# ForgeAI loading-dock sample · v1.0.0

One shift as the supervisor of a cross-dock distribution centre, priced in dollars.
Released by ForgeAI Data Labs (https://forgeai.gg) under **CC BY 4.0**.

## What's inside

| File | What it is |
|---|---|
| `scenarios.json` | All 12 scenarios of the frozen `dc-v1` study bank: trailers, doors, crews, delays, no-shows and door outages, plus each scenario's reference costs. |
| `episodes/bank-3-harness.json` | Qwen 3.5 9B (`qwen/qwen3.5-9b`, reasoning off, temperature 0) planning through the ForgeAI hybrid harness. $4,950, score 75.1. |
| `episodes/bank-3-expert.json` | The scripted expert controller. $5,020, score 74.6. |
| `episodes/bank-3-naive.json` | First come, first served. $16,080, score 0. |
| `episodes/bank-3-alone.json` | The same 9B model choosing every move itself. $23,660, score 0. |
| `manifest.json` | Version, licence and a SHA-256 for every episode file. |

All four episodes run the same scenario (`bank-3`) and were recorded on 2026-10-31.

## Episode record (`forgeai.environment-episode.v1`)

Each episode holds the system prompt, every tick's observation text exactly as the agent saw it,
the agent's reply or plan, the actions it submitted, any actions the engine rejected (with the
reason), the cost change per tick, and for the harness run the options the per-step model
(JEV) chose between, with their probabilities. `actionLog` is the authoritative input.

## Scoring

Cost is the sum of detention, overtime, late outbound, short-ship, spoilage and end-of-shift
charges. Lower is better. The score puts the cost between two references on the same scenario:

    score = max(0, 100 × (naive − cost) ÷ (naive − oracle))

`naive` is first come, first served (0) and `oracle` is a hindsight plan search that sees the
whole shift in advance (100).

## Verify

The engine draws no random numbers and reads no clock. Replaying `actionLog` against the
scenario reproduces the recorded cost exactly. You can do it on the replay page
(https://forgeai.gg/environments/loading-dock/replay) or with the engine, on request.

## What this sample doesn't include

- The scenario generator, or any scenario outside the published `dc-v1` bank. Private banks
  and custom environments are available to research partners under a written data agreement.
- Human preference data or expert annotations.
- Any claim about performance outside this environment. A score here says how an agent ran
  this shift, nothing more.

## Cite

ForgeAI Data Labs, *Loading-dock sample v1.0.0* (2026), CC BY 4.0, https://forgeai.gg/environments/loading-dock/replay