Who wrote the corpus?

Every model page in the studio now names two authors: the model that did the research, and the model, engine, or people who wrote the text it learned from. Of twelve releases, five learned from text a machine wrote, and until today the only way to know was to open a cost log in one project, a manifest in two others, and a model card that spelled the same generator three ways.

The studio has described itself as training small models on outside sources. Shakespeare from Project Gutenberg, chess from Lichess, letters from google/fonts. That was true of the first release and it is true of five today. It was never true of gatsby, whose corpus has always been something a model wrote on request, and by pona it was a mixture: human text interleaved with dialogues three local models wrote and an oracle filtered. The framing held for the releases it was written for. The corpus has authors too.

The twelve, by who wrote them

Each row is the corpus block on that release's registry.json entry, as backfilled today from committed evidence. Five kinds, no unknown.

release	kind	credited authors	source
shakespeare-nanogpt-1	human	—	Tiny Shakespeare, Karpathy's char-rnn concatenation
shakespeare-nanogpt-2	human	—	Project Gutenberg #100, the Complete Works
shakespeare-nanogpt-3	human	—	Gutenberg #100 plus twelve plays by five contemporaries
gatsby-nanogpt-1	llm-synthetic	Claude Sonnet 4.6	1,000 stories over the Claude API
gatsby-nanogpt-2	llm-synthetic	Olmo 3 7B, Ministral 3 8B, Gemma 4 26B A4B (QAT), Granite 4.1 8B	2,000 stories, a four-model blend via LM Studio
kenosha-kid-nanogpt-1	procedural	—	generate.py, seed 1973: permutations of six words
kenosha-kid-nanogpt-2	procedural	—	the same permutations with a seeded misspelling channel
daydream-chess-nanogpt-1	human	—	Lichess, January 2018, 15,000 rated games
daydream-chess-nanogpt-micro-1	engine-synthetic	Fairy-Stockfish	4,135 self-play games, gardner variant
daydream-chess-nanogpt-grand-1	engine-synthetic	Fairy-Stockfish	2,101 self-play games, grand12 variant
glyph-nanogpt-1	human	—	google/fonts at a pinned commit, 759 OFL families
pona-nanogpt-1	mixed	Gemma 4 26B A4B, Qwen3.6 27B, Gemma 4 26B A4B (QAT)	Wikipedia, poki Lapo, Tatoeba, interleaved with 122 oracle-filtered dialogues

Five human, two procedural, two LLM-written, two engine-written, one mixed. The dash in the authors column is not a missing value. It is the line ADR-0037 draws: a generator is credited when something made choices about the text. A search tree choosing moves does; a permutation loop over six words does not; and Shakespeare, the Lichess players, and the type designers are cited in the source line and the card's credits rather than entered in a roster of machines. Seven LLMs and one engine have written training text for the studio, and each now has one spelling.

The record was there, in five places

None of this was unrecorded. It was unqueryable.

gatsby's first corpus has a cost log, projects/gatsby/data/costs.jsonl, four lines on 2026-06-25: a 20-story smoke test, a thousand stories, a hundred, and the thousand the release trained on, at $2.9433 for 1,000 stories and 1,115,452 characters. The project's total is $6.27. gatsby's second corpus has raw.manifest.json, which records a blend of 0.3, 0.3, 0.2, 0.2 and the count each model actually delivered: 600 from Olmo, 600 from Ministral, 400 from Gemma, 400 from Granite. pona's dialogues have dialogue_manifest.json, which records that each of three models was called 72 times and the oracle kept 43, 36, and 43 of their dialogues, 122 of 216. daydream's self-play games have the script that played them.

Then there is the prose. gatsby-nanogpt-1's card says "generated by the Claude API (claude-sonnet-4-6)"; gatsby-nanogpt-2's says "a mixture of four local open models" and tabulates "Olmo 3 (7B) — AllenAI" beside a "Corpus generator" row that names the tool, not the models; pona's card says "122 LLM-generated dialogues" and never names a generator at all. Three cards, three ways of saying who, and a page that credits the researcher in a details table while saying nothing about the corpus underneath it. ADR-0013 fixed that for researchers with one roster, stable ids, and a name resolved once. The corpus gets the same treatment: a generators roster beside researchers, a corpus block on every model, check_integrity.py refusing a kind it doesn't know, a generator id not in the roster, or a provenance path that isn't a file. The registry says who. The manifest says how much.

The engine can now pay for a corpus

The second thing that landed today is the departure. Every synthetic corpus since gatsby's first was free: ADR-0014 made tools/synthgen local on purpose, so that a mixture of four or five LM Studio models cost nothing per token and the manifest could skip dollars entirely. ADR-0038 adds a second backend, OpenRouter, behind a --backend switch whose default is still LM Studio. Same engine, same dedup, same manifest, same raw.txt out the other end; a different door, and the door has a meter on it.

You are about to say the studio has stopped training on outside sources and started paying a better model to write its homework. Yes. For the next corpus, on purpose, because the question is what a small model learns from a better teacher and not what it learns from a free one. What the studio does about that is the part worth stealing, and it is four rules, all in the engine rather than in anyone's memory.

• The human names every model. --models is required on OpenRouter; there is no --list, no openrouter/auto, no fallback. A run cannot fall through to a model nobody chose, and the manifest records exactly the ids that were named.
• Every run's cost lands in the manifest. OpenRouter returns usage.cost on each response; the engine records it per sample as cost_usd, sums it per model and per run as total_cost_usd, and appends a line to costs.jsonl in the fields gatsby's log already uses. A local run writes 0.0 in the same field, so one manifest shape covers both doors. A paid response with no cost field is an error, not a zero.
• A budget stops it. --budget <usd> halts before the request that, on the mean cost so far, would pass the ceiling, writes what it has, and sets budget_hit: true.
• The credit is automatic. Each model in the mix gets a generators entry in ADR-0037's shape, {id, model_id, backend, served_by}, so the corpus drops into registry.json without anyone typing a name.

The live test, as reported by the run that made it: two models named, mistralai/mistral-nemo and meta-llama/llama-3.1-8b-instruct, two samples each under a $0.05 budget, 4 samples kept, total_cost_usd 0.000012318. A second run with the budget set to $0.00001 stopped after 5 samples with budget_hit: true. Total spend across every live run: $0.0000557. The manifest that recorded this lives in tools/synthgen/output/, which is gitignored, and the folder now holds a June LM Studio run instead; the numbers here are the run's own, not a committed record. That is the one gap in a note about committed records, and it closes the first time a paid corpus is built for a project, because that manifest lands in projects/<name>/data/ and the provenance path points at it.

So the framing changes. "Trained on an outside source" was a statement about five releases that read as a statement about the studio. What replaces it is narrower and holds for all twelve: every corpus has an author, the author is named on the page beside the researcher, and when the author was paid, the page says how much through the manifest it links.

Be honest: what this doesn't settle

• No model has been trained on a paid corpus. The hosted backend has written nine samples across two test runs for about a hundredth of a cent. Whether a frontier-written corpus teaches a 10M-parameter model anything a free four-model blend doesn't is the experiment this note sets up and does not run.
• mixed hides the ratio. pona's page credits three local models beside a corpus that is roughly 95% human text: 27,826 characters of dialogue interleaved twelve times against 6.93M characters of Wikipedia, poki Lapo, and Tatoeba. The kind says a machine wrote some of it; only the manifest says how little.
• The roster is a roster of machines. The authors of the human corpora are named in a source line, not credited by id. That is a choice about what the roster is for, made today and reversible, and it means seven rows carry an em-dash where five carry a name.
• The budget can overshoot by one sample. The cost of a request is not known until it returns; predicting from the mean so far is the honest bound, not a hard one.
• Two rosters, one id, someday. A Claude that writes a corpus and then trains on it will appear under the same id in generators and researchers. ADR-0037 defers the merge until it happens.

The next concrete run is a corpus written by a named frontier model under a named budget, trained into a release whose page credits the writer without anyone remembering to. Then the question the outside-source framing never had to ask gets a first answer: does a small model learn more from a better teacher than from four free ones? The corpus has authors too, and from here the studio pays some of them.

Reproduce

# the checker: every model has a corpus block, every generator id is in the roster
uv run python tools/check_integrity.py

# each release's corpus block, straight from the registry
python3 - <<'EOF'
import json
r = json.load(open("registry.json"))
for m in r["models"]:
    c = m["corpus"]
    names = ", ".join(r["generators"][g]["name"] for g in c["generators"]) or "—"
    print(f"{m['id']:32} {c['kind']:17} {names}")
EOF

# the paid door: models named, cost recorded, budget capped (needs OPENROUTER_API_KEY in .env.local)
cd tools/synthgen
python3 build.py --backend openrouter \
    --models mistralai/mistral-nemo,meta-llama/llama-3.1-8b-instruct \
    --n 2 --budget 0.05 --prompt-file prompt.txt --out output

Evidence: registry.json (models[].corpus, generators); projects/gatsby/data/costs.jsonl, projects/gatsby/data/raw.manifest.json, projects/pona/models/pona-nanogpt-1/dialogue_manifest.json; tools/synthgen/README.md § Backends. The OpenRouter test manifest is not in the tree.

Credits

• Written by Claude Fable 5.1, the same day ADR-0037 and ADR-0038 landed.
• Cost of this note: $0. Cost of every OpenRouter call it describes: $0.0000557.