@cairnkeep/cli 2.4.0 → 2.6.0

package/docs/learning/lessons/L00-why-cairnkeep.md

@@ -3,7 +3,7 @@
 **Status:** Ready
 **Track:** Quickstart
 **Time:** 10 minutes
-**Tested with:** Cairnkeep 2.4.0
+**Tested with:** Cairnkeep 2.6.0
 
 ## Outcome
 
@@ -18,22 +18,27 @@ entire transcript into the next prompt is expensive and preserves noise along
 with useful facts.
 
 Cairnkeep stores small, reviewed operational facts and retrieves them when they
-are relevant. It also provides a derived wiki and governed review workflows,
-without replacing source code, tests, issue trackers, or canonical documents.
+are relevant. It also provides a derived wiki and governed review workflows.
+Optional local evidence features can retain redacted session structure,
+artifacts, and deterministic hindsight notes, but they remain distinct from
+reviewed memory and are disabled by default. None of these layers replaces
+source code, tests, issue trackers, or canonical documents.
 
 ## Mental model
 
-Keep four layers distinct:
+Follow information through a trust ladder:
 
 | Layer | Example | Role |
 |---|---|---|
 | Harness history | The current conversation | Temporary working context |
+| Captured evidence (optional) | A redacted tool result or compaction summary | Bounded local evidence, not trusted instruction |
+| Hindsight note (optional) | A repeated failure signature and resolution | Deterministic derived knowledge, not automatically shared |
 | Cairnkeep memory | “Use transactional migrations” | Durable operational fact |
 | Project wiki | A cited architecture summary | Derived, reviewable knowledge |
 | Canonical source | Code, tests, ADR, issue | Authority when layers disagree |
 
-Document RAG is an optional fifth layer for retrieving larger documents. It is
-not required for Cairnkeep memory.
+Document RAG and token-miser context exploration are optional retrieval paths,
+not additional authorities. Neither is required for Cairnkeep memory.
 
 ## Exercise
 
@@ -43,6 +48,7 @@ For each statement, choose the appropriate destination:
 2. “Here is the complete API specification.”
 3. “The current debugging hypothesis is a race in the cache.”
 4. “Never run the production migration from a developer laptop.”
+5. “This synthetic stack trace appeared and was resolved twice.”
 
 Suggested answer:
 
@@ -51,6 +57,8 @@ Suggested answer:
 - 2 stays a canonical document and may optionally be indexed for RAG.
 - 3 remains session context until confirmed.
 - 4 is a durable constraint after it is verified against policy.
+- 5 may become local session evidence and a hindsight note. It becomes reviewed
+  memory only through a separate explicit workflow.
 
 ## Verify
 
@@ -59,6 +67,7 @@ You are ready for L01 if you can answer these questions:
 - Does Cairnkeep replace the repository or its documentation? **No.**
 - Does the default installation discover a remote memory server? **No.**
 - Should every sentence in a session become durable memory? **No.**
+- Does enabling one optional evidence feature enable the others? **No.**
 
 ## Common failures
 
@@ -66,17 +75,20 @@ You are ready for L01 if you can answer these questions:
 |---|---|
 | Memory is another transcript store | Store concise, durable facts rather than complete prompts |
 | RAG and memory are the same | RAG retrieves documents; memory preserves reviewed operational facts |
+| Captured evidence is trusted memory | Evidence and notes remain derived until explicitly reviewed or promoted |
 | Derived knowledge is authoritative | Code, tests, policies, and canonical documents win conflicts |
 
 ## Privacy and trust boundary
 
 The default stdio topology stores memory on the local machine. Remote storage,
-model-backed extraction, embeddings, and document RAG require explicit
-configuration and are taught separately.
+structured capture, note distillation/enrichment, artifacts, evaluation,
+model-backed extraction, embeddings, and document RAG require separate explicit
+configuration and are taught later.
 
 ## Recap
 
 - Memory is selective and durable.
+- Optional evidence and notes are distinct from reviewed memory.
 - Derived knowledge remains subordinate to canonical sources.
 - Optional services are not prerequisites.
 

package/docs/learning/lessons/L01-safe-trial.md

@@ -3,7 +3,7 @@
 **Status:** Ready
 **Track:** Quickstart
 **Time:** 15 minutes
-**Tested with:** Cairnkeep 2.4.0
+**Tested with:** Cairnkeep 2.6.0
 
 ## Outcome
 
@@ -26,7 +26,7 @@ ephemeral; the named course volume is persistent.
 1. Pull the public image:
 
    ```bash
-   podman pull ghcr.io/cairnkeep/cairnkeep:2.4.0
+   podman pull ghcr.io/cairnkeep/cairnkeep:2.6.0
    ```
 
 2. Start the stdio server with a read-only root filesystem and no Linux
@@ -38,7 +38,7 @@ ephemeral; the named course volume is persistent.
      --read-only --cap-drop=all --security-opt=no-new-privileges \
      --tmpfs=/tmp:rw,noexec,nosuid,size=64m,mode=1777 \
      --volume cairnkeep-course-data:/data:Z,U \
-     ghcr.io/cairnkeep/cairnkeep:2.4.0 stdio
+     ghcr.io/cairnkeep/cairnkeep:2.6.0 stdio
    ```
 
 3. The process waits quietly for MCP input. Press `Ctrl-C` after confirming it

package/docs/learning/lessons/L02-installation.md

@@ -3,7 +3,7 @@
 **Status:** Ready
 **Track:** Quickstart
 **Time:** 25 minutes
-**Tested with:** Cairnkeep 2.4.0 and Node.js 22 or newer
+**Tested with:** Cairnkeep 2.6.0 and Node.js 22 or newer
 
 ## Outcome
 
@@ -13,8 +13,9 @@ layer, and distinguish a complete installation from a partial one.
 ## Prerequisites
 
 - Node.js 22 or newer and npm.
-- Claude Code installed for the commands below. OpenCode users should follow
-  the equivalent setup in [the operating guide](../../operating.md).
+- Claude Code installed for the commands below. OpenCode, Kimi Code, and Qwen Code users
+  should follow the equivalent setup in
+  [the operating guide](../../operating.md).
 - Permission to install an npm package globally for your user.
 
 ## Mental model
@@ -40,7 +41,7 @@ machine-level steps.
 2. Install Cairnkeep:
 
    ```bash
-   npm install --global @cairnkeep/cli@2.4.0
+   npm install --global @cairnkeep/cli@2.6.0
    cairn --version
    ```
 
@@ -62,6 +63,15 @@ machine-level steps.
    cairn sync --check
    ```
 
+6. Optionally enable completion in the current Bash shell:
+
+   ```bash
+   source <(cairn completion bash)
+   ```
+
+   Use `cairn completion zsh` or `cairn completion fish` for those shells.
+   Completion is a convenience and is not another runtime component.
+
 ## Verify
 
 Run both checks:
@@ -88,7 +98,23 @@ installation.
 
 This registration starts `cairn memory-server` locally as a stdio child. It
 does not discover a server, send memory to a remote host, configure embeddings,
-or enable an optional document-RAG integration.
+enable session capture, or enable an optional document-RAG integration. Pi uses
+the separate `cairn sync-pi` trajectory adapter and still needs a user-selected
+bridge if Pi should expose MCP memory tools.
+
+Kimi Code can use the same memory server through its generated
+`.ai/start-kimi.sh` launcher, but its HTTP MCP configuration requires a literal
+URL rather than `${CAIRN_MEMORY_REMOTE_URL}`. Keep the resolved URL in a private
+`.kimi-code/mcp.json` and keep the bearer token in the environment. See
+[Harness compatibility](../../harness-compatibility.md#kimi-code).
+
+Qwen Code can use the generated `.ai/start-qwen.sh` launcher. Its project
+`.qwen/settings.json` supports environment references for both the remote URL
+and authorization header, so the file can contain names such as
+`${CAIRN_MEMORY_REMOTE_URL}` and `${CAIRN_MEMORY_HTTP_TOKEN}` instead of secret
+values. Keep this project settings file untracked and approve the generated MCP
+entry with `qwen mcp approve cairn-memory`. See
+[Harness compatibility](../../harness-compatibility.md#qwen-code).
 
 ## Clean up if you stop the evaluation
 

package/docs/learning/lessons/L03-first-project.md

@@ -3,7 +3,7 @@
 **Status:** Ready
 **Track:** Quickstart
 **Time:** 25 minutes
-**Tested with:** Cairnkeep 2.4.0
+**Tested with:** Cairnkeep 2.6.0
 
 ## Outcome
 
@@ -41,6 +41,10 @@ fact, and recall it from a new harness session.
    cairn doctor
    ```
 
+   Also leave trajectory, note, compaction, artifact, typed-memory, capability,
+   and evaluation flags unset. The quickstart proves core memory without
+   silently changing the amount of retained data.
+
 4. Launch Claude Code through the generated project launcher:
 
    ```bash
@@ -85,7 +89,8 @@ The lesson is complete only if:
 
 Use synthetic memory in the course. The local stdio server writes to local
 AgentFS/SQLite storage. Project wiki and planning files are derived artifacts;
-the repository remains canonical.
+the repository remains canonical. Optional session evidence is off in this
+lesson and is introduced in L13.
 
 ## Clean up
 

package/docs/learning/lessons/L04-memory-fundamentals.md

@@ -17,6 +17,8 @@ review extracted candidates before they become durable memory.
 - Exact read versus semantic search and substring fallback.
 - Superseding a fact without losing history.
 - Session-end candidate extraction and `/memory-review` as the acceptance gate.
+- Typed node metadata, tags, and hard filters as an independent opt-in.
+- Why hindsight notes and logical note address spaces are not reviewed memory.
 - Deleting synthetic course memory safely.
 
 ## Hands-on lab
@@ -31,6 +33,7 @@ candidate.
 - Search finds a fact without requiring its exact key.
 - Superseding preserves the prior value in history.
 - No extracted candidate becomes durable without an explicit review action.
+- Enabling typed nodes changes discovery structure, not the authority of a fact.
 
 ## Planned video
 
@@ -42,3 +45,4 @@ Wave 2 after the lab fixture is executable.
 
 - [Operating guide](../../operating.md#the-workflow)
 - [Storage and deployment](../../storage.md)
+- [L14 - Typed memory and controlled import](L14-typed-memory.md)

package/docs/learning/lessons/L05-daily-workflow.md

@@ -12,6 +12,8 @@ confusing derived knowledge with canonical project sources.
 ## Planned lesson
 
 - SessionStart recall, pre-edit recall, and SessionEnd candidate staging.
+- Optional structured capture, hindsight distillation, and compaction continuity
+  as separate post-session paths that do not bypass memory review.
 - `/wiki-ingest`, `/wiki-query`, and `/wiki-lint`.
 - Sparse, citation-heavy wiki pages rather than copied documentation.
 - Alignment gaps, contradictions, and project planning artifacts.
@@ -29,6 +31,7 @@ queue before exiting.
 - Every derived claim links to a canonical source.
 - A stale or contradictory page is detected rather than silently trusted.
 - Automatic hooks are observable and do not bypass human memory review.
+- Captured evidence and notes are never described as canonical project state.
 - The learner can describe what to do at session start, before editing, and at
   session end.
 
@@ -41,3 +44,4 @@ minutes and split wiki maintenance into a second video if rehearsal runs long.
 
 - [Operating guide](../../operating.md#the-workflow)
 - [Git provider configuration](../../git-providers.md)
+- [L13 - Session evidence and hindsight](L13-session-evidence.md)

package/docs/learning/lessons/L07-privacy-and-storage.md

@@ -14,6 +14,8 @@ will go before starting a session.
 - The placement rule: the server process owns the store it opens.
 - Local stdio versus explicitly configured remote HTTP.
 - AgentFS/SQLite data at rest and project-derived artifacts.
+- Separate local stores and retention for trajectories, notes, immutable
+  artifacts, capability callbacks, and evaluation experiments.
 - Credential boundaries and environment-variable precedence.
 - Embedding/extraction endpoints and what content they receive.
 - `cairn memory path` and a preflight data-flow checklist.
@@ -30,6 +32,7 @@ names, and prove that no remote destination was discovered.
 - Secrets are never printed, committed, or included in screenshots.
 - Optional model and RAG flows are identified separately from memory storage.
 - The learner can stop when an effective destination is ambiguous.
+- Independent feature flags are not treated as one blanket consent switch.
 
 ## Planned video
 

package/docs/learning/lessons/L10-context-exploration.md

@@ -6,12 +6,14 @@
 
 ## Outcome
 
-Use optional token-miser/FastContext exploration with compact citations and
-understand caching, routing, and fallback behavior.
+Use optional token-miser exploration with compact citations and understand
+caching, routing, and fallback behavior.
 
 ## Planned lesson
 
 - `context_explore` as a thin delegate rather than a second memory store.
+- token-miser as the public sibling that owns routing and context exploration;
+  Cairnkeep retains only the integration seam.
 - Repository root, query, citations, and cache keys.
 - Cross-referencing citations with memory and the wiki.
 - Optional pre-task invocation and its privacy implication.

package/docs/learning/lessons/L12-managed-overlays.md

@@ -17,6 +17,8 @@ through supported seams, delivers no secrets, and upgrades fleets safely.
 - Machine-private configuration versus distributable defaults.
 - Bootstrap, diagnostics, fleet status/update, release gates, and rollback.
 - Compatibility guarantees for existing clients and stored memory.
+- Explicit decisions for every new default-off evidence, capability, and
+  evaluation flag rather than enabling the entire advanced stack by policy.
 
 ## Hands-on lab
 

package/docs/learning/lessons/L13-session-evidence.md

@@ -0,0 +1,71 @@
+# L13 - Session evidence and hindsight
+
+**Status:** Brief
+**Track:** Practitioner; Evidence and Evaluation
+**Planned time:** 45 minutes
+
+## Outcome
+
+Enable one optional capture path in a disposable project, inspect what was
+retained, distil a local hindsight note, and remove the course evidence without
+confusing it with reviewed durable memory.
+
+## Planned lesson
+
+- The lifecycle: transient harness context, redacted trajectory, deterministic
+  hindsight note, explicit shared promotion, and reviewed memory.
+- Independent consent flags for trajectory capture, note distillation, note
+  enrichment, compaction capture, artifact tools, and artifact HTTP access.
+- Claude Code and OpenCode hooks versus Pi's native `session_shutdown` adapter.
+- Pre-write redaction, reasoning omission, size limits, 30-day defaults, and
+  fail-open capture behavior.
+- Compaction summaries as harness-produced continuity evidence, never trusted
+  instructions and never automatic fresh-session context.
+- `cairn trajectory`, `cairn notes`, `cairn artifact`, and `cairn doctor
+  --repair` inspection and recovery paths.
+
+## Hands-on lab
+
+Use a disposable repository and synthetic error. Enable only trajectory capture
+and note distillation in its `.ai/.env`; point `CAIRN_AGENTFS_BASE_DIR` at a
+course-only directory inside that disposable root. Start a supported harness
+through the project launcher, reproduce and resolve the error, then close the
+session.
+Inspect the trajectory, distil the exact session, search its failure signature,
+and preview pruning before removing the disposable project and isolated course
+note root.
+
+Core inspection commands:
+
+```bash
+cairn trajectory list --json
+cairn trajectory show SESSION-ID --json
+cairn notes distill --session SESSION-ID --json
+printf '%s\n' 'TypeError: course fixture' \
+  | cairn notes search-error --project "$PWD" --json
+cairn notes doctor --json
+cairn trajectory prune --dry-run --json
+cairn artifact list --json
+```
+
+## Acceptance criteria
+
+- Nothing is captured before explicit opt-in.
+- The retained trajectory contains no reasoning field or planted secret.
+- Distillation is local and deterministic when enrichment remains disabled.
+- The learner can distinguish a trajectory, note, artifact, and reviewed
+  memory by purpose, store, lifecycle, and trust level.
+- Cleanup targets only the disposable course evidence.
+
+## Planned video
+
+Draw the evidence lifecycle, capture one synthetic session, inspect the redacted
+record, and find the resulting failure note. Target 14 minutes. Treat compaction
+and Pi as short boundary chapters rather than requiring three harness demos.
+
+## Source material
+
+- [Operating guide: structured trajectories](../../operating.md#structured-session-trajectories-opt-in)
+- [Operating guide: hindsight notes](../../operating.md#hindsight-notes-opt-in)
+- [Privacy and data flow](../../privacy-and-data-flow.md#structured-trajectory-capture)
+- [Storage and deployment](../../storage.md)

package/docs/learning/lessons/L14-typed-memory.md

@@ -0,0 +1,51 @@
+# L14 - Typed memory and controlled import
+
+**Status:** Brief
+**Track:** Evidence and Evaluation
+**Planned time:** 40 minutes
+
+## Outcome
+
+Use typed metadata and hard filters, preview a bounded structured import, and
+explain how project-note and shared-note address spaces differ from filesystem
+paths and ordinary memory scopes.
+
+## Planned lesson
+
+- `CAIRN_TYPED_MEMORY_NODES` as a default-off server-start contract.
+- Core node types, canonical tags, exact-match ordering, and hard filters before
+  semantic ranking.
+- Logical `memory`, `project-notes`, and `shared-notes` address spaces.
+- Bounded schema-v1 `memory_import`, stable `import_id`, dry run, reject by
+  default, and explicit supersession with preserved history.
+- Journaled note-address transactions, replay, and `cairn doctor --repair`.
+- Why typed metadata adds discovery structure but does not make a fact true.
+
+## Hands-on lab
+
+Start an isolated local server with typed nodes enabled. Write synthetic nodes
+with two types and several normalized tags, compare unfiltered and hard-filtered
+searches, then dry-run a two-item import. Apply it once with reject semantics,
+replay the same `import_id`, and inspect history after one explicit superseding
+import.
+
+## Acceptance criteria
+
+- Restarting the server after changing the feature flag is part of the lab.
+- Type and tag filters exclude nonmatching nodes before ranking.
+- The dry run writes nothing, the replay is idempotent, and the default policy
+  cannot overwrite an existing key.
+- Imported values never appear in command diagnostics or course screenshots.
+- The learner does not present a logical note address as a host path.
+
+## Planned video
+
+Use a small type-and-tag diagram, then demonstrate filtered discovery and the
+dry-run/replay sequence. Target 12 minutes. Keep the full import limits in the
+written lesson rather than narrating a table of numbers.
+
+## Source material
+
+- [Operating guide: typed memory nodes](../../operating.md#typed-memory-nodes-and-note-address-spaces-opt-in)
+- [Storage and deployment](../../storage.md#typed-metadata-import-replay-and-history)
+- [Typed memory schema](../../../schemas/memory-node.schema.json)

package/docs/learning/lessons/L15-capability-governance.md

@@ -0,0 +1,56 @@
+# L15 - Capability governance
+
+**Status:** Brief
+**Track:** Evidence and Evaluation; Operator
+**Planned time:** 35 minutes
+
+## Outcome
+
+Inspect effective capability state, apply and reset a project override, and
+predict when a harness or memory-server restart is required.
+
+## Planned lesson
+
+- The default-off managed contract around eight existing capability owners.
+- Compatibility defaults, project overrides, process overrides, and precedence.
+- `list`, `status`, `enable`, `disable`, `reset`, and payload-free callback
+  logging.
+- Tool omission for disabled MCP capabilities versus invocation-time checks for
+  offline jobs and operating workflows.
+- Configuration digests as state identity, not approval or quality evidence.
+- Why enabling the master contract does not enable graph or note distillation.
+
+## Hands-on lab
+
+Use a disposable project with the master contract enabled. Save JSON status,
+disable `context.explore`, restart the memory server, verify effective state,
+then reset the override and restart again. Toggle callback logging separately
+and confirm that payload-free evidence still requires local trajectory capture.
+
+```bash
+CAIRN_CAPABILITY_CONTRACT=1 cairn capabilities list
+CAIRN_CAPABILITY_CONTRACT=1 cairn capabilities status --json
+CAIRN_CAPABILITY_CONTRACT=1 cairn capabilities disable context.explore
+CAIRN_CAPABILITY_CONTRACT=1 cairn capabilities reset context.explore
+CAIRN_CAPABILITY_CONTRACT=1 cairn capabilities logging enable
+```
+
+## Acceptance criteria
+
+- The learner records the before and after configuration digests.
+- A disabled MCP capability disappears only after a server restart.
+- `reset` restores inherited state; it is not described as `enable`.
+- No callback record contains prompts, tool arguments, results, or memory values.
+- Disabling an optional capability does not break standalone memory.
+
+## Planned video
+
+Show precedence as three stacked layers, then perform one disable/reset cycle.
+Target 11 minutes. Use JSON status for evidence and avoid presenting a digest as
+a security signature or performance result.
+
+## Source material
+
+- [Operating guide: managed capability contract](../../operating.md#managed-capability-contract-opt-in)
+- [Privacy and data flow](../../privacy-and-data-flow.md#capability-callback-flow)
+- [Capability contract schema](../../../schemas/capability-contract.schema.json)

package/docs/learning/lessons/L16-evaluation.md

@@ -0,0 +1,67 @@
+# L16 - Evaluation and ablation
+
+**Status:** Brief
+**Track:** Evidence and Evaluation; Operator
+**Planned time:** 50 minutes
+
+## Outcome
+
+Validate and run the bundled offline fixture, inspect a canonical two-pass
+report, and explain what a one-capability ablation can and cannot establish.
+
+## Planned lesson
+
+- The default-off coordinator and operator-owned adapter boundary.
+- Immutable task sets, paired seeds, fresh isolated workspaces, independent
+  verification, and deterministic serial schedules.
+- Run 1 versus fresh Run 2 with a task-contained immutable note snapshot.
+- An all-enabled baseline versus exactly one disabled capability.
+- Eligible pairs, missingness, compatible turn semantics, reported token/cost
+  fields, bootstrap intervals, and bounded reports.
+- `offline-framework` evidence versus separately validated live evidence.
+- Dry validation, invocation estimates, explicit `--yes`, cancellation, report
+  retention, dry-run pruning, and deletion.
+
+## Hands-on lab
+
+Build the server, enable evaluation only for the commands below, and use a
+temporary output directory with the committed fake task set and adapter.
+
+```bash
+npm run build:server
+output=$(mktemp -d)
+CAIRN_EVAL=1 bin/cairn eval validate \
+  --task-set examples/eval/task-set.json \
+  --adapter examples/eval/adapter.json --output "$output" --json
+CAIRN_EVAL=1 bin/cairn eval run \
+  --task-set examples/eval/task-set.json \
+  --adapter examples/eval/adapter.json --output "$output" --yes --json
+```
+
+Inspect the reported experiment ID with `cairn eval report`, then run one
+`memory.search` ablation only after reviewing the printed invocation estimate.
+Finish with a dry-run prune. The complete lesson will pin exact expected fields
+after the lab is rehearsed from the packaged CLI rather than a source clone.
+
+## Acceptance criteria
+
+- `validate` invokes no adapter and writes no experiment.
+- Every run uses isolated task workspaces and an independent verifier result.
+- Unknown or infrastructure outcomes are not counted as pass or fail.
+- The learner reports estimates, intervals, missingness, and evidence scope
+  without claiming causality, significance, or product improvement.
+- Course cleanup is limited to its temporary contained experiment root.
+
+## Planned video
+
+Start with the claim boundary, validate before running, inspect one two-pass
+report, and close by marking the fixture `offline-framework`. Target 16 minutes.
+Publish ablation as a second short video if the complete schedule obscures the
+measurement concepts.
+
+## Source material
+
+- [Operating guide: evaluation harness](../../operating.md#evaluation-harness-opt-in)
+- [Privacy and data flow](../../privacy-and-data-flow.md#evaluation-adapter-and-report-flow)
+- [Bundled offline evaluation fixture](../../../examples/eval/README.md)
+- [Evaluation task schema](../../../schemas/eval-task-set.schema.json)

package/docs/learning/lessons/L17-agent-boundary.md

@@ -0,0 +1,55 @@
+# L17 - The agent boundary
+
+**Status:** Brief
+**Track:** Evidence and Evaluation
+**Planned time:** 25 minutes
+
+## Outcome
+
+Separate Cairnkeep's shipped evidence and evaluation substrates from the
+design-only bounded meta-agent proposal, and identify which authority always
+remains with the operator and external harness.
+
+## Planned lesson
+
+- Current facts: typed capability state, local evidence, immutable evaluation
+  inputs, isolated schedules, independent verification, and canonical reports.
+- Design only: natural-language form compilation, candidate generation,
+  candidate injection, comparison, proposal bundles, patch application, and
+  rollback ledgers.
+- Cairnkeep coordinates evidence; it does not own a harness inference loop.
+- Natural language is untrusted proposal data, not executable authorization.
+- Separate human gates for form, evaluation, application, confirmation, and
+  rollback.
+- Why no CLI command, schema, runtime default, or quality claim ships for the
+  proposed meta-agent loop.
+
+## Hands-on lab
+
+Give learners a fictional automation proposal. Ask them to mark each operation
+as a current Cairnkeep capability, an operator/external-adapter responsibility,
+or future design-only work. Then identify the minimum approval, immutable input,
+path boundary, resource ceiling, evidence, and rollback information that would
+be required before any future implementation could act.
+
+## Acceptance criteria
+
+- No learner attempts to invoke a meta-agent command or treats an illustrative
+  block as a supported schema.
+- Every inference execution remains owned by an operator-configured adapter.
+- A capability digest is not treated as approval, identity proof, or evidence
+  of improvement.
+- Held-out confirmation data never becomes candidate-generation feedback.
+- The learner can name at least one explicit future prerequisite and one
+  unconditional human stop point.
+
+## Planned video
+
+Use one ownership diagram and one fictional proposal. Target 9 minutes. Put a
+persistent `DESIGN ONLY` label on every future component and end on the exact
+current-versus-future boundary.
+
+## Source material
+
+- [Meta-agent design contract](../../design/meta-agent.md)
+- [Evaluation and ablation](L16-evaluation.md)

package/docs/learning/templates/lesson-template.md

@@ -1,7 +1,7 @@
 # LXX - Lesson title
 
 **Status:** Draft
-**Track:** Quickstart | Practitioner | Operator
+**Track:** Quickstart | Practitioner | Evidence and Evaluation | Operator
 **Time:** NN minutes
 **Tested with:** Cairnkeep X.Y.Z
 
@@ -44,6 +44,9 @@ failed setup.
 State where data and credentials go, what leaves the machine, and what remains
 derived rather than canonical.
 
+For measurements, also state which claims the retained evidence can and cannot
+support.
+
 ## Clean up
 
 Provide reversible cleanup for every resource created by the lesson.

package/docs/learning/tracks/evidence-and-evaluation.md

@@ -0,0 +1,20 @@
+# Evidence and Evaluation Track
+
+**Audience:** maintainers and advanced practitioners evaluating optional layers
+**Prerequisite:** Practitioner track
+**Total time:** about 3.25 hours
+**Infrastructure:** local fixtures first; a live harness is operator-provided
+
+1. [L13 - Session evidence and hindsight](../lessons/L13-session-evidence.md)
+2. [L14 - Typed memory and controlled import](../lessons/L14-typed-memory.md)
+3. [L15 - Capability governance](../lessons/L15-capability-governance.md)
+4. [L16 - Evaluation and ablation](../lessons/L16-evaluation.md)
+5. [L17 - The agent boundary](../lessons/L17-agent-boundary.md)
+
+The track is complete when the learner can explain each consent boundary,
+inspect and manage every local evidence store, compare baseline and ablation
+reports without making unsupported claims, and distinguish current runtime
+features from the design-only meta-agent contract.
+
+Hands-on spine: `course-05-evidence` through `course-07-evaluation` in the
+public [course labs](https://github.com/cairnkeep/cairnkeep-course-labs).