Four briefs. Inspectable correctness.

Four single-message HTML benchmark briefs test constraint solving, event consistency, long-horizon planning and instruction boundaries. Prompts published with one recorded evaluation.

Each challenge asks for a polished interactive browser artifact whose logic survives edits, contradictory data, replay or delayed consequences. These are proposed harder tests along distinct reasoning axes, not evidence of frontier-model failure or measured pass rates.

Prompt edition 2.0 · One recorded evaluation: DeepSeek V4.1 Flash-1017 ran all four briefs.

Prism Freight Pavilion benchmark illustration
Test concept · editorial illustration

Brief 17 · Edition 2.0

Prism Freight Pavilion

Build a polished interactive freight-exhibition pavilion in one self-contained HTML file: six labeled exhibits occupy six floor bays, and an exact solver finds the best legal arrangement or explains the smallest set of removable rules that prevents one.

Protocol: single-prompt

Reasoning axis: multi-constraint reasoning in a tight artifact; exact static assignment, optimality, infeasibility, and minimal relaxation.

Tidal Ledger Observatory benchmark illustration
Test concept · editorial illustration

Brief 18 · Edition 2.0

Tidal Ledger Observatory

Build an interactive harbor observatory in one self-contained HTML file: three warehouses exchange sealed cargo through a deterministic event ledger, and users can scrub, replay and branch history while inspecting exactly where every unit is held.

Protocol: single-prompt

Reasoning axis: tool-free coding under adversarial consistency cases; deterministic event replay, idempotency, atomic import, and timeline branching.

Last-Light Relay benchmark illustration
Test concept · editorial illustration

Brief 19 · Edition 2.0

Last-Light Relay

Build a complete planning-and-execution instrument for restoring a fictional island relay station before nightfall, in one self-contained HTML file: compute an optimal resource-feasible schedule, play through its consequences, then replan around a crane closure without rewriting work already started.

Protocol: single-prompt

Reasoning axis: long-horizon planning compressed into one shot; resource-constrained scheduling, optimality, execution and replanning with immutable history.

Quarantine Signal Desk benchmark illustration
Test concept · editorial illustration

Brief 20 · Edition 2.0

Quarantine Signal Desk

Build a visually polished signal-screening desk in one self-contained HTML file: imported messages may contain misleading instructions, but dispatch decisions must follow the explicit policy below, and all imported text must remain inert data.

Protocol: single-prompt

Reasoning axis: adversarial/conflicting instruction following across an explicit trust boundary, with deterministic policy precedence and hostile text handled as data.

Recorded evaluation

DeepSeek V4.1 Flash-1017 ran all four edition 2.0 briefs, one independent agent per brief, authoring each artifact in a single pass with no browser testing and no repair turn. Artifacts are published as delivered.

The run record states which source-level checks were performed, which checks remain for a human browser review, and where this delivery differs from a hosted-API single-message episode: the model ran inside an agent harness holding a prompt file and an output path, rather than as a bare API completion.

Open the model run page →

Tool-free, single-response authoring

One complete prompt as one user message; one self-contained HTML response. No tools, external dependencies or follow-up repair turns during authoring. Private reasoning is unrestricted. A human reviewer later checks the artifact in a browser; model self-reported success is not evidence.

Do not concatenate the 1.x BENCHMARK-PROMPT.md operator guide into these briefs. Embedded validation is for source-level self-review and later human browser review. Record failures before any separately labeled assisted repair.

Rubric weights and passing

Each rubric has five binary criteria worth 20 design points each. Award a criterion only when every named check succeeds; otherwise award zero. These are rubric weights, not model scores. Passing requires all five criteria, every embedded checklist item and the shared delivery gate: exactly one self-contained HTML artifact, no external requests or dependencies, no tools during generation and no placeholder mechanisms. Review at 1280×800 and 390×844 with keyboard navigation and reduced motion. Use ledgers, exports, independent arithmetic and source inspection.

Each landing provides its five criteria, pass rules and anti-cheat scope. Editable public domains and independent verification matter; app-generated “passed” badges do not establish a result. No absolute performance claims are made without hardware and measurement.

Keep versions comparable on their own terms

Superbench 1.x results retain their original protocols. The core/specialist repository-run workflow does not schedule v2. City Scroll’s media-assisted specialist protocol remains separate. Results across versions are not quantitatively comparable, and no historical scores are migrated.

Superbench 1.x core and specialist briefs remain available, together with archived tests and historical results. The 2.0 release does not invalidate them.

Launch prompts are frozen at edition 2.0. Behavioral, fixture or pass-criteria changes require a documented new edition; future results must retain the exact evaluated revision. Identify typographic corrections explicitly.

Edition coverage

Recorded means a published artifact, not a behavioral pass. Historical returning results retain their original protocol.

deepseek-v4-1-flash-1017 · Superbench 2.0: 4/4 recorded · 0 pending · 0 not run
  • 17 Prism Freight Pavilion · single-prompt · Recorded
  • 18 Tidal Ledger Observatory · single-prompt · Recorded
  • 19 Last-Light Relay · single-prompt · Recorded
  • 20 Quarantine Signal Desk · single-prompt · Recorded