pretty_name: RAP E1 Luna Six-Way Evaluation
language:
- en
- zh
tags:
- pcb
- agent-evaluation
- rap
- tool-use
size_categories:
- n<1K
configs:
- config_name: e1_luna_six_way
data_files:
- split: test
path: e1-luna-six-way-20261003/results.jsonl
RAP evaluation data
Current release: e1-luna-six-way-20261003. It contains 98 frozen simulated PCB test tasks × 6 conditions = 588 episodes, all with GPT-5.6 Luna, seed 1, and at most 5 model API calls per episode. No models were called and no results were regraded for publication.
| Condition | Clean pass | Total tokens | Mean wall seconds / episode |
|---|---|---|---|
| Historical R2 | 24/98 | 966,897 | 41.77 |
| Historical RAP / R4b | 94/98 | 591,725 | 9.13 |
| Native V2 + R2 | 29/98 | 2,244,091 | 22.53 |
| Native V2 + RAP | 91/98 | 1,577,621 | 15.19 |
| Harbor Codex + R2 | 24/98 | 6,992,722 | 71.78 |
| Harbor Codex + RAP | 90/98 | 5,211,733 | 62.44 |
Browse and download
- Results and latency HTML — download and open locally.
- Setup summary HTML — each row's tools, history, settings, and comparison limits.
- Per-episode metrics, summary, 98 complete tasks.
- Raw evidence archives: one compressed archive per condition, retaining per-task prompts/requests, visible responses, tool calls/feedback, artifacts, grades, usage, and available run configuration.
- Provenance and redaction manifest, SHA256SUMS, offline validator.
The historical two arms have recorded actions and usage ledgers, but not all original HTTP request/response bodies. Current four arms include the stored request/response records. A file named private/grade.json inside an archive is a formerly host-private grader artifact, now deliberately included for audit; it is not a credential file. No hidden checker was exposed as an agent tool during the runs.
Scope and limitations
This is not all E1–E5, all nine models, or 196 tasks across six seeds. The full frozen manifest includes 223 task entries; only the 98 IDs in selected-task-ids.json belong to this release. Other historical baseline summary values may appear in provenance metadata, but their trajectories are not included.
Historical Luna used output cap 1,200 and omitted reasoning effort; current runs use 8,192 and medium. Their prompts, history handling and tool serialization differ. Native and Harbor use the same RAP backend and checker, but Harbor adds a general coding-agent loop, tool discovery and scratch tools. R2 is a single-object/net geometry interface, while RAP realizes whole bundles. The 5-call constraint creates a necessary-action bottleneck for at least 47 tasks under R2. Thus these are interface/pipeline results, not an isolated proof of superior general reasoning or a pure native-vs-text/harness ablation.
Latency includes both passing and failing episodes. Harbor includes job and environment setup; Native records its loop and final grading. Timing also depends on request load and worker conditions. Tokens are input + output per provider call, including cached input and reasoning output exactly once. Costs are reference estimates, not verified invoices.
Sanitization
Public copies remove credential fields and recognizable credential strings, replace local path prefixes, and remove opaque encrypted reasoning values (including JSON embedded in strings). Duplicate SSE chunks and runtime/authentication directories are excluded. Visible tool decisions, artifacts and numeric outcomes are retained. Local originals are unchanged. The manifest contains original and published SHA256 values; selected JSONL records are hashed without their line terminator. These sanitized files are not a byte-identical full transport replay package.
Reproduce the reported aggregates
Download the dataset repository, then run python verify_release.py. It verifies every published file checksum, all archive member checksums, 588 task/condition records, token arithmetic, and the success/token/call aggregates directly from detailed episode records. This validates the stored results without new paid API calls. Re-running stochastic model inference is a separate procedure and need not reproduce the exact numbers. Selected implementation files and run configurations are included for inspection, not advertised as a complete standalone environment image.
No new license is asserted for upstream software or task sources by this data card; existing ownership and notices remain applicable.