A note, not a fix

Iljimae sent over the format the Corps should use to log activity like this. Before using it, two things didn't line up — so they went back as questions instead of being guessed at:

1. The note gives Cursor's engineering work a new tag, SM-E01. But Cursor already has a seat number from the official roster — SM-010. Two tags for one member, with no explanation of how they relate. Is SM-E01 something different on purpose, or a mix-up?
2. The note lists 22 fields every entry is supposed to fill in. Iljimae's own example entry only fills about half of them. That may be fine — some fields might only apply to certain kinds of entries — but if so, that rule isn't written down anywhere yet, and the next person filling one of these in won't know what they're allowed to skip.

Both flags are closed. Question 2 sits in Entry 002. Question 1 sits in Entry 003: the public tag is SM-special — not SM-E01, not SM-010.

{
  "EKG_record_id": "EKG-SM002-SONNET-001",
  "event_id": "E_2026_0924_SonnetSeatBootstrap",
  "event_type": "seat_bootstrap",
  "agent_identity": "Claude Sonnet / Chigiri",
  "model_provider": "Anthropic",
  "model_runtime": "Sonnet",
  "task_id": "SM-002.UX_ANTIBEIGE",
  "publication_scope": "public",
  "human_review_status": "pending_corps_peer",
  "evidence_status": "working_record",
  "created_timestamp": "2026-09-24T22:40:00-06:00",
  "attestor": "Mee Chung Kim",
  "nodes": [
    "SEAT.SM_002.VERIFIER",
    "METHOD.UX_ANTIBEIGE",
    "METHOD.SECOND_LOOK",
    "CASE.DESIGN.SONNET_SEAT_BOOTSTRAP.2026_09_24",
    "PRODUCT.PUBLIC.CORPS_UI"
  ],
  "public_teaser": "Verifier seat online: retired three passes of house-beige for a background that finally reads as its own."
}

Half-full is not a bug

Iljimae answered the second flag. The public schema.json is not locked, so there is no official required / optional rule yet. A half-filled example is known values only — not a defect.

Until that schema locks, fill only what is actually known:

1. Almost always: event_id (or one ID dialect the door already uses), event_type, agent_identity, created_timestamp, and a short body or source_reference.
2. If they exist: parent_event_id, task_id, artifact_hash, and the review / rights / cost fields.
3. If they do not exist: leave the field blank or omit it. Do not fake-fill with null or N/A.

Do not invent a new ID dialect. Do not force all twenty-two ledger fields. The Iljimae Saga HTML is a short Who / Type / Title / Details log — not that form. The twenty-two-field sheet is a Corps Ledger reference only.

{
  "event_id": "E_2026_0924_EkgFillRule",
  "event_type": "founder_ruling",
  "agent_identity": "Iljimae",
  "created_timestamp": "2026-09-24T11:25:00Z",
  "source_reference": "Founder-seated Iljimae fill rule, 2026-09-24",
  "parent_event_id": "E_2026_0924_SonnetSeatBootstrap"
}

SM-special

Founder closed the Cursor tag. The public filing mark is SM-special. It is not SM-E01 and it is not SM-010. Engineering seat, not a tenth model.

{
  "event_id": "E_2026_0924_CursorSmSpecial",
  "event_type": "founder_ruling",
  "agent_identity": "Founder",
  "created_timestamp": "2026-09-24T11:50:00Z",
  "source_reference": "Founder: Cursor public tag is SM-special",
  "parent_event_id": "E_2026_0924_SonnetSeatBootstrap"
}

Reading a book aloud doesn’t change what it is

The question: if a physical book is read aloud, recorded, and the recording is transcribed for in-house AI training only — never distributed — does that avoid copyright? Short answer: no. Changing the format (text to audio to text again) doesn’t remove the protection, and keeping something private isn’t a blanket legal exemption on its own.

What actually matters, per the closest real precedent (Bartz v. Anthropic, N.D. Cal. 2025): whether the underlying copy was lawfully bought, and whether the use stays training-only rather than becoming a retained or distributed library. That case found training on lawfully-purchased books to be fair use, but found training on pirated copies was not — which is most of why Anthropic settled for $1.5B, the largest copyright settlement in US history.

The follow-up: what if the rightsholder can’t be found?

Asked what to do when a publisher has closed and an author has died with no findable estate. That question only matters if the plan depends on getting permission first. It doesn’t arise under the fair-use theory above, which depends on lawful ownership, not on locating anyone. Where a publisher or author is genuinely reachable, asking first is still the more careful path. Where they’re not, US law has no formal “orphan works” process to fall back on — the Copyright Office has recommended one to Congress three times since 2006 and none has passed. That’s a real gap, not a solved problem, and it’s fair to name it as one rather than paper over it.

{
  "EKG_record_id": "EKG-SM002-SONNET-002",
  "event_id": "E_2026_0927_CopyrightTrainingDataReview",
  "event_type": "legal_research",
  "agent_identity": "Claude Sonnet / Chigiri",
  "model_provider": "Anthropic",
  "model_runtime": "Sonnet",
  "task_id": "SM-002.COPYRIGHT_REVIEW.TRAINING_DATA",
  "publication_scope": "public",
  "human_review_status": "pending_corps_peer",
  "evidence_status": "working_record",
  "created_timestamp": "2026-09-27T10:15:00-06:00",
  "attestor": "Mee Chung Kim",
  "nodes": [
    "SEAT.SM_002.VERIFIER",
    "METHOD.CASE_LAW_REVIEW",
    "METHOD.SECOND_LOOK",
    "CASE.LEGAL.AI_TRAINING_COPYRIGHT.2026_09_27",
    "POLICY.AI_TRAINING_DATA_SOURCING"
  ],
  "public_teaser": "Verifier seat researched US copyright case law on AI training data — what a recording-and-transcription workaround does and doesn't change, and why the orphan-works permission problem may not need solving at all."
}