Eve library

The erasure port, bound and driven live — 2026-09-12

Guide

Revision 1

On this page

Evidence for 182.C.16.09 (deletion propagation) and the two halves of 182.C.40.06 that were still missing: SUBJECT DELETION and TAKEDOWN. Four responses from a BFF this session started against the local Postgres and a real versioned MinIO bucket, and the boot line they came from.

This morning's capture (../2026-09-12-eleven-ports-bound/) ended with two ports answering 503, and said what each needed: "erasure needs something that can actually delete from an object store and report, per locator, whether the target is gone." That now exists and this is it running.

Method#

apps/oshun/bff/src/server.ts under tsx on :4010 with OSHUN_V1_DATABASE_URL against oshun_dev, and OSHUN_HUMAN_VIDEO_OBJECT_* pointed at a MinIO on :9000 with versioning enabled — versioning matters, see below. Secrets were sourced through a throwaway /tmp script, deleted at the end; none is in this directory, and the captures were grepped for key material before committing. The server was torn down and :4010 confirmed free. Tokens were minted with createDevAuthToken.

Two jobs, one authorization and two takes were seeded directly into Postgres, because an identity pack needs a face detector this deployment does not have configured and a pack is required to submit a job. Six objects were placed in the bucket by hand for the same reason: nothing renders here yet. Everything after the seed is the product's own code path.

The boot line#

boot-line.txt, verbatim. erasure is bound; releases is the last port with no adapter:

text
human-video api: 32 endpoints live; bound: preset catalogue, project drafts,
idempotency records, upload sessions, jobs, candidates, reviews (queue derived
from candidates), project listing (a project is a draft), revocation reach (jobs
and authoring drafts), erasure (takes and their derived objects; bucket
oshun-human-video-test at http://127.0.0.1:9000; no edge, no URL minting, no
search index); answering 503 until bound: identity packs (no face detector: set
OSHUN_HUMAN_VIDEO_PYTHON, _FACE_GEOMETRY_SCRIPT and _YUNET_MODEL), audio takes
(no probe: set OSHUN_HUMAN_VIDEO_FFPROBE), releases

The parenthesis on erasure is not decoration. It names the bucket, and it names the three things this deployment DECLARED it does not have — no delivery edge, nothing that mints signed URLs, no search index. Those declarations travel into every outcome's detail, so an operator who does have a CDN can see that this deployment said it had none.

What was seeded, and why each piece is there#

object why
candidate/…/hvj-ev1/<TAKE1>.mp4 take 1's master, recorded in the candidate row
candidate/…/hvj-ev1/<TAKE2>.mp4 take 2's master — must survive take 1's delete
intermediate/…/hvj-ev1/<TAKE1>/sync-proxy.mp4 a proxy NO table records: only a sweep finds it
intermediate/…/hvj-ev1/<TAKE2>/sync-proxy.mp4 take 2's proxy — must survive take 1's delete
quarantine/…/hvj-ev1/a3-c0 an orphaned attempt of a job that DID render
quarantine/…/hvj-ev2/a1-c0 an orphaned attempt of a job that rendered nothing

hvj-ev2 has no candidate row at all. That is the case the last row exists for.

The four captures#

file request status
delete-asset-200.json DELETE /assets/hvc-ev1 200, manifest, takesCovered: 1
delete-asset-replay-200.json the same request again 200, byte-identical but replayed: true
delete-asset-404.json DELETE /assets/hvc-does-not-exist 404 hv.validation.unknown_resource
withdraw-consent-200.json POST /authorizations/auth-ev/withdrawal 200, manifest, takesCovered: 2

What the single-take deletion actually did#

Four locators: the signing epoch, the edge path, take 1's swept proxy, take 1's master. Afterwards the bucket held four of the six objects, and the two that were gone were exactly take 1's:

text
human-video-candidate/tenant-erase-ev/hvj-ev1/<TAKE2>.mp4
human-video-intermediate/tenant-erase-ev/hvj-ev1/<TAKE2>/sync-proxy.mp4
human-video-quarantine/tenant-erase-ev/hvj-ev1/a3-c0
human-video-quarantine/tenant-erase-ev/hvj-ev2/a1-c0

Two properties are being demonstrated there and both were design decisions.

The sweep is attributed by DIGEST, not by job. Take 2's proxy sits under the same job prefix as take 1's and was not touched, because intermediate and released keys are addressed by the digest of the object they came FROM. A job-wide sweep would have deleted a take the creator did not ask about.

A single-take delete leaves the job's quarantine alone. A quarantine object belongs to an ATTEMPT of the job rather than to any one take, and the job's other takes are still there. It is the withdrawal that covers it.

And mc ls --versions for take 1's master returns NOTHING. On a versioned bucket an ordinary DeleteObject writes a delete marker and leaves every byte version behind it — 404 to every reader, fully present to anyone who can list versions. That state is the one this whole adapter exists to avoid, and the versioned bucket is what makes its absence visible.

What the withdrawal added#

Six locators, takesCovered: 2, and one more object destroyed than the takes account for: quarantine/…/hvj-ev1/a3-c0, the orphaned attempt the single-take delete had deliberately left. Take 1 was covered again and its master reported absent rather than missing — an erasure that died halfway has to be able to finish, so "nothing was stored under this key" is a real outcome and it says that no bytes were destroyed by that call.

After it the bucket held ONE object: quarantine/…/hvj-ev2/a1-c0.

The one survivor is the one the manifest declares#

hvj-ev2 rendered nothing, so there is no take record to attribute its half-downloaded bytes to, so no plan can carry them. The manifest says so rather than going quiet:

text
"unreachable": ["provider-side", "unrendered-attempt-bytes"]
"complete": false

with the reason in the detail — "quarantine bytes from an attempt that produced no take: with no take record to attribute them to, this deletion cannot enumerate them. The quarantine namespace retention bounds them to 24 hours — and a bound is not a retraction."

That sentence is the difference between this and a deletion that reports success. A withdrawal silent about it would have told somebody their likeness was gone with a partially downloaded face still on disk.

provider-side is unreachable for a reason that is sourced rather than assumed: OpenRouter video generation is not zero-data-retention eligible at all, and no lane in this build exposes a deletion endpoint, so the manifest carries a provider-retention-disclosure obligation instead of a claim.

The replay, and why it is not a 404#

The second DELETE returned the first manifest — same destroyedAt, same locators, every field identical except replayed, which went from false to true. replayed is not a column: it is a fact about the CALL, and a stored false read back a week later would report a fresh destruction that happened days ago.

A 404 on the retry would make a successful deletion look like a failure, and somebody who reads it as a failure either stops retrying or escalates a deletion that already happened.

The ledger rows#

text
 subject_kind  | subject_id | takes_covered | complete | locators |               unreachable
---------------+------------+---------------+----------+----------+------------------------------------------
 asset         | hvc-ev1    |             1 | f        |        4 | {provider-side}
 authorization | auth-ev    |             2 | f        |        6 | {provider-side,unrendered-attempt-bytes}

The table refuses a row that claims complete while anything is unreachable. Verified directly against this database before the run:

text
ERROR:  new row for relation "human_video_destruction_manifest" violates check
constraint "human_video_destruction_manifest_never_overclaims"

The second pass: the bytes the renders were made FROM#

The first pass above left one gap named rather than fixed — a withdrawal destroyed every render and left the photographs the creator uploaded, which live in a SECOND store under a different key scheme (the upload staging root, keyed by a tenant hash and a session uuid, which parseObjectKey refuses and the namespace sweep cannot see). A subject deletion that leaves the source photographs is not one, so the same session closed it.

boot-line-sources.txt is the boot line after that change. The parenthesis on erasure grew:

text
erasure (takes, their derived objects, and the uploaded references and audio;
bucket oshun-human-video-test at http://127.0.0.1:9000; no edge, no URL minting,
no search index)

A whole withdrawal now reaches all three stores#

withdraw-whole-sources-200.json: one authorization, revoked whole, over a subject with one stored take, one reference photograph and one approved audio take. Five locators, and afterwards NOTHING of that subject's remained — the take object gone from MinIO, both staged files gone from the upload root:

text
tenant-src-ev/hvj-src1
https://media.oshun.local/v1/human-video-released/tenant-src-ev/hvj-src1/*
human-video-candidate/tenant-src-ev/hvj-src1/<TAKE>.mp4
hv-upload/aa11bb22/session-photo-1
hv-upload/aa11bb22/session-voice-1

destroyed 2 location(s) — the origin object store and the uploaded source assets — with provider-side the only thing left unreachable, which is irreducible: the lane retains and exposes no deletion endpoint.

A second subject's photograph (session-photo-2, under a different authorization) was in the same directory throughout and survived, which is the cross-check that the source sweep is scoped to the subject rather than to the directory.

And a NEGATIVE CONTROL found a real defect#

A withdrawal of one project out of three must NOT delete the references, because the projects it did not name still render from them. That was already the behaviour — and the manifest it produced was wrong. withdraw-partial-BEFORE-THE-FIX-200.json, captured live:

text
"complete": true
"unreachable": []
"detail": "every copy of this likeness known to the system has been destroyed. …"

The photograph was still on disk. The location had been omitted from copiesPresent because this withdrawal was not going to destroy it, so the reach had nothing to be incomplete about — and the sentence the subject would be shown was the exact overclaim assertNotOverclaimed exists to prevent, arriving by a route that never went through it.

The fix separates "what exists" from "what this request destroys". takedownReach now takes a retained list, whose members are reported as still there with their own reason rather than as destroyed or as unreachable-because- unbound. withdraw-partial-retained-200.json, same request, after:

text
"complete": false
"unreachable": ["uploaded-source-assets"]
"detail": "destroyed 0 location(s); uploaded-source-assets (deliberately kept:
this withdrawal was narrower than the whole permission, and the projects it did
not name still render from this). …"

This is the argument for a negative control. Everything about the positive case was right, in the code and in three passing tests, and the manifest for the case where the code correctly did nothing was the one that lied.

The third pass: thirteen of fourteen ports, and the last two live probes#

The two passes above left 182.C.40.06 with four of its six demonstrated live and two (consent EXPIRY and tenant isolation) resting on unit rules. The whole stack turned out to be configurable on this box — the YuNet model, the face geometry script, ffprobe, ffmpeg and the forced aligner are all present — so it was configured and the last two were driven live as well.

boot-line-thirteen.txt is the result, and it is the shortest 503 list this phase has produced:

text
human-video api: 32 endpoints live; bound: preset catalogue, project drafts,
idempotency records, upload sessions, identity packs, audio takes, jobs,
candidates, reviews (queue derived from candidates), project listing (a project
is a draft), revocation reach (jobs and authoring drafts), erasure (takes, their
derived objects, and the uploaded references and audio; bucket
oshun-human-video-test at http://127.0.0.1:9000; no edge, no URL minting, no
search index); answering 503 until bound: releases

submit-403-consent-expired.json. A permission that LAPSED rather than being revoked — nobody took it back, it ran out — with a real pack and a real approved audio take behind it:

text
"code": "hv.consent.expired"
"message": "The permission to depict this person has expired and needs renewing."
"terminal": true, "blame": "person"
"detail": "… auth-lapsed: 2026-09-12T09:50:09.005Z is outside
           2026-07-14T09:49:50.879Z to 2026-09-11T09:49:50.879Z"

Expired and scope-exceeded are deliberately different codes: one is renewed and the other is renegotiated, and the same code for both tells a creator to do the wrong thing.

Tenant isolation#

cross-tenant-404.json. One tenant asking for another tenant's job by its real id gets the same answer it would get for an id that does not exist:

text
"code": "hv.validation.unknown_resource"
"detail": "no job with that identifier is reachable from this workspace — the
           same answer is given whether it does not exist or belongs to
           somebody else, because saying which confirms the identifier"

and its own listing came back with zero jobs, which is the other half of the same statement.

What this does and does not close#

Closed: 182.C.40.06's subject deletion and takedown, which were the two of its six that were not demonstrated. Both now run against a real object store and a real database, and both report what they could not reach.

Closed by the second pass: the uploaded reference photographs and approved audio, which the first pass named as a gap.

182.C.40.06 is CHECKED on the strength of this. Its six: consent expiry (live, above), consent revocation (live, here and in ../2026-09-12-jobs-adapter-live/), tenant isolation (live, above), provider retention disclosure (live, as an obligation in every manifest here), subject deletion (live, both stores, verified gone), and takedown (live). The sixth — no public candidate URL before release — is a namespace rule rather than a capture: canRead(namespace, 'delivery') is true for released and false for every other namespace, pinned by a test that iterates all six, and no route on this deployment mints a delivery URL at all.

One store a subject deletion would have to grow to reach is NAMED AND GUARDED rather than assumed irrelevant: human-video-embedding-store.ts holds encrypted face embeddings, which COPY_REACHABILITY calls the copy that matters most because an embedding outlives the image it came from. This erasure does not touch them, which is correct only while nothing writes one — so a test asserts the store has no importer, with a control proving the pattern can fire, and says what has to happen on the day it does.

Not closed: complete: true is unreachable on any lane in this build and that is not a defect — it is provider-side being honest, and the item asks for the retention disclosure to be demonstrated rather than for it to be absent. 182.C.40.04 still needs the releases port and a real render. And every take here was SEEDED rather than rendered: an identity pack needs a face detector this deployment does not have configured, so the deletion path is proven against real bytes in real stores while the bytes themselves were placed by hand.