# Task board polish audit — 2026-09-22

Scope: every open task on the Eve Task Board as the database held it at
10:30 UTC on 2026-09-22 — 25,035 tasks in 37 open families (13,436 `ready`,
9,165 `later`, 2,434 `parked`) out of 52 families and 138,654 tasks. The
question Elikem asked: is every open task granularly decomposed, well
specified, correct, and is the board as a whole polished. The board is the
authority (`authority-activated` 2026-09-20T23:21Z); the tracker files are
projections. No task was closed, reopened, rejected or parked by this pass.

Method. The whole open set was exported from PostgreSQL with body,
dependencies, links, section and file anchor, and eight mechanical passes ran
over it (section 1). The P0 families were then read item by item: every one
of the 275 gap-analysis tasks, the 117 open Eve SOTA tasks, the 260 open Isis
Chroma tasks, and Phase 0 plus one full chapter procedure and the Phase E/F
items of the presentation redesign. The 21 families with fewer than 45 open
tasks were read in full. The large trackers (workbenches ledger, phases,
V1 P2 backlog, audit remediation, V-series residue, Oya, V5, V2, Demeter,
open world) were checked mechanically and read by stratified sample. The
walkthrough coverage families (9,904 tasks) and the gates (1,299) were
checked mechanically only: their items are one-line state checks by design.

## 1. Mechanical passes

| Check                                                                 | Result                                                                                                                        |
| --------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
| Dependencies pointing at a task the board does not hold               | 0 of 9,047                                                                                                                    |
| Dependency cycles                                                     | 0 (the tool refuses them at `dep add`)                                                                                        |
| Duplicate ids inside a family                                         | 0                                                                                                                             |
| Open task whose file anchor is not an open checkbox                   | 0 of 24,728 anchored tasks                                                                                                    |
| Task body different from the tracker text at its anchor               | 8 — all in `lilith-safety-automation` (the file was rewritten 40 minutes after its last import); fixed, section 3            |
| Empty bodies                                                          | 0                                                                                                                             |
| Titles cut to a bare id (`P16.`, `S11d.`) or at an abbreviation       | 55 — 42 roll-ups of the V1 residual backlog, 4 Docs Center audit sections, Tara 14.7, 8 walkthrough items; fixed, section 3   |
| Titles cut at `e.g.`, `vs.` or `??` by the importer's sentence split  | about 60, almost all in the walkthrough coverage files; 38 completed from their bodies, the rest already read whole        |
| Parked task without a reason tag                                      | 0 (release 1,538 · external 266 · upstream 224 · governance 132 · human 114 · hardware 100 · corpus 57 · gpu 3)               |
| Lease rows left on tasks that were already shipped                    | 13, all expired on 2026-09-19/20 (closed by import, so `done` never released them); deleted                                  |
| Ready tasks whose body has no verification language                   | walkthrough and gate families aside: Eve SOTA 25 (owner/verifier items whose acceptance is in the dependency chain), Isis 25, others under 10 |
| Repository paths cited in bodies that do not exist                    | 359 distinct; on inspection nearly all are files the task itself creates (spec files named as deliverables); the ones cited as existing inputs were in the gap-analysis family and are fixed, section 2 |

`./eve check` answers ten questions clean before and after the pass. Its
sixteen standing warnings (tasks closed by import on 2026-09-19 with no
`done` or `verify`) are closed work and were not touched.

## 2. The gap-analysis family (P0, rank 1)

Created that morning: 66 epics decomposed into 209 sub-tasks. The
decomposition itself is sound — every sub-task names paths, a Do, a Done-when
with the spec to run and its dependencies; the 66 epics match the report's 66
items one for one (R01–R36, U01–U11, G01–G13, C01–C03, L01–L03); the
dependency edges in the database equal the `Depends on:` lines in every body;
every sub-task sits in its epic's section; each epic depends on its sink
sub-tasks. Every path and identifier the sub-tasks cite as an existing input
was then checked against the tree. Corrections, all applied through the
worker API with `edit-task`:

- 24 sub-tasks carried a `Parent context` block cut at 600 characters,
  mid-word. Rewritten with the epic's full `What` text.
- R08.4, R14.4, R35.2 named `tools/eve-reviewer/reviewer.mjs`, which does not
  exist; the reviewer lane is `tools/eve-reviewer/run.mjs` and its report shape
  `report.mjs`. A worker following the rules block would have parked on the
  first minute.
- R34.1 and R21.1 named `libs/oshun/eve-content-core`; the package
  `@oshun/eve-content-core` lives at `tools/eve-content` and `model.mjs` holds
  `DOCUMENT_KINDS`.
- G01.2 and R06.1 named a `buildProductGraphRuntime` export that does not
  exist; `composeProductGraph` (`src/compose.ts`) and
  `readProductGraphArtifact` (`src/artifact.ts`) do.
- R21.3 and R21.5 cited `publications.ts` and `media.ts` as if under the BFF;
  they are `apps/oshun/eve-content/src/publications.ts` and `media.ts`.
- R01.5 said the board check would gain "a ninth" question; it has ten.
- R11.1's parser file was not marked `new`.
- R33 was filed under LATER; section 12 of the report says NEXT. Moved (the
  section field has no API; `section_id` was set directly, with a journal
  note on R33 saying so).
- Tier is now carried by status as well as by section, because `./eve next`
  orders by position and the sub-tasks were appended in epic order, which put
  R13.1 (LATER) ahead of R25.1 (NEXT) and L01/L02 ahead of R01.1. The 62
  LATER-section tasks and the ten LATER-tier UI tasks (U08–U10 and their
  sub-tasks) are `later`; 203 remain actionable. `./eve next` now offers
  C01–C03 (atomic checks), then R01.1, R03.1, R04.1, R05.1, R29.1, R34.1,
  G01.1 — the report's suggested order.

Not changed: the bodies' `Owner`/`verifier` language in Eve SOTA and the
"the operator" wording in older trackers (epic U11 is the census that renames
them); the C items ahead of R01.1 (they are hour-sized checks with no
dependencies, which is what the position order says).

## 3. Other families

- **`lilith-safety-automation`** (9 `later`). SA.1–SA.8 held the bodies from
  before the 2026-09-19 rewrite while their titles were current, so SA.6 told a
  worker to mount routes that are mounted and SA.2 to fix a `this` bug that is
  fixed. Bodies refreshed from the tracker text (the only body drift on the
  board).
- **`audit-remediation`** (342 open). The 42 roll-up parents of the V1
  residual backlog were titled `P16.`, `T2.`, `Q1.` — the importer cut the
  title at the first full stop. Retitled from their source lines. Each
  roll-up "closes only when every child under it is closed" but declared no
  dependency, so `./eve next` offered a roll-up as work; each now depends on
  its open children (224 edges, all in by 15:55 UTC; the per-agent daily
  write allowance spread them over two credentials and several hours).
- **`small-trackers`**: S11, S11c, S11d, S14 of the Docs Center audit
  retitled. **`tara-content-workbench`**: 14.7's title was cut at an HTML
  comment; completed. **`walkthrough-customer`**: six titles cut at `III.`,
  `Vatic.`, `J.`, `conf.`, `w.`; completed from the bodies.
  **`audit-remediation`** H1.h and one V1 remaining-work item likewise.
  Thirty-eight more titles the importer cut at `e.g.`, `vs.`, `??` or an
  opening symbol (walkthrough coverage items, two Euterpe audit lines) were
  completed from their bodies where the body began with the title; the rest
  already read as complete phrases and were left.
- **Worker-filed follow-ups.** A remote worker filed eight follow-ups during
  the day (Isis T.01.19, T.03.08, T.16.07, T.09.07; one audit-remediation
  review-persistence item; one mobile and two Telegram walkthrough findings)
  and left them `later`, because the CLI help did not say that `add` defaults
  to `later` or that `edit --status ready` promotes a task. All eight are
  `ready` at Elikem's request. The mobile and Telegram walkthrough families
  were still marked `unstarted` while holding hundreds of ready items, which
  made the API refuse ready work in them; both are `open` now (a family-row
  change, which has no API). The help lines and the worker API guide were
  corrected (commit 83ea21fb912).
- **Expired claims**: the thirteen lease rows on shipped tasks were removed
  (the sweep only frees `leased` and `in-progress` tasks).
- **Eve SOTA** (117 open, `append` policy): read in full; owner, verifier,
  dependencies and acceptance on every item; the dependency graph consistent;
  the 18 parked items name their gate (13.7, 12.7, the human label sheet, the
  After Effects host). Nothing to change, and the policy refuses edits anyway.
- **Isis Chroma** (260 open): read in full; every item has a Verify clause and
  dated measurements. Nothing to change.
- **Presentation redesign** (1,491 open): 654 items are the eleven-step
  chapter procedure repeated per chapter, each step depending on its chapter's
  setup item; Phase 0's open items carry Verify clauses and dated dependency
  notes. Nothing to change.
- **Large trackers** (workbenches, phases, P2 backlog, residue, Oya, V5, V2,
  Demeter, open world): the 2026-09-18 re-scope pass left them with
  `Verify`/`Done when` clauses and today's paths; samples confirmed it.
  Nothing to change.
- **Walkthrough coverage and gate families**: one-line state checks by
  design; their families are typed `coverage` and `gate`. Nothing to change.

## 4. Counts after the pass

Read at 20:45 UTC; other sessions kept working the board through the day, so
these include their closures and parks, not only this pass.

| Family                   | Actionable | Later | Parked |
| ------------------------ | ---------: | ----: | -----: |
| eve-gap-analysis         |        193 |    72 |      3 |
| lilith-safety-automation |          0 |     9 |      0 |
| audit-remediation        |        296 |    12 |     37 |
| walkthrough-mobile       |        265 |     0 |      0 |
| walkthrough-telegram     |         44 |     0 |      0 |
| Board                    |     13,324 | 9,237 |  2,437 |

Every task write went through `./eve edit` and `./eve dep add` in API mode
with `--expected-version`, one task per call, under the `eve-codex-worker-01`
and `remote-worker-01` credentials. The direct database changes are named
above: three `section_id` values, thirteen expired claim rows, two family
status values and one family note. The
export, the check scripts and the per-operation log are in the session
scratchpad, not the repository.

## 5. Left for Elikem

Nothing blocks. One choice rather than a defect: whether the C01–C03 checks
should stay ahead of R01.1 in `./eve next` (position order says yes; they are
hour-sized and dependency-free).

The same day exposed the per-worker daily write cap (1,000) as the limit on
any board-wide pass; it is now configurable (`EVE_API_DAILY_MUTATIONS`,
10,000 on Dev) and goes live with the next Dev release.
