@@ -10,6 +10,7 @@ Included:
 - `planner/`
 - `decision/`
 - `modules/`
+- `modules/control_plane/`
 - `evolution/`
 - `self_model/`
 - `integrations/arc_agi3/`
@@ -22,6 +23,7 @@ Included:
 - `scripts/run_local_machine.py`
 - `scripts/run_webarena.py`
 - `scripts/eval_metrics_panel.py`
+- `scripts/cognitive_loop_ablation.py`
 - `scripts/check_conos_repo_layout.py`
 - `scripts/check_runtime_preflight.py`
 - `tests/test_public_repo_smoke.py`
@@ -54,6 +56,24 @@ The runtime marks execution isolation as best-effort. Execution tickets carry a
 credential/secret-lease signals, but this package does not claim OS-level
 filesystem or network sandboxing.
 
+Action authority note:
+
+Action governance lives in `modules/control_plane/action_governance.py`.
+Tool calls are not automatically trusted: mirror writes require evidence refs,
+code source sync requires passing validation, source sync requires an approved
+plan, and repeated action failures can downgrade an agent's write/exec
+authority.
+
+Failure learning note:
+
+Failure learning lives in `core/runtime/failure_learning.py` and the
+`failure_learning_objects` SQLite table. Failed runs can now persist a formal
+failure mode, violated assumption, evidence refs, missing tool or bad policy,
+suggested regression test, suggested governance rule, and a future retrieval
+object. The local-machine runner keeps these objects in the object layer and
+exposes them as JSON, rather than expanding raw failure records into prompt
+text.
+
 Minimal setup:
 
 ```bash
@@ -80,10 +100,12 @@ conos mirror exec --mirror-root runtime/mirrors/session-1 --backend docker --all
 conos mirror plan --mirror-root runtime/mirrors/session-1
 conos mirror apply --mirror-root runtime/mirrors/session-1 --plan-id <plan_id> --approved-by machine
 conos mirror rollback --mirror-root runtime/mirrors/session-1 --plan-id <plan_id>
+conos llm control-plane --route structured_answer --required-capability structured_output --permission generate_text
 conos ui runtime reports audit
 conos eval runtime reports audit
 conos dashboard runtime reports audit
 python scripts/build_macos_app.py --output-dir dist
+python scripts/cognitive_loop_ablation.py --task-count 25 --seed 17
 python scripts/local_mirror.py init --mirror-root runtime/mirrors/session-1
 python scripts/run_arc_agi3.py
 python scripts/run_local_machine.py --instruction "inspect README" --candidate README.md
 
@@ -54,10 +54,54 @@ conos ui runtime reports audit
 conos ui runtime reports audit --output runtime/ui/dashboard.html
 conos eval runtime reports audit --output runtime/evals/eval_metrics_panel.json
 conos dashboard runtime reports audit
+conos llm control-plane --route structured_answer --required-capability structured_output --permission generate_text
 conos preflight --strict-dev
 conos layout
 ```
 
+## Model-agnostic agent control plane
+
+Con OS treats models and agents as governed execution resources, not as one
+hard-coded coding assistant. Profile-backed LLM route policies can be combined
+with a generic agent registry for coding agents, browsers, CI runners, VM
+backends, or other tool executors:
+
+```bash
+conos llm profile \
+  --base-url http://127.0.0.1:11434 \
+  --route-policy-output runtime/models/llm_route_policies.json
+conos llm control-plane \
+  --route-policy-file runtime/models/llm_route_policies.json \
+  --agent-registry runtime/models/agent_registry.json \
+  --task-type code_review \
+  --route coding \
+  --required-capability coding \
+  --permission propose_patch \
+  --risk-level medium
+```
+
+The control decision records selected agent id, capability match, permission
+gate, approval-required permissions, blocked candidates, and an audit event id.
+
+## Action capability governance
+
+Con OS separates tool availability from action authority. `modules/control_plane/action_governance.py`
+checks whether the selected agent may read, propose a patch, write the mirror,
+run validation, or sync back to source. The local-machine adapter now enforces
+that mirror writes have evidence references first, code syncs have passing
+validation first, source syncs carry an approved plan, and repeatedly failing
+agents are downgraded away from write/exec authority.
+
+## Failure learning ledger
+
+Failures are stored as structured runtime objects, not chat memory. `core/runtime/failure_learning.py`
+normalizes failed actions into failure mode, violated assumption, evidence refs,
+missing tool or bad policy, suggested regression test, suggested governance
+rule, and a future retrieval object. These objects are persisted in SQLite by
+`RuntimeStateStore` and are injected into later unified context as object-layer
+evidence; the local-machine runner exposes them as JSON artifacts instead of
+dumping raw failure objects into the prompt.
+
 ## Desktop app
 
 The local desktop app is the non-terminal product shell for this alpha. It
@@ -166,6 +210,25 @@ The panel reports:
 - `recovery_rate`: recovery-attempted runs that resolved or later succeeded.
 - `verifier_coverage`: success, completion, or verification-required runs covered by verifier authority or verifier-result evidence.
 
+## Cognitive loop anti-cheat benchmark
+
+The controlled cognitive-loop benchmark is a negative-control suite, not a
+success demo. It compares the full loop against component ablations and
+anti-cheat controls:
+
+```bash
+python scripts/cognitive_loop_ablation.py --task-count 25 --seed 17
+```
+
+Core controls include `NoPosteriorUpdate`, `NoDiscriminatingExperiment`,
+`NoHypothesisCompetition`, `NoSemanticStrictness`, `RandomProbe`, and
+`BaselineLLM`. Anti-cheat controls include `NoMemory`, `NoEvidence`, `NoStep9`,
+`NoStep10`, `ShuffledEvidence`, `WrongBinding`, and `FreshBaseline`.
+
+The report fails unless `Full` clears the hard thresholds for success margin,
+wrong commits, false rejections, posterior accuracy, semantic mismatch, evidence
+binding integrity, and formal commit completeness.
+
 ## What this repo contains
 
 ### Stable-enough public surface