pilot-feedback

pilot-feedback

The compounding loop that turns real pilot-user usage (René today, Marcel next) into structured improvement signal. Both good and negative feedback are captured — negative feedback is the high-value ΔU/surprise signal (it points at what's broken/wrong/rejected); positive feedback confirms what to keep + compound.

When to invoke

• After fetching a pilot's branch (git fetch, review their recent commits) → ingest-git.
• When the operator says "good"/"bad" about a session/output/feature → add explicit.
• Periodically (or before a design pass) → report to see where the friction concentrates.

The mechanism (capability: pilot-feedback-capture)

Script: node _SYSTEM/Scripts/pilot-feedback.mjs <cmd> Ledger: _SYSTEM/state/pilot-feedback.jsonl (append-only JSONL via _SYSTEM/lib/jsonl.mjs).

command	what it does
add --pilot <name> --polarity good|bad|neutral [--source explicit|commit|session] [--tag <area>] [--note "..."]	record one explicit feedback event
ingest-git --pilot <name> --branch <ref> [--since 1d] [--author <name>]	parse a pilot's branch commits; classify each subject → good/bad + tag; record signal-bearing ones (neutral skipped)
report [--pilot <name>]	aggregate by polarity + tag, print recent 5

Tags (auto-classified)

voice · file-io · safety · memory · launcher · persona · tool · automation · other

Polarity heuristic (commit subjects)

• bad = the prior state was wrong/rejected: replac…/reject/broken/wrong/fix/rollback/revert/crash/hang/fail/missing/dead
• good = addition/build/enable: add/new/improve/ship/feature/enable/wire/scaffold/launch/builds
• A commit with BOTH (e.g., "Kokoro replaces robotic SAPI") → bad (it's a rejection of the prior).

How it compounds

1. Capture — ingest-git after every pilot fetch + add for in-session explicit feedback.
2. Report — before a design/build pass, run report to see where friction concentrates (high bad count in a tag = priority fix).
3. Route — negative feedback in a tag feeds the corresponding roadmap phase (e.g., voice bad-count → A0 voice quality; memory → A3 NEURO_CORE). Negative feedback = the ΔU/surprise that NEURO_CORE weights heavily (write_strength = |ΔU| × precision).
4. Synthesize (periodic, GLM lane) — a higher-level pass over the ledger produces improvement proposals → memory proposals → skill updates. (Future wiring; the deterministic report is the v1.)

Worked example (2026-07-05, René's branch)

Manual review of origin/rene surfaced: robotic TTS rejected (bad/voice), PDF/Word/Excel text extraction added (good→fix, bad/file-io on the prior raw-bytes), folder-scoped safety (good/safety), local FTS5 second-brain (good/memory), one-word launcher (good/launcher). That manual review is exactly what ingest-git automates:

node _SYSTEM/Scripts/pilot-feedback.mjs ingest-git --pilot rene --branch origin/rene --since 1d --author "René"
node _SYSTEM/Scripts/pilot-feedback.mjs report --pilot rene

Session Notes

• 2026-07-05: built v1 (add + ingest-git + report + hermetic tests 5/5). Ledger at _SYSTEM/state/pilot-feedback.jsonl. Feeds A0–A5 priority via the negative-feedback signal. Next: GLM synthesis pass for cross-session pattern extraction.