Eve library

V1–V9 content ledger: re-audit what September changed (2026-09-18)

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

V1_V9_AUTONOMOUS_CONTENT_SOTA_GAP_CLOSURE_TODOS_2026-07-15.md has 145 open boxes and its own machine audit of them, dated 2 August 2026: evidence/v1-v9/remaining-actionability-manifest.json, generated and checked by scripts/audit/remaining-actionability-manifest.mjs. The audit's verdict is that nothing in the ledger is agent work any more: 8 boxes are standing evidence rules, 12 are portfolio gates, and 125 wait on state outside the repository (licensed human gold, cohorts, live providers, production access, release events). The Eve Task Board therefore parks the whole family and points at the manifest for the reason behind each box.

Do not edit an open box of that ledger by hand. The manifest pins the text of every unchecked box by SHA-256 (EXPECTED_UNCHECKED_DIGEST); adding a note or a tag to one fails node scripts/audit/remaining-actionability-manifest.mjs --check, its test and node scripts/verify-remaining-actionability-audit.mjs. The task-board audit of 18 September 2026 did exactly that to twenty boxes in the morning and restored them, byte for byte, in the evening.

Five facts have changed since 2 August, and five blocker profiles may no longer hold in full. Each task below re-audits one package through the ledger's own procedure. The procedure for every task:

  1. Read the package in the ledger, its evidence file under evidence/v1-v9/ and its profile in BLOCKER_PROFILES.
  2. Do the part of the work that the changed fact makes possible, under the ledger's evidence-pack rule (its section 0.3) and the repository's model and spend rules (CLAUDE.md: cheapest adequate model through OpenRouter for anything that is a test or a harness; quote a spend and ask before a paid render).
  3. If a box is now satisfied by current, committed evidence, check it in the ledger. If it is not, leave it and correct the profile's reason and requiredExternalEvidence so they say what is still missing.
  4. Regenerate the manifest, update EXPECTED_UNCHECKED_DIGEST and AUDITED_REPOSITORY_TIP, and run the three checks named above. Verify for every task: all three pass, and the evidence file says what was run.
  • LRA.1 V1-003, "real provider-backed artifacts for a bounded representative set of text and media slots". The profile's blocker was "no detected live provider credential names". OPENROUTER_API_KEY has been in ~/.config/oshun/env-master.env on the Linux server since 10 September 2026 (checked by name on the 18th; no RunPod key name is in that file). Produce the text slots on the cheapest adequate route and record the route on every artifact. The media slots go through the Isis lanes, which need the RunPod credentials wherever the Isis work keeps them: find them by name, quote the spend and ask before rendering, or leave the media half profiled as external and say so. The other five V1-003 boxes (live operator receipts, licensed calibration, outcomes, experiment, canary) are untouched by this fact and stay as profiled.
  • LRA.2 V5-001, "integrate real TTS, recap, STT and translation providers". Kokoro and Whisper already run on the Linux server's CPU for other work in this repository, and translation and recap are text calls. Under the install-first rule, bind the four seams to those, keep the vendor routes as configuration, and record latency and cost. Whether an open local model satisfies the ledger's word "real provider" is the judgement this task must write down; the human receipts box is untouched.
  • LRA.3 V8-005, "generate cases through real model/media providers under the solve-first truth kernel", then the Unreal box that follows it. Same changed fact as LRA.1 for text; media needs a spend ask. The executing machine has Unreal 5.5 (ue55-linux, run as ueagent), so the place, cook and play box can follow once cases exist. The four human, comparison, drill and sign-off boxes stay as profiled.
  • LRA.4 P1-004, "publish model, rubric, prompt, policy, provider and data cards with known limits and last calibration date". Cards are documents an agent can write from the registries and evidence that exist; where no calibration has happened the card says so. The cadence-review and owner-and-deadline boxes stay as profiled: they need the owner.
  • LRA.5 V2 and V3 editor builds. V2-001, V2-002, V3-001 and V3-002 record that Unreal 5.5.4 is present and that the retained V2 editor build id is stale and V3 has no UnrealEditor.modules. Rebuild both editors on the executing machine, resource-safely (check free -m; nothing else heavy running), and refresh evidence/v1-v9/V2-unreal-host-preflight.json and its V3 twin. This closes no ledger box by itself: V2 still has no playable match (V2/V2_TODOS.md section 0), so correct the profiles to name that, not the build, as what is missing.
  • LRA.6 Re-run the whole audit once LRA.1 to LRA.5 are done: every remaining profile re-read against the repository tip, the manifest regenerated, and the task board's registry row for the ledger updated with the new counts. If any box has become agent work, change the family from parked to open and say which boxes in its focus.