# Canon Evaluation > Score any piece of Arcanea lore against a ten-dimension rubric with an adversarial refutation pass, before it is staged or promoted. Use when evaluating a new character, god, godbeast, creature, artifact, faction, place, era, magic mechanism, Library text, legend, or parable — and whenever asked whether lore is good, canon-consistent, on-voice, ready to stage, or ready for /lock-decision. Produces a scored verdict with per-dimension evidence, not an opinion. Invoked as phase 5 of the lore-release-gate. - Skill: `frankxai/canon-evaluation` (Agent Skill) - Install (CLI): `npx skillmds@latest add frankxai/canon-evaluation` - Raw SKILL.md: https://api.skillmd.com/api/skills/frankxai/canon-evaluation/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: frankxai (https://skillmd.com/u/frankxai) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/frankxai/canon-evaluation --- # Canon Evaluation The rubric that replaces "this looks good." Every dimension is scored 1–5 against stated anchors, every score carries a one-line piece of evidence, and the whole thing survives an adversarial pass that is actively trying to kill it. Two rules make this worth running: - **Evidence or no score.** A dimension scored without a quoted line, a file path, or a command output is unscored. Write `unscored — no evidence` rather than guessing; an honest gap is information, an invented 4 is noise. - **The adversary goes last and wins ties.** If the refutation pass lands a real hit, the artifact is revised. You do not get to out-argue it in the summary. ## The ten dimensions Score 1–5. Anchors are deliberately unsentimental: **3 is competent**, 5 is rare. | # | Dimension | 1 | 3 | 5 | |---|---|---|---|---| | 1 | **Canon fit** | Contradicts LOCKED canon | Consistent, sits beside existing systems | Consistent *and* makes an existing locked system make more sense | | 2 | **Mechanism** | Vibes; nothing works a particular way | Rules stated, mostly consistent | A reader can predict outcomes and be right; the rule constrains the author too | | 3 | **Cost** | Power with no price | A price is named | The price is paid by someone specific, on the page, and it hurts | | 4 | **Consequence** | Changes nothing | Changes one situation | Reframes something the reader already believed | | 5 | **Naming** | Register-wrong or collides | Register-correct, collision-checked | Register-correct, etymology carries meaning, sounds inevitable aloud | | 6 | **Voice** | AI-slop present; register blended | Clean, in-register | Prose a reader would quote from memory | | 7 | **Viewpoint** | Single omniscient assertion | One narrator, acknowledged bias | Three viewpoints that disagree, and the disagreement is the point | | 8 | **Mystery budget** | Everything explained, or nothing is | Withheld deliberately, logged | Withheld, logged, and the withholding actively generates story | | 9 | **IP distance** | Recognizable as franchise X, renamed | Architecture borrowed, bricks original | Synthesized across sources into something with its own logic | | 10 | **Story surface** | A closed fact | Implies one scene | Implies conflicts, factions, and questions others can build on | **Hard gates** — any one of these blocks staging regardless of composite: - Dimension 1 scores 1 → contradicts the vault. Not a revision; a rewrite. - Dimension 9 scores 1 → IP exposure. Rewrite or drop. - Dimension 5 scores 1 → renaming touches everything downstream. Fix before anything else. - `node .claude/ci/lore-lint.mjs ` exits non-zero → fix first, score after. **Composite bands** (sum of 10 dimensions, max 50): **If any dimension is `unscored`, the verdict is `incomplete — evidence required`.** Do not compute a composite, do not stage, do not promote. Otherwise "evidence or no score" becomes a way to buy a band: omit evidence for the one weak dimension, sum the other nine, and a 38/45 reads like a 38/50. An unscored dimension is a missing answer, not a zero and not an exemption. | Band | Verdict | |---|---| | 42–50 | Stage it. Nominate for `/lock-decision` if it closes a known gap. | | 34–41 | Stage as STAGING with the weak dimensions named in the STAGING LOG. | | 25–33 | Revise before staging. Name the two lowest dimensions and fix those. | | ≤24 | Not ready. Return to phase 2 of the gate and pick a different pattern. | ## The adversarial pass After scoring, argue the other side. Four lenses, each asked to *succeed* at killing the work, not to be fair to it: 1. **The canon lawyer** — find the contradiction. Not with the vault (the lint covers that), but with STAGING content, with the `book/` corpus, with something established three files away. Grep before concluding there is none. 2. **The IP auditor** — name the franchise this is closest to, and state the specific element. If the honest answer is "this is X's mechanic with new nouns," say so. Closeness to *one* source is the danger; synthesis across several is the defense. 3. **The bored reader** — why would anyone care? What question does this make them want answered? If the answer is "it's lore, it's just true," the artifact is inert and dimension 10 is a 1 regardless of what was scored. 4. **The continuity auditor from three years out** — what does this make impossible later? What does it force every future writer to work around? Cheap now, expensive forever is the failure mode worth catching early. A lens that lands a hit produces a revision, not a footnote. Record hits that were considered and deliberately accepted — with the reason — so the next reviewer does not re-litigate them. ## Output format ```markdown ### Canon Evaluation — **Composite: NN/50 — ** Lint: `lore-lint.mjs` exit 0 (N files) | # | Dimension | Score | Evidence | |---|---|---|---| | 1 | Canon fit | 4 | Uses the locked Starweave Gate (852 Hz) without redefining it — CANON_LOCKED.md:62 | | … | | | | **Adversarial pass** - Canon lawyer: - IP auditor: - Bored reader: - Continuity auditor: **Verdict**: — ``` ## Scoring the world, not the artifact For whole-system reviews — "how good is Arcanea's magic system now" — use the twelve-dimension world scorecard in `docs/worldbuilding/research/SYNTHESIS.md` §2 instead, which benchmarks against Tolkien, Elder Scrolls, FF, Marvel, HP, Warcraft, and the anime canon. That file lands in a follow-up PR; until it does, score per-artifact with the rubric below and defer the whole-system verdict rather than improvising a scorecard. That scorecard is the strategic instrument; this rubric is the per-artifact one. Re-score the world scorecard when a Ten Upgrade lands, and record the movement — an upgrade that does not move a dimension either was not needed or was not finished.