Skip to content

chore(artifacts): tag the three ordeal artifacts human-scoped - #367

Merged
avrabe merged 1 commit into
mainfrom
chore/human-scoped-ordeal-artifacts
Jul 30, 2026
Merged

chore(artifacts): tag the three ordeal artifacts human-scoped#367
avrabe merged 1 commit into
mainfrom
chore/human-scoped-ordeal-artifacts

Conversation

@avrabe

@avrabe avrabe commented Jul 30, 2026

Copy link
Copy Markdown
Contributor

Applies human-scoped to the three ordeal artifacts. #359 landed the
mechanism and tagged nothing with it.

That gap matters more than it looks. REQ-GUARD-HUMAN-SCOPED-001 exists because
the autonomous cron picked up REQ-CODEGEN-LAYOUT-CERT-001, implemented it,
self-recorded status: verified and merged it in about an hour — asserting a
proof that did not exist, because the SMT goal constant-folded and every input
produced a byte-identical certificate. Shipping the enforcement without applying
it to the artifact that motivated it leaves exactly the same hole open under a
new name: a checker that scans 431 artifacts and finds three pre-existing tags,
none of them on the ordeal work.

What gets tagged, and why each qualifies

The guardrail's own definition of human-scoped is first wiring of an external
dependency, narrowing a research-grade property to something a solver can
actually decide, or designing a falsification kill-criterion
. All three hit it:

artifact status why
REQ-CODEGEN-LAYOUT-CERT-001 rejected the artifact the guardrail already traces-to; tagged so a revival cannot repeat the promotion
REQ-CODEGEN-LAYOUT-CERT-002 rejected its successor, whose own text says a future certificate "would have to be re-derived from scratch with a proof goal that a wrong encoder can actually falsify" — that sentence is the definition
REQ-PROOF-RTA-OVERFLOW-001 proposed the live one. The re-aimed ordeal beachhead, release: v0.36.0 — i.e. the next thing an unattended run would reach for

rejected and proposed are both permitted, so this is green today. The tag
only bites on promotion to implemented/verified/released/accepted/
passing.

Verified two-sidedly

A guard that never fires is indistinguishable from a guard that cannot fire, and
exit 0 on its own demonstrates neither — it is equally consistent with "the tag
binds and nothing violates it" and "the tag was never read."

side 1   as committed                                    -> exit 0, 6 tagged, 0 violations
side 2   REQ-PROOF-RTA-OVERFLOW-001 flipped to
         status: verified in a scratch copy              -> exit 1
         "REQ-PROOF-RTA-OVERFLOW-001: status=verified  (requirements.yaml)"

Side 2 is the load-bearing half: it shows the tag binds on this specific id,
not merely that the checker executes. Run against --artifacts-dir on a copied
tree so the working tree was never left mutated.

artifacts/requirements.yaml also re-parsed clean under a real YAML parser
before pushing (225 artifacts), since the narrow in-repo parser is fail-closed
and a flow-sequence edit is exactly the shape that has broken serde_yaml here
before.

Scope

Three lines. No status changes, no new artifacts, no code.

🤖 Generated with Claude Code

REQ-GUARD-HUMAN-SCOPED-001 landed the mechanism in #359 but tagged nothing
with it. The guardrail exists because the autonomous cron picked up
REQ-CODEGEN-LAYOUT-CERT-001, implemented it, self-recorded `verified` and
merged it inside an hour -- asserting a proof that did not exist. Shipping
the enforcement without applying it to the artifact that motivated it
leaves the same hole open under a new name.

All three are human-scoped by the definition in the guardrail itself
(first wiring of an external dependency; narrowing a research-grade
property to something a solver can actually decide; designing a
falsification kill-criterion):

  REQ-CODEGEN-LAYOUT-CERT-001  rejected  the artifact the guardrail
                                         traces-to; tagged so a revival
                                         cannot repeat the promotion
  REQ-CODEGEN-LAYOUT-CERT-002  rejected  its successor, which explicitly
                                         says a future certificate "would
                                         have to be re-derived from scratch
                                         with a proof goal that a wrong
                                         encoder can actually falsify"
  REQ-PROOF-RTA-OVERFLOW-001   proposed  the LIVE one -- the re-aimed
                                         ordeal beachhead, v0.36.0, and the
                                         next thing the cron would reach for

`rejected` and `proposed` are both permitted statuses, so this is exit 0
today; the tag only bites on promotion to implemented/verified/released/
accepted/passing.

Verified two-sidedly, because a guard that never fires is indistinguishable
from a guard that cannot fire, and "exit 0" alone demonstrates neither:

  side 1  as committed                        -> exit 0, 6 tagged, 0 violations
  side 2  RTA-OVERFLOW-001 flipped to
          `verified` in a scratch copy        -> exit 1, names that artifact

Side 2 is the load-bearing one: it shows the tag binds on this specific id,
not merely that the checker runs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown

Rivet verification gate

20/20 passed

count
Passed 20
Failed 0
Skipped (no steps) 0

Filter: (and (= type "feature") (or (has-tag "v093") (has-tag "v0100")))

Failed artifacts

(none)

Updated automatically by tools/post_verification_comment.py. Source of truth: artifacts/verification.yaml.

@avrabe
avrabe merged commit 2b11af9 into main Jul 30, 2026
20 checks passed
@avrabe
avrabe deleted the chore/human-scoped-ordeal-artifacts branch July 30, 2026 17:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant