aide-the-quad · git:20260826.8da8b03 · 2026-08-26 · sha256 4a5e153fd633fa21
aide-the-quad git:20260826.8da8b03A
Immutable. This exact content is served forever at /api/v1/blob/4a5e153fd633fa21.
--- name: aide-the-quad description: THE QUAD — closed-loop architecture binding Cockpit (trust+telemetry), Harness (discipline+verification), Orchestrator (execution+authority), and Model (cognition) with three reinforcement loops. Governs any feature touching component sync, the improvement flywheel, unified discipline sourcing, warm-start/offline readiness, or calibration data. Use before adding cross-component features or changing prompt/scaffold sources. --- # THE QUAD — One System, Four Loads, Three Loops Authoritative design: docs/THE-QUAD.md in-repo (read it first — this skill adds the LAW layer). ## Laws 1. **Single discipline source**: chat scaffold AND agent-loop system prompts compose from common/harness/credocore.md via ONE loader. Two hand-written prompt strings = drift = forbidden. 2. **Loop C capture is mandatory**: every approve/reject/ship emits an event to .aide/library/events.jsonl. Features that skip capture break the flywheel. 3. **[learned] blocks are earned**: only entries with >=5 paired outcomes and non-negative net effect inject; provenance count always shown; operator can purge the library. 4. **Warm-ready is a feature**: every engine path must trigger post-spawn warmup; AUTO_START is opt-in, RAM-guarded, session-flagged. 5. **Calibrate, don't guess**: thresholds (drift %, delegation defaults) become data-derived once N>=20 transcripts per model; policy defaults until then are labeled as such in code comments. ## Slice map S1 unified credo loader · S2 Loop C v0 (capture/distill/inject) · S3 fast-ready pack (warmup/auto-start/bench chips) · S4 calibration · S5 capability-aware SOP emphasis. Each slice: build -> live verify -> evidence -> journal -> commit. ## Integration - Consumes: credocore.md (credo-guardrail), ships.log + battery outputs, profile sidecars (inference-control), served-ctx probe (model-runtime). - Feeds: fine-tune dataset builder (post-training-closed-loop format), MODELS panel backend/verdict chips, rail badges. - Conflicts resolve toward: operator authority > deterministic gates > harness prose > model preference.