Audited 2026-09-11 against the repository, item by item. Forty-four of the 380
items in TODOS/phase-182.md are unchecked. They are not a backlog of similar
work: they fall into six groups with six different blockers, and only two of
the six can be advanced by writing code at all.
This document exists because the list does not say that. A session that opens the phase sees forty-four boxes and picks one, and most of them cannot be closed no matter how much effort is spent — while two groups can, and are easy to walk past.
Each verdict below names what was checked, not what was assumed.
Deferred by a release gate — 2 items#
182.C.13.02, 182.C.31.07 (Veritas provider consolidation).
Verified in apps/oshun/bff/src/middleware/release-scope.ts: "V1.0 ships four
rooms — Tara, Nyx, Arete, Nisaba. Veritas and Metis are deferred to V1.2." The
deferral is enforced at the BFF edge, not merely planned. Doing this work now
would consolidate providers for a room no member can reach.
Verdict: correctly deferred. Not a blocker on anything else.
Waiting on a consented corpus — 10 items#
182.C.17.08, 18.08, 19.08, 19.09, 20.01, 20.07, 33.01–33.05.
182.C.33.01–.05 are the corpus: consented, licensed footage of real people
across gender presentation, complexion, language (English, Ghanaian English,
Akan/Twi, Ga, Ewe), pose, motion and lighting. Nothing in a repository produces
that.
The other five are calibration and testing against it. 182.C.17.08 says so
in its own text — "fitting thresholds to invented ratings would be the
fabrication". 182.C.19.08 is explicit that it is "blocked on a labelled corpus
(182.C.33), not on hardware", which matters because the Linux box can run
the models; what it cannot do is invent the labels.
Verdict: blocked on consented subjects and real clips. The machinery is built; the boxes are deliberately unchecked.
Waiting on hosting and GPUs — 3 items#
182.C.32.03, 32.04, 32.06. Packaging self-hosted candidates with
pinned weights and GPU requirements, benchmarking full-frame performance, and a
secure RunPod/sovereign GPU deployment.
Verdict: blocked on hardware and a hosting decision.
One item that is a rule, not a task — 1 item#
182.C.38.08: "Update the capability matrix and public product copy only
after the promoted route has real production-like evidence."
It is a prohibition with a trigger. There is nothing to do until the evidence in §40 exists; doing it sooner is the failure it describes.
Verdict: gated on §40.
Final evidence — 13 items#
182.C.40.01–.12, and 182.C.36.06–.09 in part.
Every one is a demonstration against a running, production-configured system:
active-scene renders, identity continuity, rejection and reroute behaviour,
consent expiry, held-out quality results, archived test evidence. 182.C.40.12
closes only when all the others do.
182.C.36.06–.09 (Playwright web, Playwright admin, accessibility automation,
mobile automation) additionally need a surface to drive and a dev server.
Neither exists: see the two environment gaps below.
Verdict: gated on everything above, plus the surfaces below.
The two groups that can be advanced — 9 items#
Studio (182.C.23.02–.05, 23.13, 29.08) and mobile (182.C.24.02,
.03, .04, .06).
These are the only items where the blocker is that somebody has not written the
code yet. Both apps exist (apps/oshun/web, apps/oshun/mobile); neither has a
human-video surface.
As of this audit seven of the nine have had their controls or their policy built, in this order:
| Item | What exists | What is missing |
|---|---|---|
23.02–23.05 |
four Studio controls, browser-verified with axe | a route, project state, and anything that calls the BFF |
24.02 |
resume, recovery, playback and deep-link decisions | a binding to expo-file-system / expo-notifications |
24.03 |
the containment policy for all five leak channels | the one layer that calls it before touching an OS API |
24.04 |
the six accessibility rules and a checker for them | a screen for the checker to check |
23.13, 29.08 |
23.13 closed 2026-09-11; 29.08 has the primitives |
an admin human-video surface (29.08 only) |
24.06 |
— | a surface to automate |
Every one of these items nevertheless remains unchecked, because the item
text asks for a working surface and what exists is what a surface will be built
from. Per-item detail is in implementation-tracker.md.
The two that have not been started are the two that cannot be: 182.C.24.06
(mobile automation) needs screens to drive, and 182.C.29.08 needs an admin
human-video surface — apps/oshun/admin/src/app has no human-video area at all.
So the next piece of work on this phase is a route, in either app, that holds project state and feeds the controls that already exist. Not another control.
That route was built on 2026-09-11, and it found the next blocker#
/studio/human-video now renders the five regions from a derived project state
(projectState.ts, HumanVideoStudio.tsx, projectPort.ts). 182.C.23.13
closed with it — the states exist per region, browser-verified. 182.C.23.02–
.05 did not, and the reason is no longer "no surface":
The API lists projects and has no operation that returns one.
list-projects is the 23rd endpoint of the declared surface and answers
{ projectId, name, createdAt }. The authoring detail the screen is made of —
references and their consent, the script, the audio and its approvals, the
shots, the takes — has no read at all; it is spread across upload sessions,
identity packs, audio, jobs and candidates, and nothing aggregates them. So a
listed project can be shown and cannot be opened, and the route says exactly
that rather than assembling one out of five list calls and presenting the guess
as the project.
The consequence for the two acts that are records: adding a reference and approving audio are disabled with the reason attached, because a drop zone wired to nothing loses the file and a button wired to nothing looks like it worked. Everything else authors normally into a draft held in the tab, which says so.
The next piece was therefore a service, not a screen — and it was built the
same day. POST/GET/PATCH /projects keep the authoring work across the days
this flow waits for a subject; GET /audio/:audioId makes approve-audio
reachable from a page that did not submit the take (its If-Match tag had
appeared in exactly one response, ever); human_video_project_draft and
human_video_idempotency are real tables with a compare-and-set write; and
server.ts now mounts the surface, fail-closed, with a boot line naming every
port that will answer 503.
What remains for 23.02–.05 is the rest of the adapters — identity packs (for
consent and reference quality), audio takes, candidates — and an ingestion path
for reference images. The surface already renders their absence honestly rather
than as emptiness, so each adapter closes a named gap rather than changing the
design.
The last link is a scope nobody is issued#
Measured 2026-09-11 against a complete local stack — BFF booted with the draft
store bound to Postgres, web dev server, real member session. The Studio surface
calls GET /v1/human-video/projects and gets 401, and the surface says why:
"this session does not carry the scope the surface requires".
buildAuthSession in customer-auth-store.ts issues exactly one scope to every
customer session: ['domain:*']. The human-video surface declares
hv:projects:read|write, hv:audio:read|write|approve, hv:jobs:*,
hv:candidates:*, hv:assets:*, hv:evaluations:read, hv:release:write,
hv:consent:write — and nothing in the repository issues any of them.
So even a fully-adapted deployment answers 401 to its own Studio surface. Which
role or entitlement grants which human-video scopes is a product decision rather
than a wiring detail — hv:consent:write and hv:assets:delete are not things
to hand out with a login — and it is the last link in the chain the surface
needs.
Three measurements taken while looking for the next adapter#
Each one is a reason the next vertical is bigger than it looks, found by reading the code rather than by estimating.
1. The upload vertical cannot be made durable by adding a store. The obvious
next adapter is UploadSessionStore, and two things in
human-video-upload-session.ts stop it from meaning anything on its own. The
running SHA-256 lives in a module-level Map (line 309, "kept beside the
session so chunks stream"), so a resumed upload after a restart cannot finish
its digest — the session row would survive and the hash would not. And the chunk
path has no byte sink at all: appendUploadChunk sniffs, counts and hashes,
and the bytes are discarded. So resumable ingestion needs object storage and a
resumable-digest strategy first; a durable session store shipped before them
would look like working resumption and break at completion.
2. Consent has no read anywhere on the surface. GET /identity-packs/:id
returns detection and mapping state — that is quality, not permission. The
authorization itself can be withdrawn
(POST /authorizations/:id/withdrawal) and never read, and
human-video-likeness-authorization.ts is explicit that the enrolment record
lives in @aja/consent-management rather than here. So the Cast region's
consent indicator (182.C.23.02 names it) cannot be true until either an
authorization read exists on this surface or the Studio reads the consent domain
directly. Until then the surface says not read here, which is the honest form
of the same thing.
3. The identity-pack adapter is dead on its own. create-identity-pack
needs uploads and a face detector as well as the pack store, so binding only
the store yields a read over rows nothing can create. The three verticals are
therefore ordered by force rather than by preference: storage → uploads →
detector → packs.
The environment gaps, re-measured and mostly closed#
This section used to name two blockers as missing workspace dependencies. Both were re-measured on 2026-09-11 by starting the servers rather than reasoning about them, and most of what it said was wrong.
1. next dev could not serve any page — three causes, one an install.
apps/oshun/web/tsconfig.jsonredeclarespaths, which REPLACES the base config's rather than extending it, and was missing@oshun/resilienceand@oshun/embodiment-psyche, both reached transitively fromShellLayout. Every route under the shell 500ed on modules that exist in this repository. Fixed.@oshun/v10-rail-channel-ori-presencehad no mapping anywhere and@oshun/v10-rail-threadhad no subpath mapping although the BFF imports@oshun/v10-rail-thread/program. Fixed intsconfig.base.json.html2canvas-prois declared and was not installed. Apnpm installprovides it.
2. The BFF could not boot, and the mobile suite could not run — one install.
pnpm install --frozen-lockfile refused because the committed lockfile was out
of date with apps/oshun/bff/package.json: phase-182's own recent commits added
fast-check and prettier without updating it. Regenerating the lockfile is a
15-line diff; the install then takes about two minutes from a warm store. After
it: the BFF boots, apps/oshun/mobile's jest suite runs (60 human-video tests
pass), the Prisma client regenerates, and the migrations apply to the local
Postgres.
3. What a complete local stack then showed. With the draft store bound to
Postgres, a project created through the API lands as a row whose cast_ids
projection is written from the document; a stale save answers
409 hv.validation.draft_conflict; a draft carrying a consent field is
refused with the sentence that says why.
apps/oshun/web/e2e/human-video-studio.spec.ts passes 7 of its 8 tests against
that stack, and the eighth skips itself for the reason below.
4. The one that is still open: the surface's scopes are issued to nobody.
See the section above. A member session carries ['domain:*'], the surface
requires hv:projects:read, and the 401 is what the eighth test skips on.
The honest summary#
| Group | Items | Can code move it? |
|---|---|---|
| Release-gated (Veritas) | 2 | no — V1.2 |
| Corpus | 10 | no — consented subjects |
| Hosting / GPU | 3 | no — hardware |
| Capability-matrix rule | 1 | no — gated on §40 |
| Final evidence | 13 | no — gated on all the above |
| Studio and mobile surfaces | 9 | yes — 1 closed, 7 advanced, 1 needs an admin surface |
| Playwright / automation | 4 | partly — needs the surfaces and a dev server |
Nine of forty-four. That is the number worth knowing before planning a session on this phase.