Compare the packet, not just the search result.
Ninai should be measured against full history, lexical top-k, semantic top-k, and future graph retrieval using the same questions and source evidence.
Answer accuracyDid the model answer correctly?
Context precisionHow much released context was useful?
Temporal accuracyDid it prefer the current fact?
AbstentionDid it stop when evidence was missing?
Scope leakageDid any forbidden memory appear?
Token reductionHow much context stayed local?
ProvenanceCan each claim be traced?
LatencyIs the policy layer fast enough?
Next experiments / 04
Earn complexity one benchmark at a time.
- 01 Build deterministic long-session fixtures.
- 02 Measure lexical PACT against full history.
- 03 Add a rebuildable local vector adapter.
- 04 Test temporal supersession and conflict handling.
- 05 Publish results—including failures.