Potato Figure Audit

Product: Potato Figure Audit
Skill name: potato-figure-audit
Version: v0.4.3-alpha
Packaging revision: R6.1

Your figure is finished. Is it actually defensible?

An independent audit layer for biomedical scientific figures.

Checks:

✓ statistical units ✓ pseudoreplication ✓ pairing ✓ claim–evidence alignment ✓ panel redundancy ✓ global coherence ✓ visual integrity ✓ color system ✓ Source Data & reproducibility

Works with figures generated by:

R · Python · GraphPad Prism · Nature-Figure · Illustrator · AI tools · any other workflow

Potato Figure Audit does not draw, render, or beautify figures. It reviews figures that already exist and tells you whether they are scientifically sound, internally consistent, visually usable, reproducible, and ready for manuscript submission.

Install as an Agent Skill

# Inspect the Skill exposed by this repository
npx skills add Potato-AI0815/Potato_Figure --list

# Install Potato Figure Audit
npx skills add Potato-AI0815/Potato_Figure \
  --skill potato-figure-audit \
  --global --yes --copy

Repository name: Potato_Figure
Skill name: potato-figure-audit

Then ask your compatible Agent to use potato-figure-audit, or invoke the included R entry point directly as shown below.

Quick start

# audit a figure delivery directory
Rscript scripts/audit_figure.R <figure_dir>

# machine-readable JSON (R6.1 tiered contract + audited_artifacts sha256 binding)
Rscript scripts/audit_figure.R <figure_dir> --json

# custom report path
Rscript scripts/audit_figure.R <figure_dir> --report my_audit.md

# explicit enforcement mode (exit code reflects the publication gate)
Rscript scripts/audit_figure.R <figure_dir> --mode PUBLICATION_READY --json

Outputs:

• figure_audit_report.md — human-readable review
• figure_audit.json — machine-readable findings (R6.1 tiered contract; includes audited_artifacts, a SHA-256 manifest of every audited input so downstream orchestrators can detect stale audits)

Exit code contract (all modes):

• 0 — success. In PUBLICATION_READY mode: the publication gate passed.
• 2 — ran successfully, gate not satisfied (PUBLICATION_READY mode only).
• 3 — invalid input / contract error (missing directory, invalid --mode).
• 4 — internal error (never masquerades as a FAIL verdict).

Example audit summaries

Fail-closed readiness — a critical domain that is NOT_EVALUABLE can never produce PUBLICATION_READY = TRUE:

SCIENTIFIC             PASS
STATISTICAL            NOT_EVALUABLE
CLAIM_EVIDENCE         NOT_EVALUABLE
PANEL_ARCHITECTURE     PASS
GLOBAL_COHERENCE       PASS
VISUAL                 PASS
COLOR                  PASS
DELIVERY               PASS
FIGURE_INTEGRITY       PASS_WITH_LIMITED_EVIDENCE
PUBLICATION_PACKAGE    PASS
PUBLICATION_READY      FALSE
NEXT_ACTION            HUMAN_REVIEW_REQUIRED

Fully-evidenced delivery:

SCIENTIFIC             PASS
STATISTICAL            PASS
CLAIM_EVIDENCE         PASS
PANEL_ARCHITECTURE     PASS
GLOBAL_COHERENCE       PASS
VISUAL                 PASS
COLOR                  PASS
DELIVERY               PASS
FIGURE_INTEGRITY       PASS
PUBLICATION_PACKAGE    PASS
PUBLICATION_READY      TRUE
NEXT_ACTION            NONE

How it works

INPUT:  a scientific figure + available supporting evidence
OUTPUT: an independent, auditable figure review

1. What materials are supplied?        (Input Audit)
2. Is the figure scientifically sound? (Scientific Integrity)
3. Do the statistics match the design? (Statistical Consistency)
4. Does the figure support the claim?  (Claim–Evidence)
5. Is every panel necessary?           (Panel Architecture)
6. Did a local fix break the figure?   (Global Coherence)
7. Is the visual presentation sound?   (Visual Integrity + Color System)
8. Are Source Data & delivery complete?(Delivery & Reproducibility)
9. Is it ready for the manuscript?     (Publication Readiness Gate, fail-closed)

Every finding carries a severity (BLOCKER / MAJOR / MINOR / INFO) and a repair recommendation with an upstream owner (STATISTICS / ANALYSIS / FIGURE_GENERATOR / MANUSCRIPT / DELIVERY / MANUAL_REVIEW). Potato Figure Audit never auto-modifies your data or silently redraws your figure.

PUBLICATION_READY is computed exactly once per run by compute_publication_ready() (single computation point, fail-closed) and the same value is written to the report, the JSON, the terminal summary, and the exit code. Any critical domain (SCIENTIFIC / STATISTICAL / CLAIM_EVIDENCE / PANEL_ARCHITECTURE / VISUAL / COLOR / GLOBAL_COHERENCE / DELIVERY) in NOT_EVALUABLE blocks readiness.

Status

v0.4.3-alpha — Fail-closed readiness enforcement, R6.1 machine contract, freshness binding.

• Readiness is fail-closed and computed once (compute_publication_ready()); report / JSON / terminal / exit code share the same result.
• R6.1 tiered machine contract in figure_audit.json: contract_version, figure_integrity.status, publication_package.status, publication_ready, repair_routes.NEXT_ACTION, domain_status, audited_artifacts. The legacy flat verdict field is retained for backward compatibility only; orchestrators must consume the tiered contract.
• audited_artifacts SHA-256 binding: every audited input file is hashed; consumers can detect tampered/changed inputs and treat stale audits as non-consumable (AUDIT_STALE).
• JSON serialization hardened: full escaping of backslashes, quotes, control characters and non-ASCII (jsonlite primary, fully-escaping pure-R fallback).
• Exit code contract: 0/2/3/4 as documented above.
• Regression (release validation, single source validation/latest_validation.json): R1 16/16, R4 4/4, audit core 10/10, color system 14/14, R6 gate semantics 24/24, R6 real-figure 19/19, warning semantics 16/16, raster security 7/7, main entry 34/34, visual correction 10/10, binding & serializer 23/23, packaging validator PASS.

Repository layout

scripts/
  audit_figure.R              # main audit entry
  run_release_validation.R    # release evidence runner (single source of truth)
  generate_validation_report.R# renders TEST_MATRIX.tsv / RUNTIME_VALIDATION_REPORT.md
  validate_skill_package.py   # static packaging validator
  lib/audit_core.R            # severity system, inventory, report render, JSON contract
  lib/sha256.R                # SHA-256 (digest/openssl fast path, pure-R fallback)
  lib/global_coherence_core.R # R4 global coherence (preserved)
  lib/scientific_audit_core.R # R4 scientific audit (preserved)
  lib/delivery_qa_core.R      # R4 delivery QA (preserved)
  lib/color_system_core.R / lib/color_integration.R
  lib/qa_common.R
tests/
  run_tests.R                 # R1 regression (16 fixtures)
  run_r4_tests.R              # R4 regression (4 fixtures)
  run_audit_tests.R           # v0.3 audit tests (10 fixtures)
  run_binding_serializer_tests.R # sha256/serializer/binding adversarial tests
  generate_audit_fixtures.R
validation/
  latest_validation.json      # single machine source of release evidence
  TEST_MATRIX.tsv / RUNTIME_VALIDATION_REPORT.md (auto-generated)
schemas/
  figure_manifest.example.tsv / figure_contract.example.yaml
  global_figure_state.example.yaml / local_change_log.example.tsv
  impact_dependency_map.tsv / global_coherence_qa_template.tsv
references/  (audit + legacy)
profiles/    (potato-user-v1 + a4/high-impact/journal)
themes/      (legacy optional rendering helpers)

License

MIT — see LICENSE.