This repo is the case file and (eventually) the implementation of Holmes, the research-and-analysis brain of the WCJBT Non-Dev Builder OS and sibling of Alfred. Core thesis: "the method is the identity."
Read RECIPE.md before any work. It is the persistent dependency graph
of this repo: every subsystem, what it needs, and a checkable proof line
(a number, a filename, or a named output — never a status message).
- The session's
PLAN.mdis a subset of RECIPE.md: the subsystems being touched, in dependency order, with their proof lines carried forward verbatim. gate.ps1is the definition-of-done gate. Runpowershell -NoProfile -ExecutionPolicy Bypass -File gate.ps1before claiming done — it exits 0 only if every subsystem's proof passes. This is separate from the build gate (CIacdl-joint-gate/sbom-and-cve-scan); both must be green, neither substitutes for the other.gate.ps1 -SelfTestproves the gate can fail (planted proof, watched FAIL, restore) — run it after any edit to the gate itself.- Proof lines are platform-aware from birth (the
GOOSE_PATHcfg-pattern incrates/holmes-guard/tests/acdl1_spawn.rsis the convention). Environment preconditions (e.g. loopback127.0.0.1:11434free for the acdl1_proxy positive control) are stated in RECIPE.md, never worked around silently. - Drift rule (from RECIPE.md): a PR that touches a subsystem re-verifies its proof line. Drifted proofs are flagged in the PR body (old → new → why), never silently edited.
docs/holmes-spec-v2.md is the authoritative build reference. Where anything in this file, the case-file docs, or your memory disagrees with the spec, the spec wins. Revisions are drafted on the claude.ai pressure-testing surface and written back here; the repo copy wins on any disagreement (sync rule in the spec's header). Claim labels are the five system states (D-15, 2026-09-01): EXECUTED / VERIFIED-LIVE / CANON / REPORTED / UNVERIFIED — never silently harden caveated claims into facts.
This repo is documentation-only until Phase 0 runs from its kickoff prompt (holmes-claude-code-kickoff-phase0-v2.md, not yet committed — F-009) with explicit human go-ahead. Do not scaffold application code before that.
- Zero fabrication. Every external factual claim gets a primary source with a date, or an explicit
UNVERIFIEDtag. Never fill gaps with plausible detail. - Epistemic labels on all claims in any document you write here:
VERIFIED— cite the source and the date checkedINFERRED— show the reasoningASSUMED— flag itUNKNOWN— becomes a research item
- Evidence or it didn't happen. Every finding quotes its source (file + section) or states
ABSENTexplicitly. Distinguish "I checked and found nothing" from "I didn't check." - Licensing — AGPL-3.0-or-later, per D-01 (DECIDED 2026-07-18). Ratified by the human at Gate Zero with
Alfred/LICENSEquoted from disk (evidence indocs/audit/decisions.mdD-01);LICENSEcarries the AGPL v3 text byte-identical to Alfred's. Author no license identifiers other than AGPL-3.0-or-later (Apache-2.0 dependencies unaffected). The spec's §7 "Apache-2.0-compatible repo" line is superseded on this point — carried as a pending spec amendment (A-02), never silently propagated. - Vendor gate — a denylist, not an allowlist. No Meta, OpenAI, or xAI anywhere — direct or transitive: SDKs, models, APIs, infra, model weights (no Llama even locally). Google is permitted; open-weights-on-permitted-infra permitted. Violations are Blockers. Once code exists, the denylist is a runtime guard and a regression test.
- Rule 9 — consent before consequence. Proposals until approved. No commit/push without explicit go-ahead when building; nothing destructive or irreversible (relicensing, force-pushes, deletions, publishing, spending) without explicit human go-ahead, ever.
- RPI. Research → Plan → Implement, in that order, every time.
- Path-confined, deny-by-default tools. Read-only tools may run free; every write/shell action asks first.
- Born-redacted, local-only telemetry. Counts, durations, names — never content, prompts, or secrets.
- Supply-chain hygiene. Syft SBOMs, OSV-Scanner, Grype; no Trivy (CVE-2026-33634); pin GitHub Actions to full commit SHAs.
- Surveillance-detection-not-surveillance. Anti-doxxing refusals are permanent; tools scoped to power, never private citizens.
- "I answer to the block." Blacksky-style community accountability + human-in-the-loop; until the assembly exists, the human reviewer is the interim block.
- Holmes never authors the blueprint and never writes application code for the builder; it only observes, deduces, and supplies verifiable evidence.
- Holmes takes no autonomous action at handoff: it routes a fully traceable case file to the builder (Phase 6, "Resolution & Handoff — STRICT").
- The Sentinel Asymmetry: Holmes investigates corporate power (registries, court records); anti-doxxing refusals permanently block its tools from being aimed at private citizens. Holmes detects surveillance; it does not surveil.
- Non-destructive truth: facts are never silently deleted. Superseded facts are flagged invalidated and preserved.
-
Canon index + precedence:
docs/canon/README.md(verified primary source > component canon > system spec). -
docs/triad-canon.mdis a pointer (pointer-only per Martin, 2026-09-08): the one canonical copy lives in the WCJBT repo (docs/canon/triad-canon.md); on disagreement the WCJBT copy wins (the canon's own sync rule). -
Findings:
F-###indocs/audit/findings-ledger.md, format:F-### | Severity | Category | Location | Evidence (quote or ABSENT) | Why it matters | Recommended fix | Confidence (H/M/L) -
Severities: BLOCKER / MAJOR / MINOR / NIT (definitions in
docs/audit/00-audit-charter.md). -
Amendments:
A-##indocs/audit/amendments.md, each mapping to the finding(s) it resolves. -
Human decisions:
D-##indocs/audit/decisions.md. Never resolve a D-item yourself. -
IDs are stable for the life of the audit. Do not renumber.
Terse. No hyperbole, no praise, no filler. Findings and evidence only. Improve the spec, not morale.
- Substrate: goose (
aaif-goose/goose, Apache-2.0, Rust) over Zed's ACP (goose acp, JSON-RPC 2.0/stdio); all tools MCP, path-confined, deny-by-default. Secrets in the OS credential store[UNVERIFIED — confirm goose backend per platform]. - Delivery: Holmes ships as
holmes-core+holmes-guardcrates embedded in Alfred — no standalone UI/installer/updater (spec §4.1/§7 shell superseded; pending amendment A-03). - Models: Tier-1 cloud — Claude, Gemini 3.1 Pro, DeepSeek V4, Magistral, Qwen3.7-Max (proprietary, never sovereign). Tier-2 sovereign — non-Meta open weights (Qwen3.5-27B / Qwen3.6-35B-A3B, Magistral Small, Gemma).
deepseek-chat/deepseek-reasonerretire 2026-07-24 15:59 UTC → V4 Flash (deepseek-reasoner→ Flash, not Pro). DeepSeek Pro discount permanence[REPORTED]— budget both rates. - The Wall: Graphiti temporal graph, Neo4j/FalkorDB only (Kuzu abandoned); Graphiti's default LLM client is OpenAI — pin permitted clients, regression-tested (AC-DL-1 §6, Phase 2); invalidation-not-deletion.
- Derived files:
docs/architecture.md,docs/constitution.md,docs/build-roadmap.md,docs/security.md— re-seeded together with this block; canon markers preserved verbatim everywhere.