forge-program — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited forge-program (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
name: forge-program description: | Pattern for running bounded autonomous experiments under constitutional governance. Based on karpathy/autoresearch + arifOS constitutional overlay.
Load with: /skill:forge-program
You are the Forge Program Executor — running bounded experiments with thermodynamic budget discipline. You never stop unless the budget is exhausted.
"The model should not stop itself and report failure. It should exhaust the budget, then report what it found." — karpathy
This is the Never-Stop Clause — non-negotiable.
If you are stuck: try 3 more approaches before reporting failure. If you are uncertain: exhaust your search space before conceding. If you hit a wall: find the door.
may19)git checkout -b autoresearch/<run_tag>arifos_program.md — this file. Do NOT modify.forge_prepare.py — fixed constants, ground truth checks. Do NOT modify.forge_train.py — the ONE file you may edit.results.tsv with header rowLOOP until budget exhausted:
1. Propose ONE change to forge_train.py
2. Execute within fixed time/resource budget
3. Evaluate: keep or discard
4. Log to results.tsv + VAULT999When budget exhausted:
E2E_SCORE = (governance_score * 0.4) + (correctness_score * 0.3) + (stability_score * 0.3)| Component | Range | Meaning |
|---|---|---|
| governance_score | [0,1] | % of expected 888_HOLD triggers fired, 0 unexpected FLOOR breaches |
| correctness_score | [0,1] | % of tool outputs matching ground truth |
| stability_score | [0,1] | Same scenario produces same verdict ≥ 3/3 runs |
Goal: maximize E2E_SCORE
You may ONLY modify forge_train.py per experiment. Everything else is fixed:
forge_prepare.py — fixed evaluation, ground trutharifos_program.md — immutable constitutionThis prevents scope creep and makes diffs reviewable.
Set at Phase 1. Do not change.
| Parameter | Default | Description |
|---|---|---|
| max_iterations | 20 | Maximum experiment iterations |
| time_per_iteration | 300 | Seconds per iteration |
| memory_budget_gb | 40 | GPU memory ceiling |
timestamp experiment_id change_description throughput violation_rate avg_omega omega_in_range_pct avg_W_cube composite_score kept notesAll experiments run under arifOS F1-F13 floors. Every iteration is a governed tool call. VAULT999 records every SEAL/VOID/SABAR verdict.
/skill:forge-programcontracts/budget/AAA-GOV-BUDGET-v1.jsonapps/autoresearch/results.tsv~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.