Recovered scripts: originals, adapters and their limits
These files were recovered from the author’s inference host during the transparency follow-up. The recovery manifest records source/released hashes and whether a file is exact, adapted or an example. Read the code corner for worked examples and the disclosure for what is still missing.
| File / family | Status | What it supplies |
|---|---|---|
| bench_agents4.py | Byte-identical original; SHA-256 starts ab17f129… |
Synthetic text/image growth, streaming TTFT/TG, token usage, per-cell summary |
| needle160k.py | Byte-identical original | The generated long-prompt needle probe; running this sends a real large request immediately |
| run_grid.sh | Portable adapter of three recovered wrappers | Sequential n1–4, single-card caps, and non-spec image cell; requires an already configured server |
| convert_model.sh | Portable adapter | Recorded M2 conversion recipe; operator supplies image and model paths |
| Pi recipe and provider / settings examples | Recovered base recipe plus sanitized settings | Pi 1.1.0 client, 131,070 context, 32,768 max output, reserve 34,816; not the raised-budget retry |
Existing scripts/run_falling_blocks_*, prompt generator and event parser |
Matched against server copies | The agent runner now includes the recovered optional PI_NETWORK mode; other matching files were already present |
tools/editorial_fleet/fleet.py, round2.py, round3.py |
Byte-identical recovered orchestration/prompts | The initial 18 editorial tasks; code is available in the git checkout, excluded from the rendered site |
tools/editorial_fleet/documentation_followup.py |
New follow-up, separately labelled | Environment/history/code-corner source review |
Original does not mean ideal
The original v4 harness sends synthetic tool-result messages and places a system-labelled instruction inside a user message. It is agent-shaped traffic, not a real tool-executing agent. It appends text reconstructed from streamed replies and uses usage token counts where provided. Its temperature is 0.7; it supplies no sampling seed. The growth cap is checked on the previous prompt, so a final step can overshoot it.
TTFT includes the first recognized content, reasoning content or tool-call chunk. Streaming gaps are chunk timings, not necessarily one-token intervals. The original stores request errors in the summary but can still return process exit 0; inspect errors before accepting a run. The adapter adds a final recorded-error check and refuses to append into an existing result CSV.
With --csv-append, v4 also inspects the named Podman container and records selected environment variables and the full command in a command column. That can expose operational details or credentials. Its raw CSV/logs are private working outputs, not files to commit; published campaign exports omit that column. Supplying a non-existent container makes inspection fall back to client arguments; that changes provenance capture and must be disclosed.
run_grid.sh does not enable/disable speculation, select GPUs, restart inference or verify model identity. The profile name labels requests only. Match the server configuration to the campaign before scheduling it; do not run grids alongside other inference work. Its default CONTEXT_LIMIT=131070 preserves the recovered client wrapper’s guard, which is distinct from radiance’s recorded 131,072 server limit. Logs/cell IDs differ from the original wrappers; the adapter excludes their automatic game-fleet stage.
The conversion adapter requires Podman and the matching image recipe. It writes a new model bundle, mounts input read-only, and refuses to overwrite an existing output. It was syntax-checked here; no conversion was rerun. Model files and the engine build context are not included.
Pi build inputs
The base recipe pins the Pi npm package at 1.1.0 but not the Node image digest, apt packages or transitive npm dependency lock. It is therefore a recoverable recipe, not a fully frozen image. Copy the examples into a private build directory as models.json and settings.json, then set your endpoint/credentials before building. Those operator files are gitignored in this directory. On a Linux host, a loopback API requires the recovered PI_NETWORK=host option or an equivalent reachable endpoint; bridge-container localhost is not the host.
The examples preserve the older client context and budgets even if the current server supports 204,800. Changing them defines a new client configuration. The later successful single-card retry used different budgets; do not label its settings the same as these examples.
Replaying the editorial tool
Copy the three recovered editorial scripts beside each other in a private scratch directory and provide a public-only repository snapshot in its repo/ subdirectory. Each script makes live API requests to localhost port 8000 and writes drafts/tool traces under its workspace. They expect the recorded served model ID and an already running compatible endpoint; no service is created.
fleet.py runs six initial tasks; round2.py and round3.py import it and replace the tasks for their six-task passes. The follow-up file runs three additional documentation tasks. Requested settings are recorded in the source. Do not launch all passes simultaneously. Treat outputs/traces as private until reviewed.
The original prompts retain then-current mistakes, including the subsequently corrected fleet-exit statement. They are historical evidence, not current author instructions or validated facts. Read the historical account and the editorial corrections before reusing them. Prompts and tools can now be inspected, but the exact changing per-pass repository snapshots and all original outputs are not supplied; a new run is not deterministic replay.
Home · Compare · vllm-radiance · radiance · Charts · Stuff for nerds · Disclosure