srdcheck

The shared rules rail for a table of agents — one cited, deterministic verdict every model calls, so a party of AIs running a game together stays in lockstep instead of quietly disagreeing.

Machine verdicts over the rules of the System Reference Document 5.2.1: cited, reproducible, delivered in milliseconds with zero tokens, and honest enough to refuse questions that aren't the rules' to answer. The rules lawyer for agents.

Status: v0.1 — young but real, building in the open. The kill tests that shaped the product — including the one that killed half our original idea — are in eval/RESULTS-phase0.md; the with/without-rails demo is in demo/mage-hand/; the truth scorecard below is generated by CI.

Why this exists

A game increasingly runs not as one model but as several agents sharing one evolving state — a GM, player agents, a rules referee, tool-callers. The moment more than one mind adjudicates, the binding problem stops being any single agent's knowledge or attention budget and becomes consistency: five capable models will each read the SRD and rule on grapple, concentration, or a death save a little differently, and the shared game state quietly desyncs. No amount of individual intelligence guarantees they agree with each other, and every re-derivation is another chance to drift. The fix is a deterministic rail they all call — the same atoms, the same citations, the same verdict every time — so every ruling carries one provenance any agent (or a human) can audit. That is where srdcheck stops being a convenience and becomes load-bearing.

Even at a table of one, the rail earns its place. A model running a game is a brilliant improviser with a finite attention budget: every mechanical micro-check it handles in-context — is this legal, is that slot spent, does the reaction refresh this round — spends tokens and attention that belong to the only work that needs a mind, the story. And the checks a model can answer, it cannot prove, cannot reproduce, and — as our own benchmark showed — will not refuse when the question is outside the rules' jurisdiction.

We tested this before building. Frontier models answered our SRD rules questions nearly perfectly — and confidently ruled on house rules, GM discretion, and content that doesn't exist in the SRD, where the only correct answer is "not my call." Small local models got 19–30% wrong with zero refusals. So srdcheck does not compete with what models know. It is a rail: state in, verdict out, citations attached, deterministically, every time — and the same rail for every agent at the table.

What it is

srdcheck answers one kind of question: is this legal under the rules? — and one better one: what is legal right now?

• Verdicts, not vibes. Exit code 0 = legal, 1 = illegal, 2 = cannot adjudicate. Every verdict carries its chain of SRD 5.2.1 citations. A rule we cannot cite is a rule we do not have.

• Judge, never simulate. No dice, no narration, no owned game state. State comes in with the query; a verdict goes out. The kernel is a stateless pure function — embeddable in anyone's DM product, VTT, or agent.

• Deterministic and fast. No LLM call anywhere in the verdict path. Runs local and offline.

• For agents first. MCP + CLI, --pipe, --schema, tool.json at the repo root. Humans get a plain-English why in the same payload.

• Rulesets are adapters. The kernel knows no game; all rule content loads from adapter packages, each carrying its own provenance manifest — source document, hash, license, attribution — that every verdict cites through. The SRD 5.2.1 adapter ships in this repo as the reference implementation. Anyone can build an adapter for another ruleset — a community, a private table, or a publisher shipping a first-party adapter for their own IP — and their content never passes through this project. The adapter catalog points; it never hosts.

See docs/product-truths.md for the invariants this project holds itself to, and docs/anatomy-of-a-turn.md for where srdcheck sits in a game-running agent's pipeline — a combat turn, a stealth infiltration, and the Mage Hand test, worked end to end.

Show your work. A big context window can read the SRD; srdcheck's value is the layer above — a consistent, cited set of rule atoms and a shared schema that several agents can rely on without each re-deriving the rules into context and drifting apart. docs/atom-concordance.md is the audit trail for that layer: every rule the engine applies, mapped to the exact SRD text it's grounded in (forward and inverse), generated and CI-checked-fresh, with a verbatim-citation oracle proving each quote really is on its cited page. A machine-readable atom-concordance.json is the slim index an agent can load to know what the rail covers and where it came from.

What srdcheck does not check yet

Honesty is the product (truth T8), so the boundaries are stated, not implied. As of v0.1 the SRD adapter covers combat-turn fundamentals; it does not yet check:

• Feature prerequisites. turn.plan judges the turn's action economy — that you spent at most one action, one bonus action, one reaction, one spell slot. It does not verify that a feature actually grants a given action. A lone two-weapon-fighting offhand attack is "action-economy legal" even though the 2024 rules require the Attack action first; that prerequisite lives in the character's features, which this version doesn't model. The success message says so explicitly.
• Per-spell effects. Damage/HP, death saves, and saving throws are now modeled (the reducer folds damage/heal/death-save; save.check and concentration.check resolve caller-rolled saves), but what a given spell does — Fireball's 8d6, Hold Person's paralysis on a failed save — is the long-tail swamp and stays refused. srdcheck computes the save; it does not roll the dice or apply the spell's bespoke effect.
• Conditions — all 15 are fully modeled across every surface srdcheck exposes: attack rolls/legality, action economy and Speed (incl. Exhaustion's graduated reduction), condition-aware saving throws (auto-fail Str/Dex, Restrained's Dex Disadvantage) and ability checks (Blinded/Deafened auto-fail sight/hearing, Poisoned/Frightened Disadvantage), damage typing (Petrified's Resistance to all), and condition immunity (Petrified vs Poisoned). A completeness oracle enforces in CI that every codified clause is cited-and-modeled — the only clauses left deferred are the three srdcheck never owns by doctrine (T6): positional geometry (Frightened's "can't approach"), the initiative order, and stealth/perception contests. Those return exit 2 by principle, not for lack of building.
• Content outside the SRD 5.2.1 — anything not in the SRD returns exit 2 from jurisdiction; as of 0.2.0 the name registries cover spells, creatures, conditions, classes, subclasses, species, and feats (census-anchored against the source text in CI).
• Misspelled or unknown query fields refuse loudly (0.2.0): every query validates its params against the declared schema — an unrecognized field returns exit 2 naming the offender, never a silently-defaulted verdict. A drift oracle in CI keeps handlers and schemas from ever diverging again.

Nonsensical inputs (negative Speed, a 99th-level spell, exhaustion past 6) return exit 2 rather than a confident-looking answer. When in doubt, srdcheck refuses — a wrong verdict is the only unforgivable bug.

Try it now

$ pip install git+https://github.com/chaoz23/srdcheck
$ python -m srdcheck jurisdiction "Fireball"          # exit 0 — known content
$ python -m srdcheck jurisdiction "Hexblade"          # exit 2 — not in the SRD, honestly refused
$ python -m srdcheck query mage-hand.use '{"kind": "attack"}'
{
  "verdict": "illegal",
  "exit_code": 1,
  "why": "The hand can't attack.",
  "citations": [{"section": "SRD 5.2.1 p.145 'Spells > Mage Hand'", "page": 145,
                 "quote": "The hand can't attack"}],
  "rule_ids": ["mage-hand.cant-attack"],
  "adapter": "srd-5.2.1@0.2.0"
}
$ python -m srdcheck cite "Command"                    # verbatim SRD text + page for any registered name
$ python -m srdcheck --schema                          # I/O contract for agents

Deterministic, offline, no tokens, sub-millisecond. The query surface is young and growing slice by slice — the architecture (kernel + adapters, spec at v0.9 RC) is the point.

For agents (MCP)

srdcheck is an MCP server with zero dependencies — stdlib only. After pip install, the command is srdcheck-mcp; from a clone it's:

{
  "mcpServers": {
    "srdcheck": {
      "command": "python3",
      "args": ["-m", "srdcheck.mcp"],
      "cwd": "/path/to/srdcheck"
    }
  }
}

Twenty tools: jurisdiction, turn_plan, turn_options, reaction_available, roll_compose, attack_modifiers (conditions, distance, and the ranged-in-close-combat Disadvantage), mage_hand_use, event_apply (the state reducer — folds a declared event, including damage/heal/death-save with damage typing, into a verdict plus a hash-stamped next state), save_check/check_make/concentration_check (resolve a caller-rolled d20 against a DC — condition-aware saving throws and ability checks, and the concentration-save DC from damage taken — srdcheck folds, never rolls), opportunity_attack_provoked (the provoke rule and its exceptions), grapple_initiate (Grapple/Shove DC + size legality), passive_perception (10 + modifier, and it refuses the non-SRD ±5), help_assist (the Help action's proficiency gate), creature_valid/creature_stats (cited SRD 5.2.1 creature validity + CR/XP), encounter_xp_budget (the p.202 party-level XP budget), and the toy adapter's ttt_move/ttt_options (which exist to prove the adapter spec). Every call returns the same verdict object as the CLI (verdict, exit_code, why, citations with source quotes) as structured content. An illegal verdict is a result, not an error; cannot-adjudicate is an honest refusal, not a failure. Tool descriptions and schemas come from the loaded adapters, so new adapters extend the tool list without kernel changes. See also tool.json for the CLI surface.

As a Python library

Consume a ruleset's content through a stable, versioned interface — no coupling to internal file paths:

from srdcheck import load_adapter, available_adapters

available_adapters()               # ['srd-5.1', 'srd-5.2.1', 'toy-tictactoe']
a = load_adapter("srd-5.2.1")      # a versioned identifier
a.categories()                     # the content categories this adapter carries
a.names("creature")                # every creature name (326, complete from the stat blocks)
a.record("creature", "Ghast")      # {"name": "Ghast", "cr": "2", "xp": 450, "citation": "SRD 5.2.1 p.287"}
a.query("encounter.xp-budget", {"level": 3, "difficulty": "moderate"})  # a verdict dict

Cross-version edition-trap detection — catch a name that was valid in an older edition but renamed/removed in the current one (the classic LLM failure, e.g. "Goblin" in 2014 → "Goblin Warrior" in 2024):

from srdcheck import edition_check
edition_check("Goblin", "creature")          # exit 1 — a srd-5.1 name, not in 5.2.1;
                                             #   data.candidates_in_current suggests Goblin Warrior/Minion
edition_check("Aboleth", "creature")         # exit 0 — valid in both
# or from the CLI: srdcheck edition-check "Goblin" --category creature

Adapter identifiers carry their version, so load_adapter("srd-5.2.1") and load_adapter("srd-5.1") coexist without a breaking change; edition_check is caller-parameterized (current / priors), so it generalizes to any versions and any category. The handle is content-neutral: it knows about categories and records, never about a specific ruleset's vocabulary.

The benchmark

bench/ is the rules-fidelity referee: versioned question sets with SRD-cited gold verdicts, a harness that scores any model or agent (gemini:, ollama:, or cmd:your-agent on stdin/stdout), and a generated scorecard that reports wrong-rate, refusal-rate, and false-confidence separately, per category, with no aggregate number — ever. A generated leaderboard ranks subjects by wrong-count alone (the one unforgivable failure), and any third party can submit a tamper-checked result — you can't grade your own homework, because the golds are the set's. Published findings: frontier models ace codified rules and fail by false confidence exactly where the rules end — and, on a per-horizon drift lane whose gold verdicts are derived by the engine itself, they don't drift even at 30-round horizons, while a local 8B model's errors are mistakes of rule-knowledge, not memory. Benchmark your own DM product with one command.

Truth scorecard

Every tagged release publishes a scorecard against the product truths — generated by CI, never hand-edited, no aggregate score.

Generated by scripts/truth_scorecard.py — regenerated and diff-checked in CI, never hand-edited. Statuses are honest: structural and held in review mean exactly that.

truth	claim	status	evidence
T1	wrong verdicts	enforced in CI	244 tests including gold suites ported from the Phase 0 eval; any wrong verdict fails the build
T2	no citation, no rule	enforced in CI	90/90 rule atoms carry verbatim source quotes; every-verdict-cites tests on all adjudicated paths
T3	advise, never overrule	structural	the API has no blocking or veto interface to wire; verdicts are advisory by construction
T4	one payload, two audiences	enforced in CI	every verdict carries machine fields plus a templated plain-English why; schema-tested
T5	enumeration is the product	proven in CI	consistency sweeps (50 turn states + toy boards) verify enumerate/validate agreement in both directions on every push
T6	judge, never simulate	enforced in CI	determinism test plus a purity lint: no randomness anywhere in the kernel, no network or subprocess in the verdict path
T7	mechanism never knows the game	enforced in CI	kernel lint scans every kernel module for game vocabulary (it caught a real violation during development)
T8	honest boundaries	enforced in CI	refusal goldens: unknown content, unmodeled conditions, and genuinely ambiguous rules text all return exit 2 with citations
T9	never a single number	enforced in CI	bench scorecard freshness test; per-category tables, no aggregate score exists anywhere in this repository — even the leaderboard ranks by wrong-count alone (T1) and shows the other failure modes unblended
T10	stranger-agent bootstrap	enforced in CI	cold-start conformance test reaches a first verdict from tool.json/--schema/MCP alone (22 tools); live probe: a frontier model given only tool.json produced a correct first verdict in 1 attempt(s), 3.9s (2026-07-16)
T11	table speed	enforced in CI	p95 latency budget test: 100 verdicts must stay under 100 ms at p95 (typically sub-millisecond)
T12	never sell what the model has	held in review	a strategy invariant: features pitched on knowledge parity are cut in review — enforced by humans and admitted as such
T13	the benchmark is a product	shipped	bench/ publishes 5 sets across 5 subjects with cited gold verdicts, incl. a per-horizon drift lane whose golds are engine-derived; a generated LEADERBOARD ranks any subject; cmd: driver + validate command let a third party submit a tamper-checked result
T14	every state has a lineage	enforced in CI	event.apply reducer stamps every transition (predecessor hash, causing event, rule ids, rule-vs-ruling kind); tests cover replay verification, tamper detection, the schema minimality ratchet, and reducer/validator agreement; demo replays 15 rounds hash-for-hash

Licensing

• Code: MIT.
• data/: includes material derived from the System Reference Document 5.2.1 under CC-BY-4.0 — see srdcheck/adapters/srd-5.2.1/sources/README.md for provenance and the required attribution.
• srdcheck is unofficial and is not affiliated with or endorsed by Wizards of the Coast.

vs alternatives

• Just asking the LLM — frontier models know these rules nearly perfectly (we measured it and killed half our own idea). What they can't do by construction: prove a verdict, reproduce it, or refuse when the question is outside the rules' jurisdiction — they claimed rules authority in the discretion zone in 5 of 6 demo runs. srdcheck sells proof, refusal, determinism, and economy — never knowledge.
• Lookup APIs and reference MCP servers (Open5e, 5e-srd-api and their wrappers) — they answer "what does the book say," not "is this legal given this state." Complementary, not competing; srdcheck cites the same text they serve.
• Closed engines inside AI-DM products — the strongest products in this vertical run deterministic rules layers, each rebuilt privately, uncited, unbenchmarked, and locked to their platform. srdcheck is that layer open, cited, benchmarked, and embeddable — including in theirs.
• Simulator libraries (combat engines, character libraries) — they run games and own state. srdcheck judges and owns nothing (state travels with the query, stamped with lineage), which is exactly what makes it embeddable anywhere.

Prior art

srdcheck stands on lessons from Temple of Elemental Evil / Temple+ (dispatcher architecture), PCGen (prerequisite predicates), the FoundryVTT PF2e system (rules as data), Datasworn (official rules-as-JSON precedent), and FIREBALL (structured play state). Patterns were studied; no code was taken from any of them.

mcp-name: io.github.chaoz23/srdcheck

Formula blanks (v0.3)

If a rule is a formula, the verdict is the formula — a number appears only when the caller supplied every blank. grapple.initiate without str_modifier + proficiency_bonus returns dc_formula and no dc; passive.perception without perception_modifier returns score_formula and no score. Origin: a live table where the base-8 rendered as "DC 8" when the true DC was 13. srdcheck stays lookup-pure: the formula is the rule text, filling caller-supplied blanks is arithmetic — deriving the blanks from a sheet is charactercheck's lane, never this one.

Spell & feature facts (v0.4)

srdcheck query spell.facts '{"name": "Protection from Evil and Good", "cast_at": 1000}'
srdcheck query feature.uses '{"feature": "divine-sense", "charisma_modifier": 4}'

Casting time, range, components, duration, concentration — the facts a DM misremembers at 2am — as cited verdicts, census-anchored over all 336 SRD spells (the build fails if any spell parses incompletely). cast_at adds expires_at as pure arithmetic on the caller's own clock: srdcheck holds no state and no clock. feature.uses covers the SRD's use-count formulas under the formula-blanks discipline. Origin: a 10-minute ward narrated as active for seventy minutes, caught by the human player.