The Quarantine

"The pipeline finishing is a fact about the pipeline. It is not a claim about the truth of what came out."

The Promotion That Should Never Be Automatic

Recall's transcripts are tempting. There are thousands of them, they're searchable, they're full of Dad's actual thinking — exactly the raw material a knowledge system would love to retrieve from. And the pipeline just declared them complete with a clean structural score. So: ship them into the corpus, right?

No. And the refusal is architectural, not cautious hand-wringing. Every new or imported release lands with status candidate and corpus eligibility excluded. Default denied. To become corpus-eligible, a release must walk an explicit path — excluded → candidate → approved — and every step of that walk is a human decision. Neither the transcription completing nor the quality audit coming back clean can perform that promotion. The machine is structurally incapable of promoting its own output into a trusted context.

Why Default-Deny and Not Default-Allow

Because of what's actually in those transcripts. The legacy corpus is full of silence hallucinations, boilerplate stock phrases the model emits when it hears nothing, and Korean/English confusion. That contamination is fine for Recall's own search — you search, you get a hit, you open the video and see for yourself. A human is in the loop, verifying at the source. But a knowledge corpus is quoted downstream as fact, with nobody checking the video. The same text that's harmlessly noisy in a search result becomes a confidently-wrong citation in a knowledge system.

So the default has to be denial. This is the fail-safe-defaults principle from security design, applied to data trust: when the decision is 'should this be trusted,' the safe failure mode is 'no.' If a review never happens, nothing bad ships — the transcript just stays useful for search and invisible to the corpus. Flip the default and the failure mode inverts: forget to review, and contaminated text silently becomes cited knowledge. One of those failure modes is boring. The other one is garbage-in, garbage-out at the scale of an archive.

The Quarantine Costs Nothing, Which Is the Point

Here's what makes this design honest rather than merely strict: Recall is a complete product with the corpus integration entirely absent. There's no runtime dependency. Search works, the editing handoff works, the whole engine works, with zero transcripts ever approved. So the quarantine isn't a bottleneck holding back the product — it's a wall the product doesn't need to cross to be useful. If the archive never clears the quality bar, the integration simply stays absent, and nothing is weakened. That's the tell of a well-placed trust boundary: the safe side is fully functional on its own, so there's never pressure to weaken the gate just to ship.

Code

Default denied; only an explicit human walk promotes·python

# Every new or imported release lands here. No exceptions.
release.status = 'candidate'
release.corpus_eligibility = 'excluded'      # DEFAULT DENIED

# The only path out is explicit and human:
#   excluded -> candidate -> approved
#
# Neither of these may perform that transition:
#   - the transcription completing
#   - a clean structural quality audit

# The export guard refuses anything that didn't walk the path:
def serialize_for_corpus(release):
    if release.status not in ('reviewed', 'published'):
        raise NotApproved
    if release.corpus_eligibility != 'approved':
        raise NotApproved
    return export(release)

# Forget to review -> nothing ships. That's the safe failure.

Exercise

Find a place in your own work where data produced by a pipeline gets consumed by something that treats it as true — a report, a dashboard, a model's training set, a retrieval system. Ask: what marks it as trustworthy, and can the producing pipeline set that mark itself? If completion automatically means trusted, design the quarantine: a default-denied flag, an explicit human transition, and a guard at the consuming boundary that refuses anything unpromoted.

Hint

Two checks reveal whether you have a real trust boundary: (1) can the thing that produces the data also declare it trustworthy? If yes, there's no boundary — just a rename. (2) If nobody ever reviewed anything, what happens? The safe answer is 'nothing ships and the rest of the system still works.' If the answer is 'everything ships anyway' or 'the product breaks,' the gate is decorative or will be bypassed.