Configuration

Two files decide what a run is allowed to call, and what text it is allowed to serve. Both are in the repository, and neither is a secret.

writers.json

The file is the writer list. A use is allowed unless a clause in the maker’s licence or terms restricts it; silence is not a restriction. A name that is not on it is refused, and a name that contains cloud is refused even if someone adds it.

Every entry needs its proof, and the list is refused when it loads if any entry lacks one:

Field	What it holds
family	The model family, for the critic check
licence	The licence the maker’s card names
terms_url	The maker’s live page (https://)
terms_date	The day it was read, YYYY-MM-DD
training	The registry’s verdict on training with its outputs: allowed, conditions (with conditions written out) or barred (refused)
checked_by	Who read it
registry_id	Its record in rnd’s licence registry (licences/registry.json), carried onto every receipt

Writers	Critics
nemotron-3-ultra-550b-a55b, deepseek-r1-0528, nemotron-3.5-super-vl-120b-a12b, nemotron-3.5-bolt-74b-a9b, muse-glimmer-30b, nemotron-3.5-lightning-30b-a3b	muse-glimmer-30b, nemotron-3.5-lightning-30b-a3b

A critic of the writer’s own family is refused, so Muse-Glimmer reads the Nemotron and DeepSeek writers, and Lightning reads Muse-Glimmer.

The pins. Models are served by llama-server, built with CUDA 13.4, at Q8_0. A listed model is refused until it’s pinned: repo, revision, gguf_file and gguf_sha256. llama_server pins the binary by commit and sha256. The wave script hashes the GGUF and the binary before launch and passes them in (--writer-gguf-sha256, --critic-gguf-sha256, --llama-server-sha256). Any drift refuses the run before a byte is sent.

The run’s context. --partition is production or trial:<name>. --machine is the wave script’s CUDA machine line, as JSON: driver, compat (true or false), runtime, card, and where it ran (colab_release_tag, pod_id or rig).

An entry with rehearsal_only: true writes under a rehearsal/ directory, and a request to write it anywhere else is refused. An entry with role: critic-comparison runs in critic-control only, never as a run’s critic.

--writers defaults to writers.json in the working directory. Point it elsewhere only when you mean to.

Card values. --cards points at rnd’s experiments/local-builders/card-values.json. The writer and the critic each run at their card: every sampling key (temperature, top_p, top_k, min_p, repeat_penalty, presence_penalty) and the chat_template_kwargs go on every call. The context is 32,768 tokens and a reply may use 16,384. A card’s transforms list may name fold_system, which moves system text into the user turn for R1-Distill; reattach_thinking is J2’s and is recorded here as not applied. An entry’s card names its candidate when that differs from served_as. Receipts record the file’s sha256, the candidate and what was sent. A real write or critic-control without --cards is refused.

The grant

A step runs only inside a grant unless its machine line names a rented host (colab_release_tag or pod_id). A line that names the rig needs the grant whatever its card says. Off a rented host, write and critic-control require:

• --grant, a grant id for the batch window
• --heartbeat, a file whose modified time is under 60 seconds

The heartbeat file is not opened. There is no flag that turns the check off. --mock never reaches it, because the mock loads nothing.

The pack

pack sync copies the humans’ text out of a local aspire-si clone at one commit.

uv run role-set pack sync --repo <aspire-si clone> --commit <sha> --role verifier --out roles/verifier
uv run role-set pack verify --pack roles/verifier

--role defaults to verifier and --out defaults to roles/verifier. The clone path is not stored. The commit is. pack verify recomputes the hash and checks it against the manifest sync wrote.

What sync is expected to fill: the role lessons, one file per trait, the kind table, amendments, rules, an exclusion list, the puzzle-structure menu, pressure kinds, injection kinds, and exemplars taken from merged batches. This repository does not ship that text. roles/verifier here has sources.json, exemplar-exclude.json, and a README that names the layout. Run sync before a write that should see the lessons.

write does not scan the pack for operator identity. Scan a pack before you publish one.

The pack’s files, the Advocate draft, and how the hash is built are on Role packs.

What lands on an accepted receipt

An accepted receipt spreads the run’s pins. Always: the plan hash (plan_sha256) and the pack hash (pack_sha256), plus the pack commit. On a real run: the partition, the wave, the writer’s and critic’s family, repo, revision, GGUF file and hash, licence and proof URL, the llama-server commit and hash, the machine line, and the hash of writers.json itself. When you passed exemplars: exemplar_commit, and exemplar_exclude_sha256 when you also passed an exclude file. When the stuck rule was loaded: stuck_rule_sha256. The exemplar text itself is not hashed onto the receipt. The commit is the pin for which text was read.

drafts.jsonl is every writer call. It does not carry those pins on each line. batch-report.json is the run summary. It is not a second copy of the pins. A run with no accepted row has no receipts.jsonl.

The gates package version and its hash are on the receipt too, so a later reader can see which checks the row was built against.