A file format for claims that declares the class of every source, and never arbitrates truth.
A .ross file is a portable dossier about one subject. Every claim inside it carries the class of the source it rests on. The file opens with a computed card that tells you how well grounded it is before you read a word of the content.
It does not tell you what is true. It tells you what kind of ground you are standing on.
+------------------------------------------------------------------+
| AVISO ROSS |
+------------------------------------------------------------------+
| Master of Puppets / Metallica / 1986 |
| lido como: jornalista |
| |
| SUSTENTACAO MINIMA |
| [####........................] 13.7% |
| |
| 14 afirmacoes |
| 4 contestadas por outra fonte |
| 13 sem verificacao na fonte original |
| 1 sem fonte nenhuma |
| |
| nenhuma afirmacao deste arquivo foi verificada na fonte |
| original. em ross, isso rebaixa o dossie inteiro. |
+------------------------------------------------------------------+
That card is the whole idea. It is borrowed from the content advisory that used to run before a TV programme, except the warning is not about language or nudity. It is about how much of what follows is actually held up by anything.
The reference dossier in this repo, masterofpuppets.ross, was written by a language model from memory, with no source ever consulted. The format caught it immediately and scored the file at 13.7 percent. That is not a bug in the demo, that is the demo.
An encyclopedia article written the same way looks identical to a well researched one. A .ross file cannot hide it.
The moment a format says "this source is trustworthy", it becomes an arbiter of truth. Arbiters of truth are attacked, captured, and centralized. .ross refuses the job.
Instead, each claim declares only what class of source it rests on:
| class | meaning |
|---|---|
derivacao |
an argument the reader can redo without trusting anyone |
primaria |
primary document |
judicial |
court record |
revisada_por_pares |
peer reviewed work |
veiculo |
press outlet |
terciaria |
encyclopedia or compilation |
autodeclarada |
statement by an interested party |
fa_site |
fan site or community |
rede_social |
social media post |
sem_fonte |
no source declared |
The weighting lives in the reader, not in the file. A reading profile decides what each class is worth. Same bytes, different readers, different verdicts:
| profile | verdict on the same file |
|---|---|
pesquisador |
8.4% |
jornalista |
13.7% |
fa |
27.8% |
Nobody had to be declared right. The file only had to be honest about its footing.
Most knowledge systems treat a contradiction as a defect to be resolved. .ross treats an unresolved contradiction as a legitimate permanent state.
The reference dossier carries three real disputes about the album, including the death of Cliff Burton (the driver said black ice, James Hetfield said publicly that he walked the road and found none, and a Swedish court acquitted the driver). The format does not pick a winner. It shows the reader that two of the three positions are statements by interested parties and one is a court record, and it leaves the conflict standing.
The same happens with Dave Mustaine's claim of an uncredited writing contribution to Leper Messiah. Both sides get the same class: interested party. The format makes the tie visible instead of hiding it behind an editorial decision.
python3 ross.py masterofpuppets.ross # card + claims
python3 ross.py masterofpuppets.ross --tudo # plus compiler notes
python3 ross.py masterofpuppets.ross --perfil fa
python3 ross.py masterofpuppets.ross --so-cartaoNo dependencies beyond the Python standard library.
A .ross file is a per claim provenance record that a machine can read. Most model hallucination happens because the text a model ingested has no provenance at the sentence level: a rumour and a court record arrive as the same flat prose. A corpus of .ross files is a corpus where every sentence already knows where it came from and how weakly it is held.
rossgen.py turns raw text into a valid .ross file with a local model through ollama. Claims get split, source classes get inferred from what the text alleges, and an embedding pass narrows candidate pairs before the model judges which ones actually contradict each other.
python3 rossgen.py input.txt --sujeito "Subject" --tipo albumThe generator is itself exactly the kind of weak source this format exists to expose, so it is not allowed to flatter itself: everything it emits is born with verificacao: pendente and the model name stamped into coleta.por. A dossier written by a machine says so on its own first screen.
These are unsolved and stated plainly rather than hidden.
- Who writes the file. A format that discloses sourcing does not stop anyone from lying about the class of a source. Signatures constrain this (a claim signed by a named party can be held against that party) but do not eliminate it.
- Merge semantics. The trust layer wants content addressing as identity, hash-linked revisions as history, and signatures for authorship, which is essentially git's model. But git exists to make divergence converge, and this format needs to preserve disagreement permanently. Whether git's model survives that inversion is an open question.
- Convergence signal. When several independent parties sign the same claim, the file gains a grounding signal with no central authority and no blockchain of any kind. The threshold at which that signal means something is not defined yet.
- Field names are Portuguese in v0.1 because the author is Brazilian and the first reader needed to understand it on sight. v0.2 moves the keys to English and keeps a translation table.
v0.1, written in a single night. The schema is in SPEC.md. It is small enough to read in five minutes and to reimplement in an afternoon, which is on purpose.
MIT. See LICENSE.