Task board specification audit — 2026-09-18

Scope: every open item the task board (TODOS/BOARD.md) counts as work — 13,627 tracker items in 71 files, plus the coverage walks (10,930) and launch gates (1,299) at family level. The question was not "what is done" but "is what is left well specified, well designed, correctly ordered, free of duplicates and free of contradictions". No box was checked or unchecked by this audit.

Method. Every checkbox was extracted with its heading path and body using the board's own scanning rules (the extract reproduces the board's 13,627 exactly). Mechanical passes looked for duplicate ids, parent/child state conflicts, exact and near-duplicate text (TF-IDF cosine at 0.72, within and across files), declared dependencies pointing at open or missing items, and blocker notes that name a machine. The P0 and P1 trackers were then read: every open item of Isis Chroma (242), Eve SOTA (111) and phase 182 (30); the structure, rules, decision sections, dependency map and a stratified sample of the workbenches ledger (8,446); and the reconciliation rules of the two presentation trackers. The P2 and P3 trackers (about 2,300 items in 66 files) were read by one delegated read-only pass whose findings were spot-verified before being repeated here.

1. What the mechanical passes found#

Lexical duplication across trackers is essentially absent. The duplication that exists is semantic — the same system planned in two trackers in different words — and it is concentrated in the three P0 trackers (section 3).

2. Spec quality by tracker#

Isis Chroma (P0, 242 open). The best-specified tracker in the repository: every item has a Verify clause, measured facts are dated, spend is gated, and it carries its own ordering sections. Defects are at the edges: 14 items were blocked in prose but counted as actionable (tagged in this pass); T.10.07 carried a Mac-side blocker that does not hold on the Linux server (corrected); T.17.02 still required a benchmark against Meshy that the checked item T.01.15 removed by owner decision (reconciled). Five items are "keep open unless the owner asks" (A.03.05, E.04.05, E.06.04, E.01.07, L.07.01) and will read as pending work for ever; they are tagged now, but a "listed, not planned" section without boxes would be more honest. The tracker's title still says "MVP" over 586 items and six tracks; the board cannot rank a film studio (F, 119 items, opened today, none started) separately from the two CyberRealistic stills that close the arsenal track.

Eve SOTA (P0, 111 open). Owner, verifier and dependencies on almost every item, and a dependency graph that is internally consistent. Two structural facts matter more than any single item. First, the 2026-09-05 "no closure with blockers" rule means the initiative cannot close without After Effects, Unreal, declared console/VR devices, an external channel and human raters (5.10–5.29, 6.6–6.8, 12.3, 13.5, 17.11–17.20): that is a legitimate bar, but none of those items says which machine or licence is expected, so they are neither actionable nor tagged. Second, Eve's DCC, generation and 3D items (5.x, 15.7, 17.4, 17.22) are written against Bellona's Blender agent, "Isis Flux/SD3.x adapters" and "Bellona Meshy/Tripo", while the Isis tracker has since made the live lanes RunPod Chroma/Z-Image/Wan/LTX, an OpenRouter hosted lane, a Meshy lane under Isis, and its own Blender bridge (section 3).

V1 domain workbenches (P0, 8,446 open — 62% of all open work). Structurally clean: its own lint reports 11,920 cells and 0 violations. Three problems.

1. Its own sequencing rule has been broken. Rule 1 of §4 is "complete BASE, STD and Phase S architecture decisions before new workbench packages or routes". Today BASE is 21% done, the UX and standards contract (§2.3–2.4) is 1 of 163, and the capability contract (§3) is 0 of 565 — while Phase S is 97% done, Metis 89% and Isis I1–I5 about 95%. Inside Phase I the decision section I0 has 243 open cells, including the threat model (I0.10), data classification (I0.11) and emergency controls (I0.12), while the safety phase I5 it was meant to inform is 114 of 114.
2. The contract sections are acceptance criteria, not tasks, and nothing links them to the tasks. §3's 565 CAP cells each ask for a "per-domain capability-matrix row"; they can only be checked when all seven domains finish, and DONE.3 restates them as one gate. Outside their own sections the ledger references a CAP, UX or STD id twice in 90,000 lines. So 727 boxes sit at the top of the P0 queue that no phase task is traceable to.
3. The unstarted phases are not decomposed, only enumerated. Measured on open items: in Phase Y 54% and in Phase E 51% of boxes pack six or more slash-separated behaviours into one checkbox ("create/assign/status/ priority/estimate/duration/dates/owner/crew/resource/budget/asset/gate/ version and bulk operations. Evidence: task tests."); 0% name a file, package or route; the evidence clause averages 2.4 words. The checked phases (S, M) average 45–200 words of evidence per cell. About 6,300 cells in phases Y, V, E, A and B are in this state. They should not be decomposed up front; each phase's first task should be to ground its cells in the code the N0 inventory finds, and the ledger should say so.

Priority question for the owner, not a defect: release-scope.ts ships V1.0 with Tara, Nyx, Arete and Nisaba, V1.1 adds native apps, V1.2 restores Metis and Veritas. This ledger covers none of Nyx, Arete or Nisaba; 1,612 of its open cells are for the two V1.2 rooms, 1,444 for Euterpe (which its own §1 says is not a ratified V1 domain), and 3,411 for operator suites (Yemaya, Aja, Bellona). Its wave order also puts Metis and Veritas third and Isis fourth, while execution did Isis before Veritas. Meanwhile the V1.0 customer-surface walk is 51% and ranked P2?.

Presentation center pair (P1?, 1,967 open). Reconciled with each other in prose (the redesign file governs; section 10 of the older file stays as review criteria), but three open boxes in the older file still told an agent to build a design pilot that section 0 forbids rebuilding (reconciled in this pass). Section 10's 134 boxes are criteria rather than tasks, and almost no item in either file has an id. The how-to binds narration to a Mac path and --device mps; CLAUDE.md says Kokoro runs on the Linux server's CPU, and the file gives no recipe for it.

Phase 182 (P1?, 30 open). Every open item carries a dated note naming its blocker. The phase is effectively waiting on four things, none of them code: a consented corpus (15 items), a GPU host (3), the V1.2 train (2), and two ports plus an object store (4). It should read as blocked at family level, with the corpus as an explicit owner decision.

3. The same system planned in more than one tracker#

None of these pairs names the other file. Each is a place where two agents working two trackers would build the same thing twice, or differently.

4. P2 and P3 trackers#

Every open item in the 66 files was read. The headline is that about 1,270 of their roughly 2,300 open items are blocked in their own prose — an account, a licence, a device, a corpus, a release event, a person — and carry no blocked: tag, so the board reports them as actionable. Phase 27 alone is 427 of those, V2 83, phase 29 79, the V1–V9 ledger 63, phase 137 60. A further 550 or so in phase 27 and V5 wait on assets or builds that no item asks for.

Smaller trackers and the V1 lineage#

• A safety service is ranked P3 under a title that says it is finished. apps/lilith/svc-safety-automation/IMPLEMENTATION_COMPLETE.md reads "COMPLETE", "Production-Ready", "111/122 tests passing", and its three open boxes are: mount the API endpoints in app.ts, add end-to-end tests, and "Fix 11 bugs in original safety-automation service". It belongs in a Lilith tracker at a higher priority, under an honest file name.
• Blockers hidden where nothing can see them. Phases 78 and 161 and V8 keep their reasons inside HTML comments, which neither the rendered page nor the board shows; phase 180, the ambient rail and the Yemaya checklist use private words (_Blocked:_, GATED:, EXT) instead of the blocked: tags.
• Four dogfood gates depend on the phase 27 crew (phase 161 L263, 163 L250, 172 L548, 174 L635). If phase 27 is reclassified they can never close and should be reworded or removed with it.
• One missing corpus, three trackers. V8 4.6, 9.5 and 12.5, V5 L1232 and the V1–V9 ledger's V5-002 all wait on the authored V5 cases; none names the others.
• A criterion that cannot be met. Phase 125 125.11.3.2 asks for more than 10,000 Monte Carlo sweeps a second on a 32,000-atom cell on one core, about three nanoseconds per energy evaluation.
• An item that tells an agent to break a rule. Tara 14.6 asks for a hand edit of CONTENT_GENERATION_CAPABILITY_MATRIX.md, whose header says it is generated; the task is to change the registry and the generator.
• Orphaned handoffs. ISIS_GAPS_TODOS.md P2-005, P3-003 and P3-004 are "deferred to the deployment ticket" and "the frontend ticket"; no such ticket exists in any tracker.
• Two V1 lineages that never meet. docs/releases/v1/specs/todos.md (re-verified by the P2 backlog, Parts IX–X) and V1/TODOS.md (re-audited by the residual backlog, which it names operative). Quotas and budgets appear in four trackers in different states (P2 §51 open, residual T4 open, V1/TODOS.md §18.6 checked, V1–V9 COST-001 checked); i18n and observability are "Confirm …" items in the P2 backlog while the residual backlog found the layers absent; mobile store readiness is tracked five times and is V1.1 scope by the 2026-08-05 decision.
• Frontier-scale training as checkboxes. Phases 176, 97, 175, 177 and 181 hold about fifty "Train …" items (an 11B-parameter dynamics model, 1.2 billion clips, a TPU or H100 pod) with no compute, data-licence or owner decision. Where a box bundles "implement training and serving", the code half is actionable and should be split off.

5. Open items that contradict a recorded rule or decision#

6. Decisions only the owner can take, ranked by how much they release#

1. What P0 means for the workbenches ledger. It is 62% of all open work and serves no V1.0 room. Either confirm that (operator and V1.2 authoring first) or re-rank its phases against the V1.0 walk, and say whether Nyx, Arete and Nisaba need authoring phases at all. Also confirm that sections 2–3 close last, so an agent stops meeting 727 contract cells at the top of the queue.
2. Isis T.01.18: do training-data terms bind 3D weights? The current rule refuses TRELLIS.2, TRELLIS, SkinTokens, Step1X-3D and Paint3D. Milestone M1 (image to a textured GLB) cannot be met on self-hosted models until this is answered; T.02 (workers, 0 of 6) is idle behind it.
3. Which domain owns agent control of Blender, Unreal and Unity — Bellona (Eve, workbenches Phase B) or Isis (T.19–T.21, F.02). ADR-0008 Q1–Q9 have been waiting since 2026-09-13.
4. Fill the 42-transcript label sheet. One sitting closes SMX 7.2 and 8.1 and Eve 12.3, and lets the judge leg be pinned.
5. Choose and credential one external channel for Eve (7.8). It is the single lane 13.5 has not run, and watcher delivery (7.4, 7.5) needs it.
6. Phase 182's consented corpus: commission it or park the phase. Fifteen of thirty open items wait on it and nothing else.
7. Film track F.00.03 (a)–(f): spend cap, Codex allowance, pilot brief, Kitsu or not, working colour space, HDR or not. 119 items, none started; the suggested order needs them by step 2.
8. Reclassify what is not software work. Phase 27 (934), V5's content bible (about 100), V2's re-opened gates (83) and the Neith residue (about 190) are counted as committed work. As proposals, inventories or gates they would leave the queue without losing a line.
9. Confirm Hetzner is final, so the hundred AWS items collapse into one platform deploy tracker.
10. Which commercial applications Eve's charter really requires (5.12): After Effects, Unreal, Unity, Houdini, Maya, each with a licence and a host, or a ratified alternative. Twenty Eve items wait on that inventory.
11. Standalone Tara and Veritas store apps, or apps/oshun/mobile? 697 gate boxes describe the former; V1/TODOS.md names the latter as customer mobile.
12. Spend: Isis E.00.01's 8–15 USD of clips, H.06.02's 0.73 USD step against a 0.50 approval, and the default-shape LTX-2.5 clip.

7. Changed by this pass#

Commits d035631839b, eaa5c09bfad, a55a2267af7, af476bded47 and the one carrying this report, all on origin/main. No box was checked or unchecked.

• Board scanner: a prose line starting with three backticks was read as a code fence; and boxes marked [~], [/] or [!] were in no count at all. The board now lists them under "Boxes the board cannot read" (384 in work files, 58 of them in three trackers the registry calls closed). Spec 32 passing.
• Isis Chroma: 14 blocker tags, each from the item's own body; T.10.07 and T.17.02 corrected; T.01.18 now lists what waits on it.
• Eve SOTA: 12.3 cross-referenced to the SMX sheet; 5.5, 11.7 and 13.5 corrected and tagged; the header's CLOSED claim corrected.
• SMX: three notes written as checkboxes folded into the two tasks they annotate.
• Workbenches: I0.20, a crosswalk against the Isis tracker before I6–I9 (structure lint 11,920 cells, 0 violations; dependency graph no drift).
• Presentation center: the three pilot boxes reconciled with the binding rule.
• Phase 29: the verbatim duplicate of 29.19.4.3 removed. Phase 71: the retired Meshy benchmark item tagged.
• Registry: the Antigravity plan superseded by the walkthrough tree; measured notes on the workbenches and redesign rows; the SMX row corrected; this report registered as a plan.
• Board after the pass: tracker open 13,628 (13,627 before: three note boxes and one duplicate out, five cells in for I0.20), tagged blocked 39 (was 23); coverage open 10,208 (was 10,930).

8. Suggested order for the follow-up#

1. Put the decisions of section 6 to the owner; items 1, 2, 3 and 8 change what the board ranks.
2. Apply the vocabulary fixes that need no decision: convert [~] to [ ] plus a tag in V1/TODOS.md and V2_TODOS.md; turn rules, gates and rejected alternatives that are written as boxes back into prose (V1–V9 §0.3 and §13, remaining-work §2.4, ambient rail's doubled human gates); renumber phase 27 §27.27.
3. Tag, one item at a time, the trackers whose blockers are already in prose: phase 28 (30 of 30), phase 19 (21 of 21), phase 137 (60 of 60), V2 (83 of 83), ambient rail (about 37 of 41). That alone moves roughly 230 items out of "Actionable".
4. Do workbenches I0.20, then close I0's open decisions (I0.8–I0.13) before any further Phase I cell.
5. Add the rule to the workbenches ledger that a phase's first task grounds its cells in code paths found by its N0 inventory, and split compound cells then, not before.
6. Rewrite Eve 15.7, 17.4 and 17.22 against the lanes the Isis tracker shipped.

9. The overhaul that followed (same day)#

The owner then asked for the recommendations to be carried out, delegated the open decisions ("make the decision based on what you think is best"), and set the order of work. What changed, by where it now lives:

• The board is the Eve Task Board, and it is built for an agent to navigate. ./eve next [family] [n] prints the next actionable items as path:line text, live; TODOS/BOARD.md opens with how to use it and a "Start here" block per top family; TODOS/PARKED.md lists every item no agent can do, by reason, for the owner; TODOS/board.json carries the same ranking for scripts. The registry gained rank, focus, queue and deferred, and open work now splits into Actionable, Later (sections held back, each with its reason) and Parked (tagged blocked: or a family the registry marks blocked).
• Order of work (owner): 1. the Eve Task Board's own SQLite store, CLI and knowledge index (TODOS/initiatives/EVE_TASK_BOARD_TODOS_2026-09-18.md, 51 tasks); 2. the Eve V1 presentation redesign, with Phase 0 — integrate Eve into the Presentation Center first (76 tasks written against a code survey that overturned four premises of the old plan); 3. Isis Chroma; 4. Eve SOTA; 5. the V1 workbenches.
• Install before you park (owner): a missing tool is a step, not a blocker. The rule is in the board, CLAUDE.md, the checkbox rule file and both top trackers. The machine that executes tasks has Unreal 5.5, so no Unreal item is parked; Unity and a local model runtime are install tasks with one owner sign-in between them.
• Decisions of section 6, taken under the delegation and recorded where they bind: training-data terms bind delivery, not internal use (Isis T.01.18 is now engineering); one Blender operation library, with a Bellona audit (T.21.07) and the ADR's nine questions answered in the ADR (T.00.06); the six film-track choices of F.00.03; Telegram as Eve's first channel; phase 182's multi-speaker flag counts as the flagged alternative; no external CMS for room content; bulk V5 content comes through the generation lanes, not from checkboxes; phase 27 is a proposal; launch, store and compliance gates are parked until a date exists; the workbenches ledger defers its contract sections and its five unstarted phases and grounds a phase when it starts (rules 12 and 13 of its section 2.5).
• Decisions 4, 5 and 11 stay the owner's: the 42-transcript label sheet, the Telegram bot token, and which mobile binary ships to the stores.
• Missing prerequisites now exist as tasks: a playable V2 vertical slice (V2.VS.1–7) that 83 re-opened gates waited for; the three simulations the ambient rail needs (RC.2.P1, RC.3.P1, RC.4.P1); Demeter's web pages wired to their API (29.22, nineteen pages rendering placeholder constants); a jest-expo suite that runs; GPS contracts by domain (GPS.M1.a–j); the safety automation service's eight real tasks (SA.1–8), raised from P3 to P1.
• The third checkbox mark is gone. 339 items in twelve trackers carried [~], which the board cannot read, so none was in any count. Each was read and is now an open box that names its blocker, or an open box with no tag because an agent can do it, or no box at all. The notes that had justified the mark were often stale by September: "no engine on this box" (the executing machine has Unreal 5.5), "no provider credentials" (an OpenRouter key exists), "deploy-bound" (the dev stack has the object store and the mail sink the item was waiting for). V7's 51 turned out to wait on eight pieces of ground nobody had listed (a session transport, a test client, a durable store, a packaged resource, a process stack, an Unreal connection, a local cluster, a live identity binding); those are now its section 0, V7.RR.1–8.
• Item-by-item passes, by tracker. Read against the code and left in one of three states (specified for an agent with a verb, a path that exists and a Verify; tagged with its reason; or no longer a box because it never was a task): the 31 phase files under TODOS/, phase 182, the Isis Chroma tracker, both Eve trackers, the small-model tracker, V1 to V10's trackers, the eight audit checklists under docs/audits/, the older Presentation Center tracker, and the small trackers. The V1 residual backlog, the P2 backlog's actionable tail, the Oya checklist, the V1–V9 content plan and V5's 33 engine items were in progress when this section was written; section 10 records how they ended. The redesign tracker's chapter items and the workbenches ledger were sampled, not rewritten: the first lists every slide with its layout and source, and the second is held to one cell format by its own structure lint (11,920 cells, 0 violations).
• The older Presentation Center tracker had two problems of its own. Its section 6 was a coverage charter of about 110 boxes, each the size of a track; its tasks are now 32 track plans (section 6A) that turn the charter into chapters in the redesign tracker, and the charter stays as the acceptance list. Its section 12 still assumed the BFF would serve the centers on new ports with Eve docked inside the decks; nine items were rewritten to the host decision (the admin app serves both centers, no port is added, Eve beside a deck is the admin chat).
• What this pass broke, and how it was found. Editing a V-series tracker is not a documentation change: content-coverage/seed-registry.v1.json is generated from those trackers and drift-checked in CI, about 1,300 check scripts pin exact checkbox lines, and some 70 of them ask whether a section still has an open box. The morning's edits had left the content-coverage job red and nobody had run it. The afternoon's conversion would have broken 27 pinned lines and 7 section checks in V2, 14 pins and a sign-off check in V4, 8 pins and 9 section checks in V6, and V3's launch-readiness verifier. All are repaired: the generator and the section checks now read an open box that ends in a blocked tag as blocked, through one small reader per product with its own test, the pins name the new lines, and the frozen counts were re-reviewed with the reason written above them. Every gate that passed before passes now; the ones that fail (8 of 1,156 V2 checks, 8 of 118 V6 verifiers) fail on source files, as they did before. The lesson is in the agent memory and in the Eve Task Board plan: run a tracker's gates after editing it.
• The plan's own first defect. ./eve handed every call to the new CLI as soon as its first commit landed, which took ./eve next, board, check and files away from every agent for an afternoon, because the plan had never said they must keep working. The shim now falls back to the markdown board for those four while the CLI answers "not built", a spec pins that, and the plan says it.
• Who owns and who reviews. The workbenches ledger asks for owners, independent reviewers and signed reviews in an estate with one accountable person. Its rule 14 records the decision: the owner is accountable, the delivery owner is the agent session holding the task, and an independent, signed review is a second session's review recorded with session, date and commit. Named human approvals remain the owner's; seven Metis cells whose packets were ready are tagged and listed for the owner.

10. How the last passes ended, and what this work broke (19 September)#

The passes section 9 left in progress.

• V1 residual backlog. Its 42 open items were clusters from a June audit. Every finding was re-read in its evidence report and re-checked against today's source: 45 are recorded as already fixed, with path and line (the Veritas and Nyx facades and the Telegram reply path are overtaken in code), and 224 are child tasks with files, a Verify and a release tier, so V1.0 is worked first.
• P2 backlog. Its actionable tail (sections 41 to 58) was respecified section by section: bare nouns under a heading became self-contained tasks, "Confirm …" items point at the workflow that already exists or became real build tasks where the code contradicts them (11 of 658 actions SHA-pinned, no secret scan in the pre-commit hook, no admin MFA, unsafe-inline and unsafe-eval in the web CSP), and cloud-era recovery and cost items were restated for one Hetzner box. All 222 items are done: 42 children added, 54 tagged, about 39 pointing at what already exists, three no longer boxes, and no checked item touched.
• Oya. Nearly every open box said its residue was "the model", "the transport" or "the simulator", and no task provided any of them. Six ground items now come first (an ONNX runtime, a model register with licences and measured CPU speed, PX4 software-in-the-loop, a replay harness, the language-model seam, Postgres storage); 35 boxes name what they wait for and 32 are open with a Verify.
• V5. Its 27 open engine items were one line each. Every class already exists as a function library that returns a descriptor; each item now says which descriptor exists, what the residue is and which automation test proves it.
• The V1–V9 content ledger and phase 180 turned out not to be editable at all, which is the next point.

What this work broke. A tracker's text is read by gates, and nobody had run them. Nine were broken by this audit's own edits and found only by running everything that reads a tracker this session touched:

Every gate that passed before this work passes again. Gates that fail for older, unrelated reasons were compared before and after and are unchanged: 8 of 1,156 V2 checks, 8 of 118 V6 verifiers, the phase 14, 24 and 97 verifiers, the crosswalk's T.20.08, the docs source-of-truth registry, the workbench inventory drift check. One could not be run here at all (the charter-workflow inventory needs a generated product-graph file this worktree lacks); its snapshot hashes the V10 tracker, which this work changed.

The lesson is written where it binds: a section of .claude/rules/task-checkbox-verification.md ("A tracker is read by gates"), task ETB.4.08 of the Eve Task Board plan (the CLI's write-through gets a write policy per family and never writes the three procedure-only files), and the agent memory.

The plan's own defects, found by the worker using it. The ./eve shim handed every call to the new CLI and took next, board, check and files away for an afternoon; it now falls back to the markdown board while the CLI answers "not built", and a spec pins that. The CLI's table of planned commands named the wrong task ids; the plan now lists the right ones.

Where the board stands (19 September, 00:30 UTC). Tracker families: 13,375 open boxes, of which 3,810 are actionable, 8,485 are held back in a deferred section with its reason, and 1,080 are parked (910 tagged item by item: external 243, upstream 221, human 105, hardware 100, release 94, governance 90, corpus 56, gpu 1; the rest in families parked whole). On the morning of the 18th the board called 13,627 tracker boxes actionable. Coverage walkthroughs add 9,621 actionable per-view checks, and the three gate families (1,299 boxes) are parked until a launch is scheduled. The five ranked families hold 2,650 of the actionable tracker tasks: the Eve Task Board 37, the presentation redesign 1,136 (Phase 0 first), Isis 226, Eve SOTA 95, the workbenches 1,156.