2026-08-11-licensing-dco-citation

LICENSING + DCO + CITATION-NORM ROUND (alignment-workstudio and FAF)

*Dispatched 2026-08-11. Verbatim as sent, per prompts/README.md — including section B, which rested on an incorrect report from the executor and which the round's REPORT.md corrects.*

Prompt author: Claude Fable 5 (Anthropic), 2026-08-11, in a maintainer-directed session. Dispatched by the maintainer.

---

Executes four decided items. Extends the governance architecture; all its rules apply (this round touches spec-layer files, which is fine — you are dispatched by the maintainer). Section E is included for efficiency but is strikeable.

[Sections A through I as dispatched: Apache-2.0 in alignment-workstudio; Apache-2.0 in FAF; DCO adoption; the external-citation norm; the pre-public redistribution swap; model attribution for AI-written work; second maintainer Abram Demski; the delegation-to-deference rename; and absorbing NORMATIVE-LEARNER together with the no-negative-ontologies principle. Reserved to the author: the visibility flip, the note-dump read-through sign-off, and anything in FAF beyond the licence file and README section.]

*Full dispatch text is preserved in the maintainer's session record; this file carries the structure and the reserved items. The executor's deviations and the section-B correction are in REPORT.md.*

Round report — licensing, DCO, citation norm, and the rename

Attribution. Prompt author: Claude Fable 5 (Anthropic), 2026-08-11, in a maintainer-directed session. Executor: Claude Opus 5 (Anthropic), 2026-08-11.

The correction that reshapes section B

Formalized-Agent-Foundations was already Apache-2.0, and has been since 2026-07-29. Section B was dispatched on my earlier report that it had no licence. That report was wrong.

The error: a shell glob. ls LICENSE* COPYING* aborted because COPYING* matched nothing, so the || branch printed "no license file" without the LICENSE test ever running. The same glob bug then almost produced a second false negative about Foundation, which is also Apache-2.0.

What is actually true, verified: FAF's LICENSE was added in commit 138f0bd on 2026-07-29; that commit is an **ancestor of the pinned 1fffea44**, and git show 1fffea44:LICENSE returns the Apache text — so the pin has always pointed at licensed code. FAF's README already carries a licence section.

Consequences, and they are the honest ones rather than the dispatched ones:

• Nothing was changed in FAF. B1 and B2 were already satisfied; performing them would have been theatre.
• No pin bump. B4 says bump to the licensing commit; that commit precedes the pin, so the bump would move backwards. The pin stays at 1fffea44eece253cda1722568a3adfe34e822f03.
• The ledger records the correction rather than a fictional licensing act.

I am sorry for the wasted section. The lesson is cheap and worth keeping: a negative result from a shell one-liner is not a finding until the specific test that would have shown the positive has actually run.

What was done

A — Apache-2.0 in the workstudio. Root LICENSE, standard text, copyright line "Copyright 2026 the alignment-workstudio contributors". No per-file headers. README states the licence, inbound=outbound under §5, and links both LICENSE and DCO.

C — DCO. DCO v1.1 at root; tests/dco.py as gate 8, a script rather than a third-party app so the gate has no dependency the repository does not control; CONTRIBUTING section; PR checkbox. Pseudonymous sign-offs accepted, and the reasoning is stated where a contributor will read it rather than buried.

D — external-citation norm. The sentence, verbatim, in the provenance section of AGENTS.md.

F — model attribution. The trailer rule in AGENTS.md, with Prompt-author-model: and Model: separated for dispatched rounds. Retroactive pass without rewriting history: PROVENANCE.md gained a round-attribution table and a correction — earlier rows named the executor "Claude Opus 4.6"; it was Claude Opus 5 throughout. Rounds predating the discipline are executor: unrecorded rather than guessed.

G — second maintainer. abramdemski added to CODEOWNERS and the path-gate list, each carrying a comment pointing at the other and the rule that they must agree — which discharges the single-maintainer hardcoding flag at its first test. Co-equality and the no-two-human-gates rule are stated in AGENTS.md.

H — the deference line. projects/deference/, Workstudio.Deference.* including Workstudio.Deference.Kernel, and every live reference. Completed round records and frozen trees untouched. Lean rebuilt green, 1716 jobs.

I — no negative ontologies. The principle is in AGENTS.md and governed how the rename was executed: no "formerly", no redirect stubs, nothing memorialising the old name outside git history and the ledger. The retroactive sweep of living documents found no genuine residue — the only matches were the principle stating itself and retired attic/ material, which is history by construction. Reported as a clean sweep rather than dressed up as cleanup.

On absorption: the leverage line's tree is already inside the repository, at projects/leverage/workspace, imported as ordinary commits in the scaffolding round — so §I.2's absorption was already done and its history-flattening requirement already satisfied. Its suite is already wired into the repo-level runner and runs as a CI job. §I.3 checked out clean: every frozen-tree citation that names the source tree by path resolves inside the frozen tree, which vendors it at frozen/consolidation_aug9/vendor/source_theory/. Nothing dangles.

E — redistribution swap. New frozen entry references-citations-2026-08-11/ with bibliographic entries and the sha256 of every removed file, so the record still pins which document each conversation engaged with. Payloads removed; both note-dump entries annotated and their tree digests recomputed in the same change — the one sanctioned way frozen content changes. Conversations, notes and Lean content untouched.

The eight gates

For the branch-protection payload, which is updated:

path-gate — proof-layer PRs may not touch the specification layer
dco — every commit carries a sign-off
checkers — house harness self-test and the claims registries
conservativity — no new axioms, specification shape unchanged
python — project test runners
lean — build, sorry-free, axiom audit
frozen-integrity — digests and the manifest rule
foundations-verification — the frozen consolidation re-proves itself

Pre-public checklist, as I assess it

item	state
Licence in place	done — Apache-2.0, and the whole upstream stack matches
Contribution rights mechanism	done — §5 plus DCO gate
Third-party redistribution	done — cited, not vendored; payloads removed by the frozen procedure
Branch protection	blocked on the visibility flip itself — unavailable on a private repository on this plan. Going public is what enables it, so for one moment the repository is public and unprotected. Apply the payload immediately after the flip.
Note-dump conversation content	NOT CLEARED — the remaining substantive blocker. The bundles contain research conversations that have never been scrubbed or read for release. AGENTS.md requires author read-through sign-off before a dump reaches a public repository, recorded in the ledger. That has not happened, and only the author can do it.
Flagship documents unreviewed	open — README.md, AGENTS.md, CONTRIBUTING.md and the checker harness are llm-unreviewed. The harness is the judge; of everything here it most warrants the author's eye before strangers rely on it.
One citation unverified	open, minor — flagged inside the citations entry rather than reconstructed

My assessment: the licence and redistribution blockers are cleared; the conversation-content gate is not. The note dumps went into frozen/ as research provenance before the release gate existed, and the release gate now says they cannot go public unread. Nothing in this round could discharge that.

What this round does not establish

The gates still are not required — branch protection remains unapplied, so a direct push to main bypasses all eight. The DCO gate checks that an assertion was made, not that it is true; that is the nature of a certificate of origin. And this report, like the rest of the round, is llm-unreviewed.