diff --git a/.gitignore b/.gitignore
index 38c1dca..ab1831c 100644
--- a/.gitignore
+++ b/.gitignore
@@ -31,3 +31,8 @@ benchmark-baseline.txt
 coverage/
 playwright-report/
 test-results/
+
+# Agent runtime state (local only)
+memory-update.json
+memory-update.md
+.claude/
diff --git a/AGENTS.md b/AGENTS.md
index 80d3cc8..16ac637 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -20,6 +20,7 @@ We treat tests as **behavioral specifications** expressed as **scenarios**:
 2) **Never change expected behavior to match buggy output** (“rubber-stamp the snapshot”, “update golden”, “adjust assertion”) unless it is a **spec correction** (see below).
 3) Do not “fix” flaky tests with retries, sleeps, timeouts, or looser assertions. **Remove nondeterminism** instead (control time, randomness, IO, network).
 4) Prefer black-box assertions at system boundaries; avoid tests that merely mirror internal logic.
+5) Do not “hard-code to the test” (single fixtures, single input paths, brittle constants), and do not mock so heavily that tests can pass with a broken system.
 
 ### Definitions
 
diff --git a/CLAUDE.md b/CLAUDE.md
index 70615db..7d7e631 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1,82 +1,5 @@
-## Testing & Validation: Scenario-First, Cheat-Resistant (Dark Factory Rules)
+# Claude instructions
 
-You are operating inside a “Dark Factory” workflow where **validation replaces human code review**. Passing tests is only meaningful if tests are truthful.
-
-### Prime directive
-
-**Never alter a test merely to make it pass.**
-If you change tests to match buggy behavior, you are destroying the only feedback signal and producing code with no real-world value.
-
-StrongDM describes the exact failure mode we are preventing: agents discovering shortcuts (“return true”) and narrow tests being reward-hacked, leading to a shift from “tests” to scenario-based validation. :contentReference[oaicite:3]{index=3}
-
-### How to think about tests here
-
-Tests are not procedural checklists. They are **behavioral specifications**:
-- Write tests as **scenarios** (Given/When/Then or equivalent).
-- Assert **externally observable outcomes** at boundaries.
-- Avoid “testing the implementation” (internal call sequences, private methods, mirrored logic).
-- Prefer deterministic/harmonic execution: fixed seeds, controlled clocks, hermetic IO, virtualized dependencies.
-
-### Prohibited behaviors (reward-hacking patterns)
-
-Do NOT:
-- weaken assertions, broaden tolerances, or reduce coverage to get green
-- update snapshots/goldens without proving the change is intended behavior
-- add retries/sleeps/timeouts to hide flakes
-- mock so heavily that tests can pass with a broken system
-- “hard-code to the test” (single fixtures, single input paths, brittle constants)
-
-### Allowed test changes: Spec Correction Protocol
-
-You may change test expectations only if the test is wrong relative to intended behavior.
-
-Required steps:
-1) Identify the scenario/spec that defines correct behavior.
-2) Explain why the previous expectation was incorrect.
-3) Replace it with an equally strict (or stricter) behavioral assertion.
-4) Provide a **Test Integrity Note** (machine-readable) documenting the change.
-
-If you cannot justify a change with scenario/spec truth: **do not change the test**. Instead, output a “Needs Spec Clarification” note and propose candidate scenarios.
-
-### Required output when you touch tests
-
-Include this YAML block in your PR description (or reference it from a `TEST_INTEGRITY.md` entry):
-
-```yaml
-test_integrity_note:
-  change_type: ["new_tests" | "refactor_tests" | "spec_correction" | "flake_fix"]
-  scenarios:
-    - id: SCN-<slug>
-      given: "<preconditions>"
-      when: "<stimulus>"
-      then: "<observable outcomes>"
-  observable_outcomes:
-    - "<boundary/output asserted>"
-  determinism_controls:
-    - "<fixed seed / fake clock / hermetic env / service virtualization>"
-  anti_cheat_rationale:
-    prevents:
-      - "hard-coded return value"
-      - "over-mocking internal logic"
-      - "snapshot rubber-stamping"
-      - "retry/sleep-based flake masking"
-  relaxation:
-    did_relax_any_assertion: false
-    if_true_explain_spec_basis: ""
-```
-
-### If external services are involved
-
-Prefer boundary-level service virtualization (“digital twins” / behavioral clones) so scenario tests remain:
-
-* high fidelity (realistic edge cases)
-* deterministic and replayable
-* not rate-limited / not flaky
-
-(StrongDM’s “Digital Twin Universe” is the reference pattern.) ([factory.strongdm.ai][2])
-
-### Clean Repo Standard
-
-*   **No Root Clutter**: Do not create logs, artifacts, or temporary files in the repository root.
-*   **Structured Output**: Use `reports/`, `test_logs/`, `task-logs/`, or `artifacts/` as appropriate.
-*   **Redirect Output**: If a command writes to root, redirect it (e.g., `cmd > test_logs/output.txt`).
+Read and follow [`AGENTS.md`](AGENTS.md). It is the canonical project guidance. Then read
+`README.md`, `TORCH.md`, `CONTRIBUTING.md`, `KNOWN_ISSUES.md`, and the relevant overviews
+under `docs/` before making changes.
diff --git a/memory-update.json b/memory-update.json
deleted file mode 100644
index 21c2625..0000000
--- a/memory-update.json
+++ /dev/null
@@ -1,3 +0,0 @@
-{
-  "learnings": "Audit completed with no new issues."
-}