GOVERNED SIMULATION SUPPLY

ScenarioForge

Turn an approved simulation demand into traceable scenarios, executable evidence, and a governed return receipt.

Keep the real seed, synthetic change, browser precheck, and authoritative run visibly separate—at every step.

▶ OPEN THE LIVE WORKBENCH  ·  中文文档  ·  Product brief  ·  Test evidence

REAL SEED  ＋  SYNTHETIC DELTA  →  2D PRECHECK  →  AUTHORITATIVE RUN

1 approved route SIMULATION only

6 bounded scenarios one governed family

6 / 6 prechecks deterministic browser runs

0 false assets fail closed on runner gaps

_{Actual product capture · 1440 × 900 · click the image to open the live Cloudflare Pages demo}

Public-demo boundary. The hosted build is an explicitly labelled deterministic Mock. It contains no API key, does not call MiniMax-M3, and never presents a browser precheck as authoritative PNC evidence.

The product decision

Simulation supply is not “generate a scene and call it done.” A data team must prove which demand was approved, where a seed came from, what was synthesized, what actually ran, which agent version was evaluated, and who accepted the result. ScenarioForge makes that chain inspectable before an asset can enter a regression set.

For	Decision supported	Product control
Simulation engineer	Is this job executable on the qualified runner?	Template bounds, preflight, six concrete plans, runner capability
PNC owner	Does this evidence apply to my exact model version?	agentRef, acceptance criteria, report lineage, human decision
Data platform owner	May this result be published and reused?	Demand/route gate, provenance, receipt, optional asset reference

Simulation supply line

01 · INTAKE Approved task data-supply-task/1.0 Human-approved SIMULATION route	➜	02 · PREFLIGHT Qualify dependencies Seed · label · model · budget Any unresolved gate blocks execution	➜	03 · PLAN Expand six scenarios Approved template + bounded axes Unique IDs and explicit provenance
				▼
06 · RETURN Issue governed receipt Result + limitations + optional asset No authoritative run, no assetRef	◀	05 · DECIDE Human acceptance gate ACCEPT or acknowledge the gap Decision is bound to the exact report	◀	04 · EXECUTE Precheck, then runner Browser 2D → MetaDrive Capability and execution states stay distinct

The implemented family is a right-lane closure with a left merge in a temporary work zone on an urban arterial. The demonstration binds DEM-WZ-001@1.0, TASK-SF-WZ-001, and nuScenes v1.0-mini scene-0061.

Evidence ladder

PROVENANCE IS ADDITIVE; CLAIMS ARE NOT INTERCHANGEABLE

01 REAL SEED Observed source context

＋

02 SYNTHETIC DELTA Declared scenario change

➜

03 2D PRECHECK Cheap deterministic filter

➜

04 AUTHORITATIVE RUN Qualified runner + exact agent

Layer	What ScenarioForge records	What it may support	What it must not claim
REAL SEED	nuScenes source, scene/sample identity, file digest and licence	Traceable visual context	That a lane closure or target-PNC event existed in the source
SYNTHETIC DELTA	Lane closure, cone spacing, target gap and bounded parameter source	A reproducible scenario specification	Ground truth observed by the collection vehicle
2D PRECHECK	Collision, TTC, lateral acceleration, merge completion and trace digest	Whether the simplified contract is worth sending to a runner	MetaDrive success, safety certification or a PNC pass
AUTHORITATIVE RUN	MetaDrive/adapter/map versions, agentRef, six run IDs, metrics and traces	Human review against the approved acceptance criteria	Any result when the adapter, report or agent binding is incomplete

The real open-data seed

nuScenes v1.0-mini · scene-0061 CAM_FRONT frame 14 1280 × 720 Sample token 2afb9d32310e4546a71cbe432911eca2 Integrity SHA-256 is recorded in DATA_LICENSE.md. Boundary The image is a seed—not temporary-lane-control ground truth and not target-PNC evidence.

Honest current state

Stage	Actual result	Interpretation
Browser 2D precheck	6 / 6 PRECHECK_PASSED	Four simplified assertions passed for each bounded instance.
Authoritative MetaDrive run	0 / 6 · CAPABILITY_GAP	MetaDrive is unavailable in the verified environment; no authoritative run occurred.
Published scenario asset	0 · NOT PUBLISHED	The product fails closed and returns a limitation-bearing receipt without assetRef.

This is intentional product behaviour—not a cosmetic fallback. A complete authoritative report fixture tests the state machine, but is explicitly excluded from real-run evidence.

AI, tools, and human authority

Actor	Allowed responsibility	Hard boundary
MiniMax-M3	Map an approved demand to a registered template, propose bounded parameters, cluster deterministic failure codes, draft reproduction steps	Cannot create map, trajectory, collision, TTC, coverage, runner-state, PASS/FAIL, or publication facts
Deterministic tools	Schema, map and physics checks; fixed-step 2D precheck; report validation; metric recomputation; capability probe	Browser 2D remains a precheck and cannot impersonate MetaDrive
Human owner	Approve the upstream route, own acceptance criteria, accept a complete authoritative report, acknowledge gaps	Cannot accept when agentRef, criteria, lineage, or runner qualification does not match

The public site never calls the model. The optional server-side MiniMax-M3 track accepts closed-set JSON only and fails closed on unknown fields, out-of-range values, timeouts, or request failures. ScenarioForge does not use visual-language output to create simulation truth.

Try the guided path

1. Open the live workbench and review the approved inbound task and dependencies.
2. Inspect the nuScenes REAL SEED, then compare it with the separately declared SYNTHETIC DELTA parameters.
3. Review the six concrete scenarios and their deterministic browser-precheck traces.
4. Continue to the runner stage: the verified public environment reports MetaDrive CAPABILITY_GAP for all six.
5. Acknowledge the gap and inspect the returned DataSupplyResult: zero authoritative executions and no published asset.

Contract surface

Boundary object	Fixed version	ScenarioForge responsibility
Inbound task	data-supply-task/1.0	Validate task, route, dependencies, and budget
Provider spec	simulation-supply-task/1.0	Bind seed, synthetic deltas, coverage axes, and evaluation target
Runner report	runner-execution-report/1.0	Verify MetaDrive identity, agentRef, and all six runs
Provider result	simulation-supply-result/1.0	Report planned, executed, and blocked totals plus limitations
Return receipt	data-supply-result/1.0	Return to SceneQL; publish no asset while a capability gap exists
Asset	simulation-scenario-asset/1.0	Bind seed/delta, runner, agent, run, and human decision

SceneQL can pass the public demonstration task through a URL fragment or JSON file. The fragment is not sent to the static server. An enterprise deployment should replace it with an authenticated, signed Task Reference.

Run and verify

Local deterministic showcase

Requires Node.js 20 or newer. The browser build has no runtime dependency.

npm install
npm run dev

Open http://127.0.0.1:4173. The default experience is explicitly labelled Deterministic Mock and automatically runs all six browser prechecks.

npm run check
npm run runner:probe
npm run preview

The current suite reports 36 passed / 0 failed and creates dist/. npm run runner:probe returning CAPABILITY_GAP is an auditable business terminal state: it means no authoritative MetaDrive run occurred on that machine.

Optional MiniMax-M3 server track

SCENARIOFORGE_ENV_FILE=/absolute/path/to/private/.env \
SCENARIOFORGE_AI_MODE=remote \
npm run dev

SCENARIOFORGE_ENV_FILE=/absolute/path/to/private/.env npm run test:llm

The server accepts the OPENAI_API_KEY, OPENAI_BASE_URL, and OPENAI_MODEL aliases, while the allowlist is fixed to MiniMax-M3 (case-insensitive). A live smoke test verifies endpoint connectivity, closed-set output, and tool-chain integration only; it does not establish visual or production robustness. The key never enters the browser bundle.

Cloudflare Pages

npm run build writes the static artifact to dist/ and marks the runtime as static. The hosted browser remains on the deterministic path, records the authoritative runner as disconnected, and makes no failing /api/* request. src/_headers supplies production security headers.

Setting	Value
Build command	npm run build
Build output directory	dist
Node.js	20 or newer
Runtime secrets	None

Do not configure OPENAI_*, SCENARIOFORGE_LLM_*, or private environment files in the public Pages project. Remote-model and authoritative-runner adapters belong behind an authenticated server boundary.

Capability boundaries

• The browser 2D engine is a product-loop prechecker—not MetaDrive, a production vehicle-dynamics model, or a safety-certification tool.
• PRECHECK_PASSED means four assertions passed under the simplified contract; it does not mean the target PNC passed.
• The in-house ScenarioSpec DSL 1.0 does not claim compatibility with ASAM OpenSCENARIO or CARLA.
• Without an agentRef, ScenarioForge can produce only reference-controller execution sanity, not an algorithm-regression claim.
• The nuScenes seed has neither temporary-lane-control ground truth nor target-PNC logs. Those conditions must be supplied through simulation, not inferred from the image.

---

Inspect the demand. Preserve provenance. Publish only earned evidence.

Live demo · Product solution · Verification report · Open-data notice

_{Source code and original public documentation are licensed under the MIT License.}