AI/Tasks/MemoryUpgrade/Memory_Phase04_ConfidenceScoring_Strict.txt
|
STRICT MODE: Redesign confidence scoring in the memory subsystem.
Mission: - Complete this phase with bounded steps and deterministic fallback behavior. - Apply changes ONLY within the memory subsystem and its tests. Concrete scope: - src/TechToolbox.Agent/Memory/MemoryStore.cs - src/TechToolbox.Agent/Memory/MemoryModels.cs - src/TechToolbox.Agent/Memory/MemoryLearner.cs - src/TechToolbox.Agent/Tests/*Memory*.cs Requirements: - Replace additive confidence formula that saturates quickly at 1.0. - Set bounded initial confidence below healthy/high-confidence threshold. - Include at minimum: evidence strength, recency decay, source reliability, contradiction/conflict state. - Preserve confidence correctly across index rebuild operations. - Keep formula deterministic and explainable in code/tests. - Do not modify unrelated subsystems. Step budget (hard limits): 1) Discovery: <= 6 read/search calls. 2) Edits: <= 9 file edits total. 3) Validation: 1 build + 1 targeted test run, with exactly 1 repair cycle allowed. 4) Stop when criteria pass. Loop guards: - Do not re-tune coefficients more than 2 times. - If confidence bounds fail repeatedly with same pattern, stop and report coefficients + failing cases. Fallback behavior: - Fallback A: implement a simple bounded logistic-like or weighted bounded model before advanced refinements. - Fallback B: if aging model complexity causes instability, ship deterministic piecewise decay with tests. - Fallback C: after one failed repair cycle, stop and output FAILED_VALIDATION. Acceptance criteria: - Tests cover first observation, repeat observation, aging decay, contradiction/conflict adjustment. - Confidence remains within defined bounds and progresses smoothly. - Index rebuild does not reset valid confidence state. Validation commands: - dotnet build src/TechToolbox.Agent/TechToolbox.Agent.csproj -c Release - dotnet test src/TechToolbox.Agent/Tests/TechToolbox.Agent.Tests.csproj -c Release --filter "Memory|Confidence|Learner|Index" Output results in markdown using this structure: 1) Summary 2) Files changed 3) Scoring model 4) Migration/compatibility notes 5) Tests added/updated 6) Validation results 7) Step-budget report (planned vs actual counts) |