forge-estimate

When this fires

Phase position: ESTIMATE — after forge-architect, before forge-provision. The output funds the cost-cap check that forge-provision makes against spec.constraints.budget_cap_usd.

What it does

1. Look up the recommended instance type for plan.target_params + plan.regime
2. Estimate GPU-hours using regime-specific heuristics (lora-sft → fewer hours per token than full-sft; qlora-sft → larger model fits but throughput is lower)
3. Multiply by the cached hourly rate from config/pricing.json
4. Add storage + transfer assumptions (small for the M2 size class)
5. Write manifest.estimate.* with a confidence band + assumptions[]
6. Advance state to BUDGET_GATE (legacy) — the v2 gate is PLAN_GATE

Inputs

• $1 = forge-id
• manifest.plan (base_model, regime, target_params)
• manifest.spec.constraints (budget_cap_usd, max_wall_clock_hours)
• config/pricing.json (ca-central-1 GPU instance rate snapshot)

Outputs

• manifest.estimate.{gpu_hours, instance_type, cost_per_hour_usd, total_cost_usd, wall_clock_hours, confidence, assumptions}
• manifest.state.current_phase = BUDGET_GATE

Idempotency

If manifest.estimate is already populated, exits 0.

Failure modes

External resources

• Local disk only — config/pricing.json lookup

Cost class

free — pure arithmetic.

Depends on

• forge-architect (populated manifest.plan)
• forge-shape is optional but refines the token estimate when run before this; without it the estimate uses a coarse corpus-size proxy.