Lava lamp v1 · OpenCode · run c5946c9b · 11.11.2026 06:18 Tallinn

Qwen3.8 27B

Extra high effort

Works

7 of 8 checks passed. All 5 blocker checks passed.

What it built: 3 captures

Desktop capture of OpenCode (Qwen3.8 27B)'s output on Lava lamp v1, xhigh effort, 11.11.2026 06:18 (run c5946c9b)
Desktop · 358.9 KB · 2992d4882933…
Phone capture of OpenCode (Qwen3.8 27B)'s output on Lava lamp v1, xhigh effort, 11.11.2026 06:18 (run c5946c9b)
Phone · 183.9 KB · a3a83b7cd1dd…
Motion capture of OpenCode (Qwen3.8 27B)'s output on Lava lamp v1, xhigh effort, 11.11.2026 06:18 (run c5946c9b)
Motion · 360.4 KB · cda6342756fb…

Along the row

Qwen3.8 27B at every effort.

  1. LowNot run
  2. MediumNot run
  3. HighNot run
  4. Qwen3.8 27B at extra high effort: desktop capture of run c5946c9b Extra high20m 19s

    Works

    This effort
  5. MaxNot run

Down the column

Other models at extra high effort.

Fastest median first · this run: 20m 19s

  1. Muse Ember at extra high effort: desktop capture of run 8970cfdfMuse Ember52.1 s

    Works

    23.4× faster than this run
  2. GPT-6.1 Terra at extra high effort: desktop capture of run 3ae8f53fGPT-6.1 Terra1m 15s

    Works

    16.3× faster than this run
  3. GPT-5.6 Luna at extra high effort: desktop capture of run c312c80fGPT-5.6 Luna1m 25s

    Works

    14.4× faster than this run
  4. GPT-5.6 Terra at extra high effort: desktop capture of run 7eac2b69GPT-5.6 Terra3m 50s

    Works

    5.3× faster than this run
  5. GPT-5.6 Sol at extra high effort: desktop capture of run 0e1f62c1GPT-5.6 Sol4m 05s

    Fails

    5.0× faster than this run
  6. GPT-6 Luna at extra high effort: desktop capture of run 61c1a78eGPT-6 Luna4m 24s

    Works

    4.6× faster than this run
  7. GPT-6 Sol at extra high effort: desktop capture of run 10bc68ffGPT-6 Sol7m 11s

    Works

    2.8× faster than this run
  8. Claude Haiku 5.5 at extra high effort: desktop capture of run 90ef705aClaude Haiku 5.58m 28s

    Works

    2.4× faster than this run
  9. GPT-6.1 Sol at extra high effort: desktop capture of run 0b6358d6GPT-6.1 Sol11m 37s

    Works

    1.7× faster than this run
  10. GPT-6.2 Astra at extra high effort: desktop capture of run 4716ed3fGPT-6.2 Astra15m 14s

    Works

    1.3× faster than this run
  11. GPT-6.2 Sol at extra high effort: desktop capture of run 9a6843d2GPT-6.2 Sol16m 03s

    Works

    1.3× faster than this run
  12. Claude Sonnet 5.5 at extra high effort: desktop capture of run a46b8eabClaude Sonnet 5.521m 18s

    Works

    about the same
  13. GPT-6 Astra at extra high effort: desktop capture of run 47bb0692GPT-6 Astra25m 55s

    Works

    1.3× slower than this run
  14. Claude Opus 5.5 at extra high effort: desktop capture of run 499f4389Claude Opus 5.528m 09s

    Works

    1.4× slower than this run
  15. Claude Fable 5.1 at extra high effort: desktop capture of run bcfef55bClaude Fable 5.132m 58s

    Works

    1.6× slower than this run
  16. Claude Fable 5.2 at extra high effort: desktop capture of run e47eb2feClaude Fable 5.240m 03s

    Works

    2.0× slower than this run

Hidden checks

7 of 8 checks passed.

  • Self-contained index.html is presentblocker
  • Visible page contentblocker
  • Visible autonomous animationblocker
  • Accessible on/off controlextra
  • 390 px mobile layout has no horizontal overflowextra
  • Local animation frame performanceextra
  • No browser errorsblocker
  • No attempted external or linked-resource requestsblocker

It works when every blocker passes. How it looks is reviewed blind: pending.

Run detailsagent config, what ran, tokens, source and revisions
  • Video (video): not published, videos are withheld. SHA-256 0f952a76958b35202df899b668241e671d3d7673ef4ba1d38a2e756682c96e44, 1511.4 KB.
Run ID
c5946c9b-21bd-4c70-9e62-61edc6f7706f
Benchmark
lava-lamp v1
Experiment
Single run
Started
11.11.2026 06:18 Tallinn
Graded
revision 1 · 11.11.2026 06:39

Works and checks, in full

All 5 blocker checks passed.

The rule: a run works when every blocker check passes. If any blocker fails it does not work. If none fails but one has no result yet, it stays pending, never counted as a pass or a fail. Extra checks are reported but do not decide it. Works is not a judgement of looks: that is the visual review.

This is a legacy contract v1 run: the task has no public requirements, and these are the original lava-lamp checks, not a versioned hidden verifier. It is never compared with contract v2 runs.

Contract v1 checks for this run, with blocker and result
RequirementCheckBlockerResult
none (v1)Self-contained index.html is presentBlockerpass
none (v1)Visible page contentBlockerpass
none (v1)Visible autonomous animationBlockerpass
none (v1)Accessible on/off controlExtrafail
none (v1)390 px mobile layout has no horizontal overflowExtrapass
none (v1)Local animation frame performanceExtrapass
none (v1)No browser errorsBlockerpass
none (v1)No attempted external or linked-resource requestsBlockerpass

Verification completed; eligible for benchmark results.

Visual review

Pending.

Pending: the captures have not been reviewed blind yet. Checks decide whether it works; how it looks is decided only in that review.

What ran

Requested and detected.

Detected values come from the run’s own output, never from our settings.

This run could use only file tools (read, write, edit, glob, grep): no shell.

Ran locally, every layer on one NVIDIA GeForce RTX 3090 24 GB: Qwen3.8-27B-UD-Q4_K_M.gguf from unsloth/Qwen3.8-27B-GGUF (Q4_K_M) in llama.cpp llama-server 0.9.0-dev (commit 494d0f4), with a 131,072-token context. No API, so no cost.

Requested agent config beside what the run reported
FieldRequestedDetectedEvidence
Modelqwen/qwen3.8-27bqwen3.8-27b (differs from the request)harness result, 7 of 7 API requests
Effortxhighxhigh7 of 7 API requests
HarnessOpenCodeopencode v2.2.7Version as the CLI reported it; binary SHA-256 b9d576000191…
Deadline60 min21m 17s usedCompleted

Tool decisions

5 calls · 5 accepted · 0 rejected

Each tool call is allowed or blocked by the harness’s permission settings before it runs. Allowed is not the same as succeeded: an allowed call can still fail.

  • edit3 calls · 3 accepted
  • read1 call · 1 accepted
  • write1 call · 1 accepted

Telemetry

Time and usage.

Generation time
20m 19s
First content
16m 28s
Output speed
34.8 tokens/s, whole run including tools
API-equivalent cost
none, local model (no API)
Actual charge
not reported

Input tokens

248,193 tokens
  • Cache reads204,32982%
  • Cache writesnot reported
  • Uncached input43,86418%

Output tokens

42,471 tokens
  • Output42,471100%
  • Reasoningnot reported

Usage as the agent reported it at the end of the run. Cache reads are part of input; reasoning is part of output.

Transcript and source

Plain-text downloads.

The redacted transcript and the frozen output bundle, with LF line endings. Hashes are of the original files.

Redacted agent transcript: 3.2 KB, 0 redactions. Download transcript.txt

Source files

1 file · 17.2 KB
Files in the output bundle
FileSizeSHA-256Redactions
index.html17.2 KBed83eb4b06ad6d543676e0e19bb484e77481ade9083b14a46874553265c0e5980
Bundle SHA-256 ed83eb4b06ad6d543676e0e19bb484e77481ade9083b14a46874553265c0e598

Revision history

This page shows revision 1.

Grades are append-only: a regrade adds a revision and never rewrites an earlier one.

Grading revisions of this run
RevisionGradedWorks
1 (current)11.11.2026 06:39works

Redaction: automated scan run; checked by Jaanus on 11.11.2026 06:43. Nothing needed redacting.