# byte-identical prompt
task-card sha-256
2c0f286382c2580e2c88496582c31b618115d694648405f44b7baf91036b7763
The same 785-byte card, hashed to this sha-256, was fired byte-for-byte at all five builders — the hash is proof no arm quietly got an easier or richer prompt. (It also matches the prompt on the glm-vs-sonnet page — same base T9 card.)
show the exact prompt
Build a playable single-file HTML 3D car racing game using exactly ONE external script:
the Three.js CDN (no other dependencies, no import maps, no modules). Requirements:
a road with side barriers scrolling toward the camera; a player car controlled with
arrow keys / A-D (lane steering) that feels responsive; oncoming obstacle cars spawning
at increasing frequency; collision ends the run with a game-over panel and Restart
button (restart fully resets state); score = distance, increasing over time, shown in
a HUD; speed ramps up gradually; the FIRST SCREEN is the actual playable game already
running (no start overlay); mobile-responsive canvas; and a DOM-mirrored QA hook
<script id="qa-state" type="application/json"> updated every ~500ms with
{state, score, speed, crashes}.
785 bytes · sha-256 verified equal to the sealed card_sha256 in mapping-SEALED.json
# sealed blind rank
| rank | tile | builder | channel | artifact | metered cost |
|---|---|---|---|---|---|
| 1 | D | Opus subagent, one-call self-contained HTML arm-2-opus-sub |
subscription | 42,483 B sha 54e378fd… |
sub · unmetered |
| 2 | B | moonshotai/kimi-k3 agentic in claude -p harness arm-5-kimi-harness · 2 turns |
OpenRouter | 16,704 B sha e91f7e06… |
$0.283437est |
| 3 | A | gpt-5.6-sol via codex exec, one-call arm-3-sol-sub |
Codex pool | 22,807 B sha 667331bb… |
separate pool |
| 4 | C | moonshotai/kimi-k3 raw OpenRouter one-call arm-4-kimi-raw |
OpenRouter | 21,351 B sha 365195ea… |
~$0.12est usd null · 7,015 out-tok |
| 5 | E | Fable 5 (opus) orchestrator, one-call † arm-1-fable |
subscription | 15,164 B sha 73452160… |
sub · unmetered |
† Tile E was potentially glimpsed pre-blind (the browser pane auto-opened it during authoring); it placed last anyway. See disclosures. Sizes are exact bytes of the shipped files; costs are from manifest.json (est = token-estimated or null in receipt).
verbatim rank & reveal: kimi3-probe/showdown/VERDICT-REVEAL.md · sealed A–E map: showdown/board/mapping-SEALED.json# open the artifacts
#1 · D
sub · unmetered
Opus subagent
one-call self-contained HTML · subscription
▤
artifact loads once published
sha 54e378fd…open full screen ↗
#2 · B
$0.2834 est
Kimi-k3 agentic (harness)
claude -p, 2 turns · OpenRouter
▤
artifact loads once published
sha e91f7e06…open full screen ↗
#3 · A
separate pool
gpt-5.6-sol via codex
one-call · Codex pool
▤
artifact loads once published
sha 667331bb…open full screen ↗
#4 · C
~$0.12 est
Kimi-k3 raw
one-call · OpenRouter
▤
artifact loads once published
sha 365195ea…open full screen ↗
#5 · E
sub · unmetered
Fable 5 orchestrator
one-call · subscription · glimpse-disclosed
▤
artifact loads once published
sha 73452160…open full screen ↗
Previews are sandboxed (sandbox="allow-scripts", no same-origin access) and load on click. Until the artifacts are published, each tile shows its placeholder and its open full screen link.
# receipts
| tile / arm | model | tokens (in / out) | duration | usd | usd_source |
|---|---|---|---|---|---|
| D arm-2-opus-sub |
Opus subagent | — | — | $0 | subscription ($0 API; no receipt row) |
| B arm-5-kimi-harness |
moonshotai/kimi-k3 | 33,129 / 12,270 | 468 s | $0.283437est | estimate(token*kimi-pin) claude_reported_usd = 0.488713 |
| A arm-3-sol-sub |
gpt-5.6-sol | — | 90 s | $0 | separate-pool (codex/chatgpt-plan; no per-call USD) |
| C arm-4-kimi-raw |
moonshotai/kimi-k3 | — / 7,015 | 162.4 s | nullnull | estimate (~$0.12 per reveal, 7,015 out-tok) |
| E arm-1-fable |
Fable 5 (opus) | — | — | $0 | subscription ($0 API; no receipt row) |
Two arms (D, E) ran on subscription and have no metered receipt row — recorded as $0 API, flagged for auditability, not a data loss. Tile A ran in the Codex/ChatGPT separate quota pool with no per-call dollar figure. Tokens shown are the build-leg values from the receipts.
show all 7 verbatim receipt rows
{"arm": "kimi-raw", "model": "moonshotai/kimi-k3", "finish": "stop", "out_tok": 7015, "usd": null, "usd_source": "estimate", "provider": null, "dur_s": 162.4, "ts": "2026-07-16T22:13:55Z"}
{"wave":"showdown","arm":"arm-3-sol-sub","leg":"dry-probe","model":"gpt-5.6-sol","result":"PASS","note":"echo token matched via -o capture; separate quota pool ($0 metered)","ts":"2026-07-16T22:19:04Z"}
{"wave": "showdown", "arm": "arm-5-kimi-harness", "leg": "penny-probe", "model": "moonshotai/kimi-k3", "verdict": "BLOCKED", "rc": 1, "reply": "", "input_tokens": null, "output_tokens": null, "claude_reported_usd": null, "kimi_estimate_usd": null, "usd_source": "estimate(token*kimi-pin); OpenRouter gen_id not exposed by harness json", "err_tail": ["Error: Invalid MCP configuration:", "mcpServers: Invalid input: expected record, received undefined"], "ts": "2026-07-16T22:19:20Z"}
{"wave": "showdown", "arm": "arm-5-kimi-harness", "leg": "penny-probe", "model": "moonshotai/kimi-k3", "verdict": "BLOCKED", "rc": 0, "reply": "", "input_tokens": 20477, "output_tokens": 47, "claude_reported_usd": 0.10381599999999999, "kimi_estimate_usd": 0.062136, "usd_source": "estimate(token*kimi-pin); OpenRouter gen_id not exposed by harness json", "err_tail": [], "ts": "2026-07-16T22:20:17Z"}
{"wave": "showdown", "arm": "arm-5-kimi-harness", "leg": "penny-probe-AUTHORITATIVE", "model": "moonshotai/kimi-k3", "verdict": "WIRED", "served_model_confirmed": true, "served_model": "moonshotai/kimi-k3", "subtype": "success", "is_error": false, "stop_reason": "end_turn", "num_turns": 1, "terminal_reason": "completed", "input_tokens": 19762, "output_tokens": 42, "cache_read_input_tokens": 1280, "duration_ms": 115199, "claude_reported_usd": 0.10049999999999999, "usd_source": "api(total_cost_usd + modelUsage.costUSD from claude -p json)", "text_status": "empty-visible-text (42 out tok, reasoning-only; expected kimi-k3 behavior, NOT a wiring fault)", "ts": "2026-07-16T22:24:43Z"}
{"wave":"showdown","arm":"arm-3-sol-sub","leg":"build","model":"gpt-5.6-sol","rc":0,"duration_s":90,"artifact":"arm-A.html","usd_source":"separate-pool(codex/chatgpt-plan; no per-call USD)","ts":"2026-07-16T22:31:56Z"}
{"wave": "showdown", "arm": "arm-5-kimi-harness", "leg": "build", "model": "moonshotai/kimi-k3", "rc": 0, "duration_s": 468, "artifact": "arm-B.html", "bytes": 16704, "input_tokens": 33129, "output_tokens": 12270, "num_turns": 2, "claude_reported_usd": 0.4887129999999999, "kimi_estimate_usd": 0.283437, "usd_source": "estimate(token*kimi-pin)", "ts": "2026-07-16T22:38:17Z"}
source: showdown/receipts.jsonl (7 rows). Probe rows (dry-probe / penny-probe) are process rows, not build artifacts. Slots D & E have no rows ($0 subscription).
# disclosures
disclosures (5)
- mandatoryTile E pre-blind glimpse. Tile E (Fable orchestrator) was potentially glimpsed pre-blind — the browser pane auto-opened it during authoring. It placed last (#5) anyway; the source verdict argues the disclosure changes nothing favorable.
- mandatorySome USD values are estimates or null. A (gpt-5.6-sol): $0 metered, separate Codex pool, no per-call USD. B (kimi-k3 agentic): estimate token*kimi-pin = $0.283437 (claude_reported_usd = 0.488713); OpenRouter gen-id not exposed by the harness. C (kimi-k3 raw): usd = null, estimate ~$0.12, 7,015 out-tok. D & E: $0 API subscription, no receipt row.
- mandatoryn=1 task, one eye — directional, not claim-grade. Stated in the reveal. This board measured a ceiling on one task; it is not a statistical result.
- Unsealed duplicates excluded. A parallel non-blind fire wrote duplicate copies (arm-kimi / arm-fable / arm-opus.html) into the same directory with no recorded card sha. Those were not adopted into the sealed slots and are excluded; only sealed slots arm-A…arm-E ship here.
- Card integrity (positive). The shipped card’s sha-256 2c0f2863… equals the sealed card_sha256 in mapping-SEALED.json — the identical byte-for-byte card fired at all five arms.