Measurable improvement for AI-operated systems.
An AI agent operating a system cannot improve what it cannot measure. loss-forge teaches agents to define loss functions for any project — operational loss that goes beyond test coverage to measure: are the end missions being accomplished? Is the agent getting more efficient? Is complexity growing or shrinking?
| Skill | What it does |
|---|---|
/loss-init |
Define 5-8 mission scores + 5-10 loss functions for a project. Writes the loss module, CLI command, and baseline snapshot. |
/loss-check |
Run before/after a change. Computes deltas. Returns verdict: IMPROVEMENT, REGRESSION, TRADEOFF, or LATERAL. |
/loss-audit |
Audit existing loss functions for coverage gaps, gameability, and missing agent metrics. |
Every project gets two composite scores:
- Loss (lower is better) — weighted average of everything that's broken: stale data, dead rules, unresolved findings, codebase complexity, agent token cost
- Mission (higher is better, max 1.0) — weighted average of end-mission completion: bills on time, transactions categorized, books balanced, net worth tracked
============================================================
REEVES-FINANCE INSTRUMENT PANEL | 2026-04-01
============================================================
LOSS: 6.21 (lower is better)
MISSION: 0.57 (higher is better, max 1.0)
============================================================
████████████████████ 100% M1_bills_on_time (w:10)
███░░░░░░░░░░░░░░░░░ 15% M2_cash_position (w:8)
████████████████████ 100% M3_categorization (w:6)
██████████░░░░░░░░░░ 50% M4_reconciliation (w:4)
███████████████░░░░░ 75% M5_tax_readiness (w:5)
░░░░░░░░░░░░░░░░░░░░ 0% M6_net_worth (w:4)
░░░░░░░░░░░░░░░░░░░░ 0% M7_cash_flow (w:3)
# Clone the forge
git clone https://github.com/eidos-agi/loss-forge ~/repos-eidos-agi/loss-forge
# In any project, tell Claude Code:
# "Use loss-forge to add loss functions to this project"
# or invoke directly:
# /loss-init"If loss goes up and missions don't improve, the change is wrong. That's the only rule."
MIT