Skip to content

evolution-reference.md

Latest commit

 

History

History
65 lines (48 loc) · 3.01 KB

evolution-reference.md

File metadata and controls

65 lines (48 loc) · 3.01 KB

Evolution Reference

Slim reference for still-relevant decisions from the original evolution plan (Dwarf in the Glass, approved 2026-03-23). Phases 1-4 are complete and implemented.

Phase 5: Domain Generalization (Conditional)

Status: Not started. Proceeds only if generalizability test passes.

Test Criteria (ALL must hold)

Criterion Requirement
Non-absence triggers At least two automated positive triggers that are NOT "absence of correction within time window"
Frequency sufficiency Projected trigger frequency sufficient to reach n=50 within 6 months per rule
Noise rate Estimated false-positive/false-negative rate below 20%
Signal latency Outcome signal arrives within 24 hours
Automation Triggers can be evaluated without domain-expert involvement

Defined Failure Modes

Failure Example
Absence-only signals "User doesn't object within 5 minutes" as positive trigger
Insufficient frequency Rule applied only during quarterly deploys
High noise ratio User accepts mediocre suggestion because correcting takes effort
Expert-dependent Evaluating correctness requires a compliance officer
Latency mismatch Outcome visible 90 days later

If any criterion fails, that domain requires a different feedback architecture -- not a parameter tweak.

What changes if test passes

  • New domains added to graph (no schema changes needed)
  • Domain-specific scope values via extensible enum
  • Domain-specific increment triggers
  • Per-framework enforcement wrappers (built when adopted, not speculatively)

Design Decisions Log

Decision Chosen Rationale
Scope generalization Extensible enum per domain "Entity" means nothing. Domain-specific values are honest.
Origin context storage SQLite, not Neo4j Write-once blob with no relationships. Graph provides no benefit.
Authority enforcement Hard preference rule, not weight-based 0.099 weight creates 5% tiebreaker, not suppression. Hard rule makes invariant explicit.
Rejected rules logging Delete, don't log No concrete consumer. Add logging when a deliverable needs the data.
Graduation threshold n=50 At n=10, Wilson 95% CI is [0.35, 0.93]. At n=50, CI is [0.61, 0.86]. Minimum for statistical meaning.
Gate naming "Structural pre-filter" Gate catches structural failures, not correctness. Honest naming.
Phase ordering Gate design before memory plumbing Resolves bootstrapping paradox -- define the filter before the plumbing it filters.
Preference threshold Empirical median adjacent-candidate gap Data exists in 83 queries. Don't guess when you can measure.

What Not to Build

Item Why
Scope renames (file->entity, module->component) False generality
Frequency tracking for non-coding domains (before Phase 5 test) Feedback signal problem unsolved
Multi-user / multi-tenant No concrete second user
"AI learning mirrors human learning" framing Metaphorical, not mechanistic