AGENTIC.a2ml instructs agents to 'Never use AGPL' — contradicts Rules 3, 4 and 5

[hyperpolymath]

An agent instruction that contradicts three licence rules

Every generated .machine_readable/6a2/AGENTIC.a2ml carries this line in its
"never do this" list:

# - Never use AGPL license (use MPL-2.0)

AGENTIC.a2ml is read by AI agents as a behavioural constraint. This one is now wrong
three times over.

What it contradicts

Rule	Says	The instruction says
Rule 3	Projects co-developed with the owner's son are AGPL-3.0-or-later	never use AGPL
Rule 4	007 — the Agent Meta-Language — is AGPL-3.0-or-later	never use AGPL
Rule 5 (added 2026-08-26)	Games are AGPL-3.0-or-later as a class	never use AGPL

The sharpest case

007-lang's LICENSE is the GNU Affero General Public License. Its own agent
instructions say - Never use AGPL-3.0.

So the single repo that Rule 4 names as AGPL instructs every agent reading it to never
use AGPL. An agent doing as it is told would "correct" the licence the policy requires.

Scale

• 899 local AGENTIC.a2ml files carry the line
• GitHub code search returns 100+ hits across 82 repos (capped, so the live figure is
  higher)
• It appears alongside ECOSYSTEM.a2ml, META.a2ml, NEUROSYM.a2ml, PLAYBOOK.a2ml,
  STATE.a2ml — a generated 6a2 manifest set, not hand-written

Why this is filed rather than swept

Same reasoning the pmpl-to-mpl-sweep-runbook §7a gives, and it was right there:

Fix at the source, not 236 times. Doing manual edits invites immediate re-drift on
the next scaffold run.

899 edits would be undone by the next 6a2 regeneration. The generator must be found and
fixed first. It is not in rsr-template-repo, scaffoldia, or
rsr-julia-library-template-repo — all three return 404 for that path — so the emitter is
somewhere else and needs locating.

Suggested wording once the source is found

# - Never change an existing file's licence (owner-only, manual — see LICENCE-POLICY A2)
# - New code defaults to MPL-2.0; AGPL-3.0-or-later is CORRECT under Rules 3, 4 and 5

That preserves the intent — MPL is the default — without instructing agents to undo
deliberate policy.

Related

This is the third instance today of the same failure mode: an automated artefact whose
self-description has drifted from what the estate actually wants.

• scaffoldia's rhodibot ran a sed sweep rewriting AGPL headers — fixed, fix(ci): rhodibot must never relicense AGPL — warn only scaffoldia#36
• hyperpolymath/.github carried a funding file pointing at the wrong org — fixed, fix: remove shadowed CODEOWNERS + stray funding file, repair dependabot stub .github#9
• and this

Horizon: 899 is a count of local checkouts under hyper-repos/ and meta-repos/, which
include vendored sub-copies; the number of distinct estate-authored files is lower. The
GitHub figure of 82 repos is capped at 100 results and is a floor, not a total.

[hyperpolymath]

Investigated 2026-08-26 — there is no generator, and this issue's remediation premise is inverted

1. Where the files come from

hyperpolymath/rsr-template-repo ships .machine_readable/descriptiles/AGENTIC.a2ml verbatim. It is not written by code; it is copied by GitHub's "Use this template", and just repo-init's token-substitution pass does not touch the [agent-constraints] comment. Repo is_template: true, pushed today.

⚠️ An earlier pass in this session reported the template "ruled out". That was a false negative — the directory was renamed 6a2 → descriptiles on 2026-06-29 (3b5b5de, PR #128). A 404 on the old path was read as repo-wide absence. Re-verified live: .machine_readable/6a2/AGENTIC.a2ml → 404, .machine_readable/descriptiles/AGENTIC.a2ml → 1,772 bytes, offending line at line 24.

The only code that actually writes the file is gitbot-fleet/robot-repo-automaton/src/skeleton.rs:59-76 — and it is dormant: grep -rn 'skeleton' gitbot-fleet/.github/workflows/*.yml returns zero hits, and no workflow in rsr-template-repo references it either, despite skeleton.rs's own doc comment claiming "rsr-template-repo's drift-CI runs robot-repo-automaton skeleton check .". That CI does not exist.

The bulk of the population came from a chain of one-off estate sweeps, not an emitter — 101 of the files were added by a single 2026-03-16 commit (chore: restructure 6a2ml files into .machine_readable/6a2/).

2. ⚠️ The remediation premise here is backwards

This issue argues that manual fixes "would be undone by the next 6a2 regeneration". There is no regeneration. The population is static. A sweep would hold.

What must be fixed at source — to stop new repos inheriting — is three files, and only three:

file	state
rsr-template-repo/.machine_readable/descriptiles/AGENTIC.a2ml	live propagator → rsr-template-repo#45
gitbot-fleet/robot-repo-automaton/templates/skeleton/.machine_readable/6a2/AGENTIC.a2ml	dormant, still on PMPL wording
standards/agentic-a2ml/examples/comprehensive.a2ml	spec example, still on PMPL wording

3. The real problem is not staleness, it is contradiction

The line reads # - Never use AGPL license (use MPL-2.0). LICENCE-POLICY.adoc now defines five rules and three of them mandate AGPL-3.0-or-later: Rule 3 (co-developed), Rule 4 (network-deployed services), Rule 5 (games).

An agent obeying this line inside a games repo would "correct" a deliberately AGPL work to MPL-2.0 — the exact regression a bot in scaffoldia was caught scheduled to perform and had removed.

4. Corrected count

The figure of 899 is wrong. Measured by enumerate-then-filter over both accounts:

	hyperpolymath	metadatastician	total
repos enumerated	384	36	420
repos with ≥1 AGENTIC.a2ml	322	28	350
blobs	615	32	647

⚠️ 647 is a floor: 1 of 420 trees (julia-ecosystem) returned truncated: true.

Files actually carrying the line: 646, not 899. The 899 appears to be a count of files at the 6a2 path, not of files containing the line.

Wording variants across all 646:

381  # - Never use AGPL license (use MPL-2.0)
253  # - Never use AGPL license (use PMPL-1.0-or-later)     <- RETIRED licence
  4  # - Never use AGPL license (use AGPL-3.0-or-later)     <- self-contradictory
  8  other wordings

The 253 PMPL copies are a second defect this issue does not mention — they instruct agents toward a licence that no longer exists.

5. Homogeneity, and the good news

141 distinct byte-contents across 1,018 local files — a small number of template vintages plus a long tail of hand-edits. Top classes: 125 × 34-line MPL, 115 × 34-line PMPL, 86 × 24-line MPL, 61 × 51-line (= template HEAD).

Unfilled placeholders are essentially absent. All 1,018 scanned for {{…}}, ${…}, PROJECT_NAME, REPLACE_ME, PENDING, TODO, FIXME and others: the only {{ and TODO hits are prose about placeholders. rsr-template-repo/scripts/check-no-placeholders.sh is doing its job.

One real residue: hyperpolymath/proof-burrower still carries project = "rsr-template-repo" at line 8 — the sole unsubstituted self-name in the population.

Not checked

Non-default branches; julia-ecosystem's full tree (truncated); local dirs outside hyper-repos/ + meta-repos/; the sibling descriptiles (ANCHOR.a2ml, CLADE.a2ml, VARIANT.a2ml), which were neither counted nor hashed.

[hyperpolymath]

Re-measured 2026-08-31. Numbers stale — the fix is half-shipped and stalled.

Sources: all fixed. rsr-template-repo (note: the file moved to machine-readable/descriptiles/AGENTIC.a2ml, no leading dot), the gitbot-fleet skeleton, standards' own agentic-a2ml/examples/comprehensive.a2ml, and 007-lang (which now states its AGPL ruling explicitly — the sharpest contradiction is cured). Population is static: nothing regenerates the old wording any more.

Deployed copies: a direct-push sweep ran 2026-08-28 00:03–05:33Z, alphabetical ("fix: AGENTIC licence line + a2ml-validate-action repin" commits, referencing this issue), and stopped after launch-scaffolder, before lithoglyph — never resumed. Remote sampling brackets the boundary exactly: 20/20 repos in the swept range carry the new wording; 24/24 in the unswept range still carry "Never use AGPL", including every metadatastician repo checked. Of 158 remotely-verifiable canonical paths, 95 remain; extrapolated against the sweep's 309-repo scope, roughly 170–185 repos (~55–60%) still carry the contradictory line (extrapolation, labelled as such — the local disk is too stale to trust on this issue; remote is the only honest plane).

This issue is now the tracker for the stalled remainder. Next step: resume the direct-push sweep at lithoglyph, alphabetical to the end, then all of metadatastician, using the same commit wording. The 08-26 census figures (646 files / 350 repos) are obsolete and should not be re-quoted.