Beads work ledger — hash-ID tasks, atomic bd update --claim, dependency-aware bd ready, intent memory bd remember/prime. PAUL emits beads instead of markdown phases.
Gate 3 verification — permission-denied holdout scenarios, write-tool-less reviewers, evidence files over self-report, veto-rate logging.
Unified bi-temporal knowledge graph over GBrain (PostgreSQL/pgvector) and Graphify (NetworkX/tree-sitter). Treats both stores as a single citation-addressable graph. Query with `brain query --unified "topic"`.
Advanced Engram patterns — progressive disclosure index→timeline→full-fetch (~10x token savings), citation IDs, Seance (interrogate dead predecessors), temporal search.
Golden task eval harness — 20 deterministic tasks with checks.py, 5-run protocol, judge calibration, CI gate on >0.3 regression vs baseline. SWE-smith for auto-generation.
Two-mode loop system — in-session /goal evaluation and overnight fresh-context runner with dual exit gate, circuit breaker, error gate, and relay notes.
Trail-of-Bits two-phase mutmut campaign — targeted first, survivors get full suite, skeptical-agent rule on survivor killing.
Delta-specs against a living source-of-truth — brownfield-native, EARS acceptance criteria, blind-judge verifiable. Use when adding or changing a feature in an existing codebase that already has documented behavior. Produces ~1/3 the markdown of a full spec-kit spec by expressing only what changes.