diff --git a/.conductor/settings.toml b/.conductor/settings.toml
new file mode 100644
index 0000000..97e64ae
--- /dev/null
+++ b/.conductor/settings.toml
@@ -0,0 +1,5 @@
+"$schema" = "https://conductor.build/schemas/settings.repo.schema.json"
+
+[scripts]
+run_mode = "concurrent"
+setup = "bash .conductor/setup.sh"
diff --git a/.conductor/setup.sh b/.conductor/setup.sh
new file mode 100755
index 0000000..c273746
--- /dev/null
+++ b/.conductor/setup.sh
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+# Conductor workspace bootstrap for builder-guild.
+# Starts the shared local Neo4j instance, then creates a per-worktree venv and installs deps.
+# Environment-ready, NOT demo-seeded: does not run etl.py / demo_seed.py — seed explicitly when needed.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+# Shared Neo4j for local development. The compose file uses a fixed container name (cb-neo4j) and
+# fixed host ports (7474/7687), so Conductor worktrees share one local DB instance. -p is for
+# consistency, not isolation. --wait blocks until the compose healthcheck reports healthy.
+(
+  cd 01-context
+  docker compose -p builder-guild up -d --wait
+)
+
+# Per-worktree venv + neo4j driver + write/read smoke test against localhost:7687.
+bash 01-context/setup_a2.sh
+
+# Project dependencies for this worktree's venv.
+./01-context/.venv/bin/python -m pip install -q -r requirements.txt -r requirements-dev.txt
+
+echo "CONDUCTOR_SETUP_DONE"
diff --git a/.gitignore b/.gitignore
index f4a3290..9b77904 100644
--- a/.gitignore
+++ b/.gitignore
@@ -55,3 +55,8 @@ audits/
 # Beads issue tracker — local project management; issues.jsonl references git-ignored audits/ and
 # internal decisions, so keep the whole tracker out of the public repo.
 .beads/
+
+# Conductor: monorepo harness is symlinked into worktrees at setup, never committed (public repo).
+# Pattern has NO trailing slash so it also ignores a SYMLINK named .claude (git treats symlinks as files).
+.claude
+.conductor/settings.local.toml
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..7caac2d
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,69 @@
+# Builder Guild — Project Instructions
+
+## What this repo is
+Builder Guild is a graph-primary, bi-temporal, role-scoped knowledge base for AI agents, with a calibrated evaluation layer. The hard product boundary is:
+- `01-context/` = online context + retrieval + enforcement
+- `02-agents/` = agent consumer layer / trust boundary
+- `03-evals/` = offline evaluation and calibration
+
+Do not blur those layers. Online enforcement lives in `01-context`. Offline evaluation lives in `03-evals` and must not become a live decision signal.
+
+## Stage (do not overstate — see README status)
+- `01-context` runs end-to-end on the synthetic demo graph, with CI gating an import smoke plus the ingest/write engine and invariant sweeps on a real Neo4j.
+- `03-evals` has run its first real calibration, which correctly **refused** to certify autonomy.
+- `02-agents` is the design + contract + a demo consumer; fleet orchestration is **roadmap**.
+
+Describe no layer as more finished than that.
+
+## Core invariants
+- Never let an LLM write a fact.
+- Namespace on node and edge.
+- Bi-temporal by default.
+- No secrets in commits.
+- Current truth, role-scoped truth, and temporal truth are separate concepts. Do not collapse them.
+- Uncalibrated roles are suggest-only. Do not imply autonomy is leased unless `CALIBRATED[role]` is actually true in code.
+
+## Repo-specific rules
+- Facts enter through deterministic writes only (`MERGE` / `MATCH ... SET` style ETL).
+- LLM-generated content belongs only in the recall layer, not in entity properties or factual edges.
+- Any new node type or relation must carry `namespace` and preserve read isolation.
+- Functional edges supersede old truth; additive edges accumulate. Check `01-context/schema/relations.yaml` before changing write semantics.
+- If you touch retrieval, preserve honest role separation:
+  - keyword = exact IDs / names
+  - graph = structural truth
+  - vector = fuzzy recall
+- If you touch serving, preserve the abstain / execution boundary. Passing the gate is not enough when the role is uncalibrated.
+
+## What to read first by task
+- For overall product and status: `README.md`
+- For invariants and contribution constraints: `CONTRIBUTING.md`
+- For near/mid/later priorities: `docs/ROADMAP.md`
+- For schema and write semantics: `01-context/ONTOLOGY_SCHEMA.md`, `01-context/schema/relations.yaml`
+- For retrieval design: `01-context/HYBRID_RETRIEVAL_ARCHITECTURE.md`, `01-context/RETRIEVAL.md`, `01-context/PAGEINDEX_PILOT.md`
+- For agent trust boundary: `02-agents/AGENT_ARCHITECTURE.md`
+- For calibration status and why autonomy is still blocked: `03-evals/CASE_STUDY_calibration.md`, `03-evals/CONTEXT_EVALS.md`
+
+## Engineering expectations
+- Keep changes focused and explain why.
+- If you touch a write path, verify determinism and idempotency.
+- If you touch isolation, verify subject + edge + object scoping, not just one surface.
+- If you touch temporal behavior, verify as-of semantics explicitly.
+- If you touch retrieval, preserve namespace safety and disclose fallback behavior honestly.
+- If you touch calibration or autonomy, separate mechanism from grant. Code may revoke autonomy; it should not silently grant it.
+
+## Things to avoid
+- Do not treat roadmap items as shipped.
+- Do not call GraphRAG / Corrective RAG the default live path unless the code path actually proves it.
+- Do not add provenance / explainability systems to `01-context` casually; current priority is trust/calibration and temporal evidence.
+- Do not move logic from `03-evals` into `01-context` just because it improves scores offline.
+- Do not weaken namespace isolation for convenience.
+
+## Testing / verification mindset
+Before claiming a change is done:
+- run the narrowest relevant proof
+- verify the right layer changed
+- confirm no invariant was broken
+- if behavior is only documented, say that; if it was executed, say what command proved it
+
+## Keep this repo publishable
+Public and AGPL-licensed. No secrets (env vars only — see `.env.example`), no personal or operator notes, no cross-project context. Durable project guidance only.
diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
index b27f8ea..732fb1b 100644
--- a/docs/ROADMAP.md
+++ b/docs/ROADMAP.md
@@ -13,10 +13,14 @@ RAG-pattern coverage (see `01-context/RETRIEVAL.md` + `HYBRID_RETRIEVAL_ARCHITEC
 - **Multimodal RAG** — ✗ NOT built (text-only) → gap **G2** below.
 
 **Gate state: suggest-only.** `01-context/src/abstain.py` has `CALIBRATED=False` with provisional
-weights, so every decision routes to a human — autonomy is not leased. The last calibration run
-(the private spine, 2026-06-12 — *last-measured, not re-run since*) reported judge κ=1.0 but a
-FAILED coverage gate (sufficiency refit weight positive yet selective gain −3.0pp), so the gate
-correctly refuses to certify → trust track **G3** below.
+weights, so every decision routes to a human — autonomy is not leased. The calibration run recorded
+in `03-evals/CASE_STUDY_calibration.md` (N=10 human-validated golden, 6-namespace graph, external $0
+judge) measured a **+10.0pp selective gain** over confidence-alone (stable across three fits) yet
+**refused to certify** — because the sufficiency proxy refit with a **negative weight** (anti-correlated,
+−3.167 → −4.089: the selective gain came from confidence alone), the serving layer was only 40% correct
+on its own golden before the fixes (80% after), and **judge κ was unmeasurable** (8/10 items scored
+deterministically — N=10 is a smoke test). The gate correctly refuses on a broken sufficiency signal →
+trust track **G3** below.
 
 ## Near