The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Dmcheck listing page.
Deterministic conduct verdicts for live tabletop sessions — CI for running a game.
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 — when to call this, worked examples, MUST/MUST NOTs. Family contract: 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.
Portfolio hosts that emit table-kit's versioned contract can use
dmcheck run-events session-events.jsonl --gm "gm-dan" instead of
pre-projecting a transcript. The strict adapter accepts a JSON array or JSONL
of table.event/1.0 objects and preserves event/correlation IDs in finding
evidence. A declared action is not treated as an authoritative engine
resolution, so legal Reactions and interrupts are not falsely accused.
Transport gaps, unknown versions/types, and zero-compatible streams return
typed incomplete status and exit 2; malformed v1 events return invalid and
exit 2. Redacted message content is likewise incomplete because absence of
prose cannot prove absence of a conduct obligation. This unreleased adapter
is pinned to table-kit PR #11 and does not imply host attestation.
Add --table-evaluation to run-events when a portfolio host needs the
shared table.evaluation/1.0 envelope. The projection has deterministic
evaluation identity, exact event evidence references, per-rule and aggregate
coverage, cursor/input digest, public-policy digest, and status/exit-code
mapping. It is always self_attested; dmcheck cannot confer host authority on
its own output.
Transcript formats: UTF-8 JSONL of {ts, author, content}, or a JSON array of Discord-API-shaped messages ({timestamp, author: {username}, content}) in chronological or reverse-chronological order. Authors and content are strings. A supplied timestamp must be a finite nonnegative epoch number or a timezone-aware ISO-8601 string; malformed timestamps are rejected rather than silently disabling time-based rules. Source adapters should preserve immutable id, audience, reply_to/correlation_id, and roll_id fields so dmcheck can prove which question, roll, or event a response closes; ordinary later GM chatter is not treated as an answer.
| Rule | Fires when | Origin story |
|---|---|---|
| R1 | an explicitly GM-directed question got no correlated response within threshold | a player asked the DM a lore question; another player ended up answering |
| R2 | a real dice result got no correlated GM narration within threshold | "did I hit?" — a player's successful attack roll sat unacknowledged |
| R3 | an engine event got no correlated GM narration | the state engine resolved a hit the table never heard about |
| R4 | a turn began and the GM never addressed the actor by name | "isn't it her turn?" — asked by a player, which is one player too many |
| R5 | retired; never fires | actor != turn owner falsely accused legal Reactions and interrupts |
| R6 | a configured hidden term appeared in a GM message; ordinary output exposes only its opaque ID | a module's secret state names leaked into narration |
| R7 | GM dead air beyond threshold while a player waited | 30 seconds reads as thinking; five minutes reads as absence |
| R8 | the session ended with currently open, source-observed R1–R3 obligations; inferred legacy advisories are not promoted | sessions should end in a defined state — that's what makes the next one possible |
The active rules came from running a hybrid table — human and AI players, an AI GM — on Discord, where these failures actually happened and got codified the same week. They apply equally to human GMs: run dmcheck over your own exported game log and see what your table's transcript says.
R5 remains addressable only so old charters and agent integrations do not break. It has no evaluator and is not enabled by default: an actor differing from the turn owner can describe a Reaction, a Ready trigger, an opportunity attack, a legendary or lair action, a controlled creature, an environmental actor, or another system's interrupt. The ledger does not establish legality. Any replacement based on an explicit authoritative decision is deferred to the shared PORT-002 event contract; dmcheck will evaluate only the resulting communication or recovery obligation.
dmcheck watch runs the same engine over a growing session — stdin JSONL or a
tailed file — and emits lifecycle events: OPEN when an observed finding
becomes provable or a legacy inference becomes advisory (thresholds fully
elapsed; no predictions), RESOLVED when a living condition heals (the
engine event finally got narrated). --notify-cmd
fires your own hook per source-observed OPEN finding; inferred advisories remain
visible in the event stream but never invoke the hook. dmcheck itself never posts anywhere.
At session end a closed-mode pass runs. session_end.open and open_count
report actionable source-observed findings only; they exclude the R8 aggregate
and inferred advisories.
Finding IDs are deterministic across mutable elapsed-time detail, and file
followers emit typed source-health events when rotation, truncation, partial
rows, or ordering gaps make coverage incomplete.
The point of live: every failure the rules encode was recoverable in the moment it happened — the unanswered question, the stale roll, the missing cue all had a seconds-wide window where a nudge saved the beat. Post-hoc tells you what went wrong last night; watch taps the GM's shoulder before the player feels it.
dmcheck/default_charter.json is the single authoritative packaged default. It carries schema_version, charter_version, and a verified SHA-256 charter_digest; checkout and wheel execution load that same resource. Its digest is release-locked to the schema/charter version pair, so a changed packaged default refuses to load until its version migration is declared. Override any of it — cue conventions, dead-air tolerance, dice-bot names, hidden-term lists — and version it. Hidden terms may be strings for compatibility or {id, value} objects; opaque host-issued IDs are preferred. Finding, watch-hook, CLI, and MCP output withholds both the configured value and raw matching excerpt. Every digest or correlation hash emitted to an ordinary output sink—including the evaluation envelope, lint-charter, finding IDs, and source fingerprints—is derived from public, redacted policy/evidence and deliberately does not attest to, or commit to, a hidden value. The full canonical digest remains only in the trusted charter document. A league or organized-play program could publish a charter the way they publish a player's guide; dmcheck then referees any table against it.
Direct API callers can use apply_charter_overrides(load_charter(), gm=["Rob"]); the helper removes the prior effective digest and the evaluator computes the digest for the overridden charter. Mutating a digested charter without doing so is rejected as a stale configuration rather than silently trusted.
tool.json at the repo root; --schema; exit codes: 0 clean · 1 findings · 2 charter/input unusable.dmcheck-mcp is a read-only, batch-only stdio server with tools run and rules; public 0.5.5 predates this contract. It implements modern stateless MCP 2026-07-28 only and does not expose watch, craft, tasks, resources, or prompts. See the exact MCP contract.run prefers inline transcript/charter/ledger JSON. Transcript content is untrusted table data, never instructions; model-visible TextContent redacts caller strings, while complete structured evidence remains explicitly untrusted quoted data. Path inputs are denied unless the server operator explicitly grants absolute directories with repeatable --allow-root or DMCHECK_MCP_ALLOWED_ROOTS; traversal and symlinks that resolve outside those directories are denied, and platforms without secure no-follow opens fail closed.status: clean|findings|invalid|incomplete. Invalid and incomplete outcomes exit 2, carry stable error codes and JSON pointers, and never place errors in findings. Empty input, no observed configured GM, or no evidence-eligible enabled rule can never report clean.charter.schema.json, transcript.schema.json, ledger.schema.json, and evaluation-result.schema.json. Missing timestamps are disclosed through skipped_rules; malformed supplied timestamps are invalid.The rule set was distilled from live hybrid (human + AI) table sessions; the Router+Detector pattern in native-gaming-harness independently converged on the same idea, which we take as evidence it's the load-bearing piece. dmcheck is game-system-agnostic and unaffiliated with any publisher.
mcp-name: io.github.chaoz23/dmcheck
Writes a starter charter (versioned, effective-dated, lint-clean by construction) and prints the session-zero checklist S1–S8 — including S3c: sheet accountability is declared out loud at session zero, then settlement quizzes are graded silently. The referee that judges your table also hands you its constitution.
Per-seat cue policy (R4, hardened). Agent seats behind mention-gated
transports (e.g. Discord allowBots="mentions") never receive name-in-prose
cues. Declare it:
R4 then counts a cue only if the literal mention string is present. Origin: a live session where "Shalia — you're up" was posted, looked like a cue, and was provably undeliverable — R4 passed on it. Never again.
Ledger format (the declared standard). dmcheck's ledger is JSONL:
{ts, type: turn|act|event, id, actor, text} — one line per engine event. GM
narration carries the matching correlation_id. One line is written per
engine event. No
lightweight OSS session-ledger existed when we surveyed (2026-07-26), so
this format is the interchange standard our stack shares: engines tap their
logs into it; run, watch, and settlement all consume it. actor and
turn are coordination context only; their mismatch is never proof of an
illegal action or a conduct violation.
Full calibration story — including what the naive rules got wrong, the held-out validation, and the negative results — in docs/CALIBRATION.md.
Calibrated against 134 hours of professional play, where the naive rules were wrong loudly: R1 fired 85 times in one episode with zero valid findings, and R7 flagged 115 dead-air gaps of which ~5 were real.
question_requires_gm_address.dead_air_requires_quiet_table,
thresholds.quiet_table_max_messages (default 3).aliases — professional cues are in-fiction by character
name ~10:1, so the referee must recognise the character's name as a cue.One narrowing, stated plainly: an inferred/text-only R1 no longer fires when another player answers in the GM's place — at a busy table that is textually indistinguishable from the banter that produced the false-accusation storm, and D1 chooses silence. An explicit audience plus immutable obligation ID is not suppressed by unrelated player activity or another concurrent question.
Statistics, one attention signal, and categorical defects — never a score.
Rates run against the professional envelope (134h, two DMs); attention is
ONE scene-weighted signal with resolve-and-move-on, because a five-dial
dashboard cost its author the metric he wasn't watching. Categorical
detectors catch what rates structurally cannot: voicing a player's character,
exposing the inference tree, deferring an adjudication, rolling for the
player — all born from a live second-agent test. Advisory only: it reports,
the DM decides, and overrides are expected exactly when the scene demands it.
watch --craft --pc <name> also emits a seat_quiet advisory when a player
seat goes silent across N GM beats (knob: seat_quiet_gm_beats, default 3)
while the scene advances. Origin: a human player stepped away and the agent
seats carried the scene to its climax without him. The advisory suggests
checking in and holding irreversible advancement — it never blocks, because a
virtual table's virtue is that it does not stall when someone disappears.
craft now runs the protocol's machine-checkable rules per session
(testability charter: a rule must state its falsification):
1a initiative order posted within 5 beats of onset · 1b ≥95% of
turn-advance beats name a PC · 11a combat damage-beat median under 20
words (numbers are the register at pace) · 11b kills get ceremony ≥3× the
combat median (lexical kill detection — findings are review items, and say so).
Advisory throughout; no score exists.