The description gate

The gate decides whether an approved record’s description can be trusted as a description of its image. It runs with no model, so its verdict can’t be shaped by a model that is later measured on the same text.

The pinned arithmetic

• Words are re.findall(r"\w+", s.lower()).
• A content word is a word not in the frozen stopword list (dataset_desk/stopwords.txt, 212 words). Its sha256 is exposed as description_gate.STOPWORDS_SHA256 and written into every audit, so two runs can show they used the same list.
• A share is the number of the description’s words (counting repeats) that occur anywhere in the prompt, divided by the description’s word count. The content-word share is the same over content words.
• Contiguous runs and the subject-line slice are matched on these word sequences.

The tests

A description fails if any of these hold.

Test	Fails when	Why
Equals or inside the prompt	its words equal the prompt’s, or appear as a run inside them	it is the prompt
Shared run	it shares a contiguous run of 6 or more words with the prompt	it copies a clause of the prompt
All-word share	0.90 or more of its words appear in the prompt	it is the prompt reworded
Content-word share	0.80 or more of its content words appear in the prompt	the same, ignoring filler
Wrapped subject	the record’s canon.subject has 4 or more words and they appear as a contiguous slice of its words	it is the subject line with padding
Too short	under 8 words, or under 5 content words	too little to pick the image out
Exact duplicate	"".join(words(a)) == "".join(words(b)) for another description in the set	the same text on two images
Near duplicate	a content-word Jaccard of 0.85 or more with another description	nearly the same text
Too long for SigLIP2	more than 64 tokens	SigLIP2 would truncate it

Notes:

• “Exact duplicate” joins words with no separator, so cobble-stone arch-way matches cobblestone archway.
• The duplicate tests compare every pair across the whole set, not within a folder.

The token count

The SigLIP2 test is the one that needs a library. It counts

len(tokenizer(s.lower(), padding=False, truncation=False)["input_ids"])

on the google/siglip2-so400m-patch14-384 tokenizer at a pinned revision, special tokens included. Counting through the processor’s defaults would pad and truncate to 64, so a description could never fail.

From the command line, pass report --count-tokens to run it: the tokenizer loads from the local cache, offline, and is needed only then. Without the flag, or as a library call with no token_counter, a record that passes every other test is unchecked, and an unchecked record is never marked as passing.

Reading description-review.jsonl

Each line is one approved record:

• gate_status: passed, failed or unchecked;
• flags and reasons: which tests failed, in words;
• the statistics behind them: the shares, the longest shared run, the largest Jaccard, the token count;
• author_kind: human for existing text, captioner_assisted for a drafted one.

What the gate can’t see

It is lexical. A description that paraphrases its prompt in different words passes. After a benchmark’s counted run, the embedding fit check in the report (does the description find its own image?) covers that case.