experiment-forensics

experiment-forensics is a Model training & evaluation skill for AI agents, published by wanshuiyin in the Anti-Autoresearch repository on GitHub. Skillspace rates experiment-forensics F (45/100) and the safety scan rated it risky; the rating is based on static analysis only.

Audit experiment integrity against the evidence ledger. At L2 (repo + result files present) a fresh cross-model reviewer reads the eval code line-by-line for fake/derived ground truth, score self-normalization, phantom results (a paper number with no backing file/key), dead/uncalled metric code, verified-scope inflation, method-described ≠ method-evaluated drift, synthesized-looking results, placeholder/fake data still wired into a released result, code-output ≠ reported-number mismatch, and missing reproducibility artifacts (an empirical/agent/LLM paper shipping neither code nor the prompts/configs its results need) — every finding span-anchored to a ledger claim_id. At L0/L1 (PDF / source only) the same patterns are surfaced as info-level 'could-not-verify' signals where the ledger gives an anchor (observability_level_required:2) — NEVER a fraud verdict from a PDF. The reviewer PROPOSES findings; tools/adjudicate_findings.py computes the verdict. Detect-only. Triggers: "experiment…

Grade
F
Quality
45 / 100
Safety
Risky
GitHub stars
103

Unproven: rated by static analysis only, not yet compared head-to-head.

Our two independent classifiers differ on this skill's category, so treat the placement as provisional.

This rating describes the version of the repository we crawled; it may have changed since.

View source on GitHubwanshuiyin/Anti-Autoresearch

Our review

Architecture
70 / 100
Maturity
Basic: a single SKILL.md
Security scan
High-risk patterns
License
MIT
Adoption
Adopted by 0 other repositories, in 0 collections

What the scanner flagged

  • executes shell/subprocess
  • over-broad allowed-tools ('*')
  • mentions secrets in text

Not checked yet

  • Not judged on the rubric yet (not a contender in its category)
  • No blind task eval yet
  • Scope (portable / vendor / repo-internal) not assessed; only the top of each category is read for scope
  • Sibling files (scripts, references) not fetched; security scan covers SKILL.md only

Better-rated skills in Model training & evaluation

Skillspace indexes and rates this skill; it does not host or endorse it. Run third-party skills at your own risk.