Adaptive Learning

Adaptive Learning · models interpret feedback; deterministic state controls durable authority

INTERPRET  EVIDENCE  HYPOTHESIS  EVAL  AUTHORIZE  ROLLBACK

简体中文 · Docs Home

Adaptive Learning

NovelForge learns from explicit user/project evidence without letting a model turn an interpretation into durable authority.

runtime/session state != learning evidence != Project Canon
semantic interpretation != promotion judgment != write authorization

learning/author_model.py remains a projection/capture layer over the existing Learning Store. It is not a second preference database.

1. Semantic interpretation belongs to the model

learning.preference_interpret interprets supplied feedback and proposes the narrowest plausible scope:

one_off | project | user_taste | general_craft

It may explain the underlying mechanism, desired/avoid behavior, exceptions, uncertainty and conflicts with prior hypotheses. It does not grant durability or activation.

The model, not Python thresholds, decides whether evidence semantically supports a stable scope/mechanism.

2. Learning Store owns durability, not meaning

The deterministic store owns:

• evidence/hypothesis IDs and provenance;
• versioned state and contradiction/supersession records;
• exact source references;
• persistence and rollback history;
• consume-once result handling;
• Project/user scope isolation.

A durable record can still be tentative/contested. Persistence does not make an inference true.

3. Promotion Gate binds semantic review to authority

learning/promotion_gate.py no longer tries to prove semantic sufficiency with arbitrary evidence-count thresholds.

The semantic promotion review decides whether the supplied evidence actually supports the proposed scope/mechanism and whether important contradictions/counterexamples remain unresolved.

The deterministic gate then verifies objective prerequisites around that review, such as:

• exact contract/result/evidence binding;
• candidate scope and identity;
• required eval/counterexample artifacts where policy requires them;
• version/rollback/CI references for General Craft;
• explicit write authorization supplied by the surrounding authority mechanism.

A passing promotion review is a prerequisite. It is not permission to write.

4. Active != relevant

An active Author Model hypothesis means it is durably eligible for future use. It does not mean every production invocation should receive it.

The Author Model exposes a compact active index. The manager/model explicitly selects the active hypothesis IDs relevant to the current task. Deterministic code verifies that selected IDs are active and scope-compatible before returning details.

This prevents context pollution from automatically injecting every learned preference.

Current explicit user instruction remains stronger than an inferred or durable preference when they conflict.

5. Scope-specific authority

One-off

Used for the current repair/task only unless new evidence is captured separately.

Project preference

May activate only under the Project’s explicit preference-write authority. It never changes Framework behavior.

Durable user taste

Requires both:

1. a current bound promotion prerequisite result for the same mechanism/scope; and
2. explicit durable-user-taste write authorization.

Neither the model nor Promotion Gate can self-grant this permission.

General Craft

General Craft remains a Framework SYSTEM-IMPROVE concern. It requires stronger counterexample/eval/compatibility/version/rollback evidence and explicit Framework promotion authority. Production feedback cannot auto-promote it.

6. Corpus/research is evidence gathering, not truth by ingestion

A semantic learning agent may identify an evidence gap and search for lawful contrast/counterexample material. Search/retrieval strategy remains model-owned inside allowed capabilities.

Corpus discovery does not imply ingestion; ingestion does not imply Canon; corpus analysis does not imply promotion.

Rights, provenance, source identity and Project/user isolation remain deterministic boundaries.

7. Contradiction and rollback are first-class

New feedback may:

• strengthen a hypothesis;
• narrow its applicability;
• mark it contested;
• split an over-broad mechanism;
• supersede an older hypothesis;
• deprecate a behavior when evidence changes.

“Strengthening” means new independent evidence, not repeated model agreement or elapsed time.

8. Production use

explicit feedback
→ semantic preference interpretation
→ source-bound evidence
→ revisable hypothesis
→ semantic promotion review when durable activation is proposed
→ deterministic authority/prerequisite validation
→ active eligibility
→ model selects relevant active hypothesis IDs for a future task
→ production observes outcomes
→ new evidence may revise/supersede the hypothesis

The Writer/Editor never receive a hidden global style profile simply because records exist in learning.db.

9. Privacy

Personal preference evidence is user-scoped and is not committed to the generic Framework repository by default. NovelForge must not infer unrelated demographic/profile attributes from fiction preferences.

Exact implementation boundaries

• learning/learning_store.py — durable evidence/hypothesis/candidate/promotion history.
• learning/promotion_gate.py — deterministic binding/authority checks around model-owned promotion review.
• learning/author_model.py — bounded feedback capture, contradiction/supersession, scope-aware activation binding, active index and explicit selected projection.
• harness/semantic_workers/contracts/production-loop.json — learning.preference_interpret.
• Framework self-improvement protocol — General Craft promotion authority.

_{Interpret semantically. Persist cautiously. Activate only with evidence and authority. 🌸}