Two Judgments

"The same sentence can be perfectly useful to me and actively dangerous to a machine that quotes it."

One Transcript, Two Very Different Fates

Take a single transcript with a real flaw in it — the model hallucinated a stock phrase during a silent stretch, and misheard a name once. Now watch that identical text meet two different consumers:

Recall search. You search for a phrase, get a hit, and click through to the video at that timestamp. The hallucinated line? It matches nothing you'd search for, so you never see it. The misheard name? You open the video, hear the real word, and move on. The flaw is harmless, because the source is one click away and a human is verifying at the point of use.
A knowledge corpus. The same text gets retrieved and quoted downstream as fact. Nobody opens the video. The misheard name is now a confident, sourced-looking claim about something Dad never said. The flaw is dangerous, because there is no verification loop — the text is the evidence.

Same bytes. Opposite verdicts. That's not a contradiction to resolve; it's the actual structure of the problem.

Fitness Is Not a Property of the Data

Here's the idea worth carrying out of this whole quest: 'is this data good?' is not a well-formed question. Fitness isn't a property of data alone — it's a property of the triple (data, consumer, verification loop). Change any one of the three and the answer changes, with the bytes untouched. A transcript with noise is fit for a consumer who checks the source and unfit for a consumer who cites it blind. So a single 'quality: good' flag stamped on the data is a category error: it answers a question that can't be answered without naming who's asking and what they'll do to check.

This is why Recall keeps the two judgments structurally separate rather than tuning one threshold. The search judgment can afford to be generous — the human closes the loop. The corpus judgment has to be strict — nothing closes the loop downstream. The structural quality score (Track 3) informs both and decides neither, precisely because it's a property of the text and the decisions are properties of the use.

Find the Verification Loop, Then Set the Bar

The practical move this gives you is a question to ask at every boundary where data leaves your hands: who verifies, and where? If a human sees the source at the moment of use, you can tolerate noise — the loop catches it. If the data will be consumed by something that treats it as settled fact, the loop is gone and your bar must rise to compensate. The bar is not set by how clean the data looks; it's set by how much verification exists downstream of it. Recall's whole quarantine is just this principle made mechanical: search keeps its loop and stays permissive; the corpus has no loop, so approval defaults to no and only a human can say otherwise.

Code

One release, two consumers, two different bars·python

# The SAME release object, judged twice, differently.

def usable_for_search(release) -> bool:
    # Human closes the loop: hit -> open the video -> verify.
    # Noise is tolerable; the source is one click away.
    return release.has_segments()          # generous, on purpose

def eligible_for_corpus(release) -> bool:
    # No loop downstream: the text IS the evidence, quoted as fact.
    # So: default denied, human-approved only.
    return (release.status in ('reviewed', 'published')
            and release.corpus_eligibility == 'approved')

# quality_score informs BOTH and decides NEITHER --
# it's a property of the text; fitness is a property of the USE.

Exercise

Take one dataset or output in your own work that feeds more than one consumer — a log that's both eyeballed and alerted on, a scrape that's both browsed and aggregated, a model output shown to users and also stored. For each consumer, name the verification loop: who checks it, and at what moment? Then set a separate bar per consumer. Note any place where a single 'validated' flag is currently serving two consumers with very different loops.

Hint

The smell is one quality gate guarding two doors. Ask, per consumer: if this record were subtly wrong, who would notice and when? Where the answer is 'a human, immediately, with the source in front of them,' you can be permissive. Where it's 'nobody — it just gets used,' that consumer needs its own stricter, default-denied gate, even though the data is identical.