diff --git a/README.md b/README.md
index 382eaa2..8678d1a 100644
--- a/README.md
+++ b/README.md
@@ -4,6 +4,8 @@
 
 Feed it a session transcript (and optionally an engine event ledger) plus a **table charter**, and it returns named findings — the player whose question was never answered, the dice roll nobody acknowledged, the turn that began without anyone being told, the spoiler that leaked into the channel, the five-minute dead air. Every finding cites the charter rule it violates, with the evidence attached.
 
+**Agents start here → [SKILL.md](SKILL.md)** — when to call this, worked examples, MUST/MUST NOTs. Family contract: [FAMILY.md](https://github.com/chaoz23/srdcheck/blob/main/FAMILY.md).
+
 > **Cold-boot probe (2026-07-24):** a fresh agent session given only this repo URL installed and refereed a session in **2 commands**, verified all three exit-code legs against the docs, and confirmed the no-false-accusation contract held (a consumed roll produced silence). Its friction notes shipped as 0.1.1 (`--dice-bot`, `dmcheck charter`).
 >
 > **The design contract:** a false accusation is the unforgivable bug. Only source-observed evidence becomes a definite finding. Legacy ambiguity is either silent or explicitly labeled `severity: advisory`, `provenance: inferred`, `confidence: low`; it is never promoted by R8 or sent to a notification hook. The verdict path is model-free and deterministic: same transcript, same findings, every time.
diff --git a/SKILL.md b/SKILL.md
new file mode 100644
index 0000000..5fe2e7c
--- /dev/null
+++ b/SKILL.md
@@ -0,0 +1,139 @@
+---
+name: dmcheck
+version: 0.6.0
+description: >
+  Deterministic conduct verdicts for live tabletop sessions — CI for running
+  a game. Use it when you want to know how a session went as a *table*:
+  "did I leave anyone hanging?", "was that roll ever acknowledged?", "did
+  the fight end without anyone being told?", "why did the table go quiet
+  for five minutes?", post-session retros, or live monitoring while a
+  session runs. Findings cite the table charter; ambiguity produces
+  silence, never accusations.
+---
+
+# dmcheck — conduct referee for live tables
+
+Feed it a session transcript (JSONL) plus a table charter; it returns named
+findings — the unanswered question, the unacknowledged roll, the turn that
+began with no cue, the event nobody narrated, the dead air. Model-free and
+deterministic: same transcript, same findings, every time.
+
+## Three things to remember
+
+1. **A false accusation is the unforgivable bug.** A rule fires only when
+   the transcript *provably* shows the violation. No findings does not mean
+   a perfect session — it means nothing was provable. Never editorialize
+   beyond the findings.
+2. **Every finding cites the charter rule and carries the evidence.** Relay
+   both; the evidence excerpt is what makes a finding fair.
+3. **Name the GM.** The packaged default charter names no GM author — pass
+   `--gm "<name>"` (or a charter with `gm` set) or you get exit 2, not
+   findings.
+
+## Exit codes ARE the verdict
+
+`0` = clean (silence) · `1` = findings · `2` = unusable input/charter (the
+honest lane — fix the input, don't retry blind). Failures print JSON on
+stderr, never a traceback.
+
+## Invocation
+
+```bash
+dmcheck run session.jsonl --gm "Greta"                      # quick: default charter + GM override
+dmcheck run session.jsonl --charter our-table.json --ledger events.jsonl
+dmcheck run-events events.jsonl --gm "Greta"                # table.event/1.0 streams (fail closed)
+dmcheck watch session.jsonl --follow --gm "Greta"           # live, during the session
+dmcheck rules                                               # the rule set with definitions (R5 retired)
+dmcheck init [--dm]                                         # starter charter (+ DM-CORE.md defaults)
+dmcheck lint-charter our-table.json
+```
+
+Transcript lines: `{"author": "...", "content": "...", "ts": ...}`. Ledger
+(optional): engine events to check narration against. MCP server included.
+
+## Worked example
+
+```bash
+dmcheck run session.jsonl --charter our-table.json --ledger events.jsonl
+```
+
+```json
+{
+ "result_schema_version": "1.0",
+ "status": "findings",
+ "exit_code": 1,
+ "messages": 9,
+ "findings": [
+  {
+   "finding_id": "r2-74e366c6f414e0dbb5cd5ffa",
+   "rule": "R2",
+   "summary": "unconsumed-roll: a dice result got no correlated GM narration within threshold",
+   "charter": "roll_ack_within_messages=4; correlation=explicit",
+   "detail": "possible dice result from DiceBot has no provably correlated GM narration within 4 messages; text-only or correlation evidence is incomplete",
+   "evidence": {"index": 6, "ts": 1700000700.0, "author": "DiceBot",
+                "content": "Bram rolls 1d20+4: [18] = 22"},
+   "severity": "advisory",
+   "provenance": "inferred",
+   "confidence": "low",
+   "charter_version": "1.0",
+   "charter_digest": "sha256:9882b2f4..."
+  }
+ ],
+ "counts": {"R2": 1}
+}
+```
+
+Exit 1. **Read `severity` and `provenance` before relaying:** only
+`severity: "finding"` + `provenance: "observed"` is a provable violation;
+`advisory`/`inferred`/`confidence: "low"` is a lead, not an accusation. A
+clean session prints `"findings": []`, `"status": "clean"`, exit 0 —
+silence is the pass verdict, by design.
+
+## MUST / MUST NOT
+
+- MUST pass `--gm` (and `--dice-bot` if a bot rolls) when using the default
+  charter; author names are how every rule tells who's who.
+- MUST include the evidence excerpt when relaying a finding to a human.
+- MUST treat silence as "nothing provable", never as "nothing wrong" — and
+  findings as "provable per charter", never as a character judgment.
+- MUST NOT paraphrase findings into accusations the tool didn't make, or
+  aggregate them into a score — there is deliberately no composite score.
+- MUST NOT present an `advisory`/`inferred` finding as a violation; the
+  no-false-accusation contract lives in those two fields.
+- MUST NOT judge the narrative quality of rulings; conduct is process, not
+  art.
+- MUST NOT invoke a model to re-check the transcript; the verdict path is
+  code, and that's the point.
+
+## Validation checkpoints (self-audit before relaying)
+
+GM/dice authors named · charter version noted · exit code read · evidence
+attached to every relayed finding · silence reported as silence.
+
+## Cross-skill workflows (check family)
+
+- **Session retro:** table-kit's session JSONL is dmcheck-compatible — run
+  dmcheck over it after close-out; findings (or silence) go to the humans,
+  who decide any process change.
+- **Live table:** `dmcheck watch --follow` alongside table-kit during play;
+  `--notify-cmd` fires per open finding.
+- **Rules vs conduct:** a ruling's *legality* is srdcheck's lane; whether
+  the table was *told about it properly* is this tool's lane.
+
+Family contract: [FAMILY.md](https://github.com/chaoz23/srdcheck/blob/main/FAMILY.md).
+
+## Changelog / stale-knowledge deltas
+
+- **0.6.0:** result envelope is `result_schema_version: "1.0"` with
+  `status` (clean|findings|invalid|incomplete); findings carry
+  `finding_id`, `severity` (finding|advisory), `provenance`
+  (observed|inferred), `confidence`, and a sha256 `charter_digest`
+  (public-policy scope — hidden spoiler values withheld). **R5 retired**
+  (actor≠turn-owner is legal: reactions, readied actions, legendary
+  actions). `run-events` ingests table.event/1.0 streams, fail closed.
+  Invalid/incomplete input is a structured exit-2 envelope, never a
+  traceback. If you remember flat findings or 0.5.x output, that's stale.
+- **0.5.5:** `init --dm` fixed for installed users (packaging omitted
+  dm_core.md through 0.5.1–0.5.4).
+
+Findings are advisory: the table's humans own every judgment about people.