Publish Readiness Rubric

publish-readiness-rubric

The rubric engine. Scores an artifact against a 15-row boolean-atomic matrix, applies criterion-importance tags from the Rubrics-as-Rewards pattern (arXiv 2507.17746), and emits a deterministic verdict.

Why this skill exists

Subjective publish-readiness assessments fail predictably. Two reviewers score the same artifact differently, a single reviewer scores the same artifact differently on different days, and the criteria drift over time as edge cases accumulate. The boolean-atomic rubric pattern solves all three: each row returns TRUE or FALSE, each row has a single objective check, and the criteria are version-controlled in this file.

The criterion-importance taxonomy (Essential / Important / Optional / Pitfall) is taken from "Rubrics as Rewards" (Viswanathan et al., arXiv 2507.17746). The hard-rules-vs-principles split is taken from "OpenRubrics" (arXiv 2510.07743). The atomic-criteria approach is the same one Anthropic, OpenAI HealthBench (openai/simple-evals), and Scale AI use internally for evaluation.

This skill is the engine. It is called by ingest at SCORE, by github-curator at STOP, and by the human directly when running an ad-hoc check.

When to use

Load this skill when the agent is asked to:

1. Evaluate any artifact (skill folder, markdown doc, packet zip) for publish-readiness.
2. Run the four-gate check (IP scrub, date consistency, no em dashes, PEL packet preparation).
3. Produce a Trinity Dialectic 3x3 report.
4. Apply criterion-importance tags to an existing checklist.

Do not load this skill for subjective review, for arbitrary code review, or to issue a PEL=1 verdict. The agent never self-issues PEL. The human runs PEL after the rubric returns PROMOTE.

The 15-row matrix

The rubric is fixed at 15 rows. New rows are added only when a real publish-time failure surfaces a gap (the evolution rule from pre_submit_gate v1.0). Speculative rows are forbidden.

Each row has four parts:

| # | Tag | Check | Expected |

Tag is one of: Essential, Important, Optional, Pitfall. Check is a single objective question with a TRUE/FALSE answer. Expected is the value that means PASS (TRUE for most rows; FALSE for Pitfall rows).

Scoring procedure

1. Load the artifact under review. If it is a directory, walk it recursively. If it is a zip, unpack to /tmp/score/<timestamp>/.
2. For each row, run the check. The check is mechanical:
   • Row 1: test -f $REPO/AGENTS.md returns 0.
   • Row 2: parse YAML frontmatter, validate against the spec.
   • Row 3: regex ^[a-z0-9][a-z0-9-]*$, length ≤64, equals dirname.
   • Row 4: char count of description string, 1 to 1024.
   • Row 5: LICENSE file exists, contains "Apache License" or "MIT License" or another SPDX identifier.
   • Row 6: grep for U+2014 and U+2013 across the artifact tree.
   • Row 7: run the IP scrub against tier-1 patterns (see references/ip-scrub-tiers.md).
   • Rows 8 to 15: see references/scoring-procedure.md for full mechanical checks.
3. Record TRUE/FALSE per row.
4. Aggregate.

Aggregation

ESSENTIAL_PASS = all Essential rows TRUE
PITFALL_PASS = all Pitfall rows FALSE
IMPORTANT_PASS = all Important rows TRUE
OPTIONAL_PASS = all Optional rows TRUE

VERDICT:
  if not (ESSENTIAL_PASS and PITFALL_PASS):
    QUARANTINE
  elif IMPORTANT_PASS:
    PROMOTE
  else:
    PROMOTE-WITH-WARNING

Output format

The rubric returns a deterministic markdown report:

PUBLISH_READINESS_RUBRIC v0.1.0
artifact: <path or identifier>
date: <ISO 8601>
scorer: publish-readiness-rubric

| # | Tag | Result | Reason if FAIL |
|---|---|---|---|
| 1 | Essential | PASS | |
| 2 | Essential | PASS | |
... (15 rows total)

Aggregate:
  Essential: N PASS / M FAIL
  Important: N PASS / M FAIL
  Optional: N PASS / M FAIL
  Pitfall: N FALSE (good) / M TRUE (bad)

VERDICT: PROMOTE / QUARANTINE / PROMOTE-WITH-WARNING

Repair list (if QUARANTINE):
  - row N: <one-line repair instruction>
  ...

The same input produces the same output. The rubric is reproducible.

The Trinity Dialectic 3x3 (escalation)

For artifacts the human tags "run Trinity" (Greenhouse submission, public paper publication, public repo release), this skill also runs the 9-cell matrix from system_directive_protocol Section 5. The 9 cells are:

Trinity is escalation, not default. The default consequential-action gate is the simpler 15-row matrix above. Trinity is reserved for explicit "run Trinity" tags.

The full Trinity output format is in references/trinity-3x3-output.md.

What this skill does NOT do

1. Does not issue PEL (Pathos-Ethos-Logos) verdicts. PEL is the human's role per pre_submit_gate v1.0 rule 4.
2. Does not score subjective quality. "Is this prompt good" is not in the matrix; "does this prompt's frontmatter validate" is.
3. Does not modify the artifact. The rubric is read-only on the artifact under review.
4. Does not invent rows. New rows enter the matrix only after a real failure documented in CHANGELOG.

Reference docs

Loaded on demand, not at activation:

• references/scoring-procedure.md the full mechanical check per row, including regex patterns and exact commands.
• references/ip-scrub-tiers.md the tier-1, tier-2, tier-3 IP scrub patterns from the system_directive_protocol governance set.
• references/trinity-3x3-output.md the 9-cell escalation format.
• references/criterion-importance-tags.md the Essential / Important / Optional / Pitfall taxonomy with examples and arXiv lineage.
• references/evolution-rule.md the procedure for adding a new row (real failure required, no speculation).

Provenance

This SKILL.md was authored 2026-04-26 by 0SxD. The boolean-atomic rubric pattern composes the Rubrics-as-Rewards importance taxonomy (arXiv 2507.17746), the OpenRubrics hard-rules-vs-principles split (arXiv 2510.07743), the HealthBench atomic-criterion approach (openai/simple-evals), the pre_submit_gate v1.0 four-rule pattern, and the Trinity Dialectic 3x3 from system_directive_protocol Section 5. All authored or composed by 0SxD; arXiv references are upstream.

END_PUBLISH_READINESS_RUBRIC_SKILL