Files

Alexander Whitestone 0d64d8e559 initial: sovereign home — morrowind agent, skills, training-data, research, specs, notes, operational docs

Tracked: morrowind agent (py/cfg), skills/, training-data/, research/,
notes/, specs/, test-results/, metrics/, heartbeat/, briefings/,
memories/, skins/, hooks/, decisions.md, OPERATIONS.md, SOUL.md

Excluded: screenshots, PNGs, binaries, sessions, databases, secrets,
audio cache, timmy-config/ and timmy-telemetry/ (separate repos)

2026-03-27 13:05:57 -04:00

2.1 KiB

Raw Permalink Blame History

Next Cycle Priorities

Date: 2026-03-24
Context: Repository-based development suspended, focus on local Timmy implementation

Immediate Actions (Next Cycle)

1. FIX SOURCE DISTINCTION BUG [HIGH PRIORITY]

Problem: Models tag training data as [retrieved] when confident, [generated] when uncertain
Root Cause: Conflation of epistemic confidence with data source
Solution: Rewrite rule to explicitly check for tool-call sources vs training data
Test Plan: Re-run source-distinction tests with corrected rule

2. FIX REFUSAL OVER-AGGRESSION [HIGH PRIORITY]

Problem: Model refuses to answer even when context contains the answer
Root Cause: Refusal rule overpowers retrieval behavior
Solution: Add context-checking step before refusal trigger
Test Plan: Re-run Test D from refusal-rule-test-001.md

3. IMPLEMENT PIPELINE PROTOTYPE [MEDIUM PRIORITY]

Architecture: Generate → Tag → Filter → Deliver
Target: Simple working prototype for qwen3:30b testing
Location: ~/.timmy/machinery/ (new directory)
Goal: Prove the architecture works before optimization

4. LOCAL BRAIN VALIDATION [LOW PRIORITY]

Model: qwen3:30b on Ollama (the intended home brain)
Test: Both source distinction and refusal rules
Budget: num_predict ≥1000 for thinking tokens
Goal: Confirm rules work on the target local model

Testing Strategy

Use existing test plans in ~/.timmy/test-results/ but with corrected implementations. Target: all tests pass before considering production deployment.

Success Criteria

Source distinction correctly identifies tool-call vs training sources
Refusal rule catches fabricated specifics but answers from context
Pipeline prototype functional on local qwen3:30b
Test suite shows green across all scenarios

Files to Create/Update Next Cycle

~/.timmy/machinery/ (new directory for implementation)
Updated rule drafts in ~/.timmy/test-results/
Re-run test results with corrected logic
Pipeline prototype implementation

Sovereignty and service always.

2.1 KiB Raw Permalink Blame History