Eve library

Review, escalation, and incident response

Guide

Revision 1

On this page

Phase 182, 182.C.39.06. Review and escalation rules, non-overridable gates, subject requests, takedowns, and incident response.

1. Who may see what#

A reviewer does not need the original to do the review. The review question is "is this output acceptable", and the output is the take. The reference photograph the subject enrolled is an input — and it is the most sensitive artifact in the system, while the review queue is the highest-traffic surface that touches this data (human-video-reviewer-access.ts).

So biometric originals and provider payloads are released only when the role and the case both require it, not when either does.

Escalation out of the building splits by audience, and the overlap is smaller than it looks (human-video-escalation-export.ts):

  • Support is inside our trust boundary and outside the tenant's. They may see which tenant, which job, and what failed. They do not need the likeness.
  • A provider is outside both. They need the request shape, the parameters and the error — and the tenant's identity is not their business. A provider who can correlate failures to named customers has been handed a customer list.

The two policies differ on the tenant identifier, which is the field everybody assumes is harmless. Every export stays audit-safe: a reader can tell what was removed.

2. The gates nobody may override#

"Non-overridable" is a list, not a principle, because every gate looks overridable at two in the morning. Four gates may be waived (human-video-quality-waiver.ts):

quality-aggregate, audiovisual-sync, encoding-conformance, output-shape-valid.

Each of those protects the buyer, who is in the room and can decide the video is good enough for them. Twelve may not be waived, and each one names the person it protects — because "not waivable" invites a request for an exception, and naming the absent person does not:

Gate Who is not in the room
consent-evidence the person whose likeness this is did not agree to be in this decision
identity-preservation waiving it ships a video of somebody who may not be who it claims
safety-scan the people a harmful video reaches are not party to this
exact-speech putting words in a real person's mouth is the harm this phase exists to prevent
synthesis-disclosure the disclosure is owed to the viewer; a tenant cannot waive an obligation that is not theirs
rights-and-license the rightsholder is not in the room
provenance-c2pa a released asset with no provenance cannot be given one later
final-bytes-binding without it the proof is about nothing in particular
provider-lineage the record of what produced these bytes is not a quality property
watermark-coverage the mark is a disclosure control rather than a quality one
policy-compliance a policy a reviewer may waive is not a policy
human-review-ready waiving readiness for review is waiving the review

A waiver on one of the four is still bounded and co-signed: it expires (seven days maximum), it carries a queryable reason code rather than free text alone — free text cannot answer "how often does this happen" — and it needs a second approver when the person asking is the person who owns the job.

3. Subject requests#

A subject is not a tenant, and every index in this system is built the other way round (human-video-subject-requests.ts). A tenant knows which packs belong to which subject because they enrolled them; we know it only through the authorization. So "everything about this person" is answerable only as far as the authorizations reach, and anything enrolled under a subject reference nobody linked is invisible to our search while being perfectly visible to whoever holds it.

The operational consequence: state the reach. A partial search presented as a complete answer is the failure mode, and the module refuses an answer that does it.

Two categories answer differently from the rest:

  • Embeddings are ours, and are returned to nobody. A vector means nothing to the subject and everything to whoever else ends up with the export, so including one in an access response hands a biometric identifier to whoever the email reaches. They are deleted on request; they are not exported.
  • Released assets cannot be recalled, only accounted for. See §4.

Subject requests do not run through the creator support queue. The subject is not the customer, and a queue ordered by customer value orders these wrong.

4. Takedowns#

A takedown destroys what we hold: origin objects, derived renditions, evaluation frames — the frames extracted for review are the likeness too — and the face embeddings, which outlive the images they came from.

It does not reach provider-side copies, CDN edge caches, signed links already issued, or anything already downloaded (human-video-revocation.ts).

Never report a takedown complete while copies remain. Telling a subject their likeness has been removed when it is still served from three caches is worse than telling them nothing, because it ends the conversation. The reach report lists the unreachable locations by name, and a zero-retention provider class is the only case where the provider-side line is genuinely empty.

5. Incident response#

Six drills, and the clock on each starts before anybody notices (human-video-incident-drills.ts): a leaked URL is reachable from the moment it leaks, a takedown obligation starts at the subject's request rather than its triage, a revocation is in force from the instant it is made.

  • Containment is timed from onset, with the detection lag reported beside the response rather than folded into it. A drill measures response and cannot measure detection, because it begins by announcing itself.
  • Four of the six need a subject, and the only honest one is synthetic: the drill must not create the harm it rehearses.
  • Two cannot be undone, so they never point at production releases.
  • A drill expires on change to the path it covers, not on a calendar.

During an incident, remember which control is which: a quarantine follows automatically from a catalog diff and nobody signs it; a kill switch has a person's name on it and blocks submission only, leaving already-paid work to finish being evaluated.

Where this is still theory#

No drill has been run, no subject request has been answered, and no waiver has been issued. The lists above are enforced in code and untested by events.