The shelf

Claims about compute, filed side by side so that someone who was not there can see what each one supports and what it does not. A card format, a checker with no dependencies, and a rule for pulling other people's shelves into yours without granting them standing. Nothing here ranks.

Everything on this page runs in your browser from committed plain text. No account, no upload, no analytics. The same checks run from the command line with python shelf.py; shared examples and generated rule-boundary cases check that both engines agree.

What is on it

Each card answers five questions: (a) what exactly is claimed, under which conditions, over what period; (b) what population, denominator and acceptance rule produced the number; (c) who observed, who controlled the conditions, who granted access, who paid, and whether the observer benefits from the outcome; (d) what a reader can obtain; (e) whether anyone inspected the evidence, recomputed the result, or independently repeated the operation, kept as three separate facts. Measured, imported, modeled and attributed are four separate flags, never one grade. Unknown is null, never a guess.

hub	card	spine	M I Mo A	inspected / recomputed / repeated

M = measured, I = imported, Mo = modeled, A = attributed. + true, - false, ? unknown. Flags describe the card's own evidence; they do not rate it.

Which cards can support a question?

State the unit the answer would need, the period it is about, and whether it must be a measurement. Each card either supports the question, bounded by its own limits and recorded checks, or cannot be used, with the rule that says why. This is the check a stranger performed by hand on 2026-09-24; here it is mechanical.

Check a filing

Paste a card. The check reports whether it carries the five answers, the four flags, the three checks, a publication-date basis, an observer-benefit record and explicit hourly units. It cannot establish that a cited source warrants a value, that a populated disclosure is complete, or that a claim is true.

Compose two cards

Records combine only when identities, periods, scopes and assumptions agree. Pick two cards and a purpose; the engine either refuses with the rule it comes from, or finds no refusal. Finding none is not admission.

File a card

Start from card.skeleton.json or copy the five sections from a relevant card, then cite the material that supports each populated value. The normative format is in SPEC.md. Three rules catch most first filings:

Dates. publication_date is null unless the source establishes publication, and it needs a publication_date_basis. retrieved_at records when you obtained the source. Experiment, artifact, index and publication dates are separate; retrieval never fills one of them.
Price units. USD / GPU-hour for one GPU's hourly price, USD / seat-hour for a whole configuration with its GPU count in conditions.gpus. USD/hour is ambiguous and refused. Accepted-work prices use USD / 1000 accepted requests. A starting price is a lower-bound offer, not an obtainable seat or a measured cost.
Observer benefit. c_parties.observer_benefit records whether the observer benefits from the outcome (true, false or null), the relationship, and the basis. A known judgment needs a basis. Funding and benefit are different facts. A relationship does not establish misconduct.

To put a card on this shelf, open a pull request that adds one line to shelf/data/cards.jsonl and passes python shelf.py validate. Review checks the cited sources; filing and machine validation do not grant standing.

Run it yourself

git clone https://github.com/BigBirdReturns/axm-tools
cd axm-tools/shelf/scripts
python shelf.py validate ../data/cards.jsonl
python shelf.py which --unit "USD / 1000 accepted requests" --period 2026-09 --measured
python shelf.py compose a.json b.json --for same_period
python shelf.py pull https://example.org/their/cards.jsonl --hub their-name
python -m unittest discover -s ../tests -p 'test_shelf.py'

Python 3.8 or later, standard library only. No package to install, no key, no service.

Hubs

A hub is any URL that serves a cards.jsonl. This page reads the hubs in data/hubs.json. shelf.py pull copies a hub's cards into data/staging/<hub>/<time>/ with the exact bytes, their hash, the time, and each card's filing errors. Imported cards stay candidates. Moving one onto your shelf requires a recorded local decision by an authorized person or procedure, bound to the exact captured input. Source checks remain attributed to whoever performed them; importing does not grant local standing. To list your hub here, open a pull request adding it to data/hubs.json.

Disclosure

On the first hub's run card, Hot Aisle's credit paid for the AMD arm and Hot Aisle is a sales prospect of the shelf's author. The H100 arm was paid in cash. One run per arm, no repeats. The card says so on its spine.