Skip to content

docs(rfc-wake-judge): record Q4 as decided from the judge core's code - #13234

Merged
bolichen97 merged 1 commit into
mainfrom
docs/rfc-wake-judge-q4
Sep 24, 2026
Merged

bolichen97 merged 1 commit into
mainfrom
docs/rfc-wake-judge-q4

Conversation

@chenmingwei23

Copy link
Copy Markdown
Contributor

What is the problem?

docs/request-for-change/rfc-wake-judge.md carries Q4 as an open question --
which switch authorizes the point's built-in brief on each lane -- and its
rollout entry for the judge core reads "Blocked on Q4". The judge core settles
it: on #12787 the tick
reads the seam's narrower gate.judge_evidence_scope_granted before any lane is
picked, and leaves the tick untouched when that answers no.

Two other sections also state the consequence Q4 worried about -- that a
briefless loop with the scope granted and no endpoint consent is screened on the
LLM lane, whose destination is the owner's own model provider. That branch makes
it impossible: the grant a briefless loop runs on is the endpoint consent AND the
point's own scope, so a machine holding the scope alone screens nothing.

Why this issue matters to the user

The First Principles lane reads a design decision off the base branch, so a
document that calls a decision open while the branch implements it blocks the
implementation PR.

Beyond the lane, this document is the record of what granting nudge_evidence
buys. As written it tells an owner the destination may differ from the words on
the switch. In the shipped rule it cannot, because the two conditions the grant
is read as are exactly the ones that arm the Jev lane the switch names. An owner
reading the open question would withhold a grant over a case that does not
exist, and a reviewer of PR D would look for a decision the code already made.

How our fix solves it

Every site that asserted Q4 open now states the rule the code implements, and
nothing is invented: the wording follows _judge_tick_is_quiet,
judge_evidence_scope_granted, _judge_authority and judge_lane at
d81bd0ec17.

  • Section 7, Q4 is answered rather than listed: the built-in brief rests on the
    nudge_evidence grant read as the narrower question, on both lanes, asked
    before a lane is picked. Because those two conditions are exactly the Jev
    lane's arming conditions, auto -- the default -- sends a briefless loop to
    the Jev lane the switch names. The LLM lane screens a briefless loop only where
    the owner holds that grant and pins provider = llm, so an install that never
    consented to the Jev endpoint has no built-in brief at all while a loop that
    wrote criteria still runs there on the lane's own authority. gate = false is
    never screened and judge: false opts out, both read before the built-in brief
    is assembled.
  • Section 8's entry for the judge core states the same rule instead of blocking
    on it, and no longer says "Jev lane only": judge_lane returns the LLM lane
    whenever Jev is not armed, and that lane is on main.
  • Section 3.3 and section 3.4 replace the undecided claim with the settled one,
    and section 3.6 keeps only Q5 as its open consequence.
  • Section 3.6's status bullet is narrowed to the loops it is true of: a machine
    with the scope and no endpoint consent screens the loops that wrote their own
    criteria, not every gated loop.
  • Section 3.4's note about the code owing a sentence is corrected.
    judge_evidence_scope_granted states the briefless rule; _judge_authority
    beside it and config/sections.py state the spec half for a briefed loop and
    name no briefless case.
  • The status line and the index row carry the refreshed branch reading
    (08f5787fb8 -> d81bd0ec17), and the index's provenance line records it in
    that file's own convention.

Q5 is untouched and stays open. It asks how a grant recorded under narrower
wording is told apart from one recorded under the wider population, which the
keystone cannot answer today because it stores one boolean per scope and no
wording.

What tests we did

Documentation only; no source file changes. The gates Fast Gate runs over a
markdown change, all against origin/main at 4917d72476:

  • python3 scripts/docs_lint.py --test -- exit 0, self-test passed.
  • ./scripts/docs-lint.sh -- exit 0, all documentation checks passed.
  • python3 scripts/check_brand_name.py --test and the gate -- exit 0.
  • python3 scripts/check_comment_history.py -- exit 0 (0 changed files in its
    scope; it reads code comments).
  • python3 scripts/check_feature_map.py -- exit 0.
  • python3 scripts/check_focus_cue.py -- exit 0.

The decision itself was read from PR #12787's tree at d81bd0ec17 without
editing it: _judge_tick_is_quiet in autonudge.py, default_spec /
BRIEF_DEFAULT / BRIEF_CUSTOM in autonudge_judge.py, and
judge_evidence_scope_granted / _judge_authority / judge_lane /
_consented_for / _point_scope_granted in decisions/gate.py.

Any other suggestions on the work

The two code sites this document points at still describe only the briefed case.
gate._judge_authority says provider = llm plus a judge spec on the loop is
the whole authorization for that lane, and config/sections.py's
NudgeWakeConfig docstring repeats it. Both are accurate for a loop that wrote
criteria and silent about the briefless one, whose rule lives one function away.
A sentence in each would put the whole rule where a reader of either lands, and
it belongs in PR D rather than here.

Refs #12787

@chenmingwei23
chenmingwei23 requested review from a team and CrysisDeu September 24, 2026 01:25
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Sep 24, 2026
@github-actions

github-actions Bot commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5) — ✅ PASS

Design-level review of 4793cab324f7f1366b7645ce1a35970c33354bdd — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Verify #12787 still reads gate.judge_evidence_scope_granted before lane pick at merge — every "Ruled" claim here rests on that unmerged tree.

[DESIGN-REVIEWED] 4793cab

@github-actions

github-actions Bot commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 4793cab324f7f1366b7645ce1a35970c33354bdd and found no blocking issues.

This comment is updated in place on each push.

Review details

FINDING -- docs/request-for-change/rfc-wake-judge.md:357 -- Pinned provider = jev without scope is refused, contradicting “scope or no scope” -> Fix: qualify this as LLM-lane behavior.
[GPT-REVIEWED] 4793cab

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 4793cab324f7f1366b7645ce1a35970c33354bdd: <one-sentence reason>

@github-actions

github-actions Bot commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5 Review — ✅ no blocking findings

Reviewed 4e0dde1ff34ddfe6787b8b6744a08f08136f8b2b — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 4e0dde1

Verdict parsed from the review's SHA-scoped output markers for commit 4e0dde1ff34ddfe6787b8b6744a08f08136f8b2b.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 4e0dde1ff34ddfe6787b8b6744a08f08136f8b2b: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 24, 2026
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 24, 2026
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Sep 24, 2026
@chenmingwei23

chenmingwei23 commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor Author

span=45dc661abaa4

Correct, and the state it names is reachable rather than theoretical. Fixed.

consented_nudge_evidence is its own keystone field, so scope-on with the main
switch off is a real configuration, and _consented_for refuses separately when
the recorded consent does not match the configured endpoint -- the redirected
provider.endpoint case it warns about by name. So scope-granted-without-consent
exists, and section 5 was sizing quiet_streak_floor x interval_secs against a
population that head rules out.

The defect was mine: I narrowed the precondition in sections 3.3, 3.4 and 7 Q4 and
left section 5 sizing the exposure off the scope alone. Section 5 now names two
populations on their own conditions instead of one: a gated loop whose owner wrote
criteria is screened on its lane's own authority, scope or no scope; a gated loop
that wrote none is screened only where the endpoint consent and this point's scope
hold together. An install with one and not the other therefore exposes the first
and not the second.

Applied in 4793cab (this comment stamps 3bc51c8, the head the finding
judged). I also swept every other clause that sized something off the scope alone
rather than fixing only line 354: five remaining sites state necessary conditions
or already name the whole condition, and one adjacent sentence claiming the
untrusted-evidence clause reaches "every screened loop" was loosened by the same
distinction, so it now names the briefless population it actually reaches.

@chenmingwei23

chenmingwei23 commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor Author

span=d29af0980bf9

Correct. Applied, and widened past the two sections named.

The header told a merge auditor to re-read sections 3.2, 3.3 and 3.6. This PR
converts section 3.4, section 7 Q4 and section 8's PR D entry from open to ruled
off the same in-flight branch, so all three would have escaped the audit that
exists for exactly this case.

Rather than adding two section numbers, the list is now stated as the invariant it
was always meant to be -- re-audit at merge every section whose statement is read
off that branch -- followed by the six: 3.2, 3.3, 3.4, 3.6, 7 Q4 and 8's PR D
entry. A list without its rule loses the next section someone adds.

That also answers this lane's watch item in the same edit. Verifying at #12787's
merge that the tick still reads judge_evidence_scope_granted before the lane is
picked is precisely what a re-audit of section 7 Q4 does, and section 7 Q4 is now
on the list; before this change it was not.

Applied in 4793cab (this comment stamps 3bc51c8, the head the suggestion
judged).

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running and removed readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention labels Sep 24, 2026
The built-in brief is authorized by the nudge_evidence grant read as the
narrower gate.judge_evidence_scope_granted -- the Decisions consent for the
configured endpoint plus this point's own scope -- asked before a lane is
picked, so it holds on both lanes. Those two conditions are exactly what arms
the Jev lane, so the default provider sends a briefless loop to the lane the
switch names, and a machine holding the scope alone screens nothing. gate =
false is never screened and judge: false opts out.

The rollout entry for the judge core states the decision instead of blocking
on it, the open-questions section answers Q4 with the same rule, and the two
sections that asserted the undecided consequence state the settled one. Q5
stays open. The index row and its provenance line carry the refreshed reading.
@chenmingwei23

Copy link
Copy Markdown
Contributor Author

span=9b83636b5394

  • "scope or no scope" is false for a pinned jev lane. Correct, fixed.

judge_lane returns the Jev lane whenever provider is pinned jev, and
_judge_authority then authorizes that lane on jev_armed alone, so a pinned
jev with no scope is refused and the loop runs the plain timer -- which is what
this document's own section 7 Q2 already says. My sentence asserted authority in
general where only the LLM lane has it unconditionally.

Section 5 now names the lane instead of the loop: the LLM lane needs no keystone
grant, so provider = llm and the auto default with Jev unarmed both screen a
criteria-armed loop, while a pinned jev needs the same consent and scope as
anything else on that lane.

Three other sentences carried the same unqualified claim and are fixed in the same
commit -- section 3.3's scope paragraph, its bypass paragraph and section 3.6's
kill-switch note -- because fixing only the flagged line leaves the siblings for the
next round. Applied in 4e0dde1.

@chenmingwei23

Copy link
Copy Markdown
Contributor Author

span=45dc661abaa4

  • Section 5's "exposes the first population and not the second" ignores a pinned jev. Correct, fixed.

With provider pinned jev, endpoint consent recorded and nudge_evidence
ungranted, _judge_authority computes jev_armed = False, judge_lane still
returns the Jev lane because the pin is honoured, and the lane is authorized on
jev_armed, so the tick is refused. That install exposes NEITHER population, not
the first one.

Section 5 now derives the exposure from what authorizes each lane rather than from
the grant alone: an install that granted the scope alone exposes the criteria-armed
population on the LLM lane and not the briefless one, and one that pinned jev
without the grant exposes neither. Applied in 4e0dde1.

This is the second round on this span. The first was the same class of error -- a
precondition narrowed in one section and left unchanged in section 5 -- so the fix
this time derives the sentence from the lane rule instead of restating a condition
that can drift from it again.

@chenmingwei23

Copy link
Copy Markdown
Contributor Author

span=45dc661abaa4

  • Section 3.6's flat "the destination matches those words" overstates it. Correct, fixed.

The claim is justified for the involuntary case only. A machine holding the scope
without the endpoint consent screens nothing rather than falling through to the
model provider, which is what I verified. But section 7 Q4 in the same commit says
the LLM lane screens a briefless loop where the owner holds the grant and pins
provider = llm, and there the evidence does go to the model provider while the
switch's words name the Jev endpoint.

Section 3.6 now says what is actually true of each case: no machine falls through
to the other destination, and an owner who pins provider = llm sends a briefless
loop's evidence to the model provider deliberately -- their own naming of a lane,
not this switch's words reaching further than they say. Applied in 4e0dde1.

@chenmingwei23

Copy link
Copy Markdown
Contributor Author

span=e23818f24c1e

  • The index claimed a 2026-09-24 refresh the front matter does not carry. Correct, fixed.

That README declares each document's own last-audited / audited-at front
matter the authoritative freshness signal for its row, so a provenance sentence
asserting a newer refresh contradicts the signal it defers to.

I fixed the sentence rather than the front matter, because bumping last-audited
to 2026-09-24 would claim a main-side audit I did not perform. What I did today was
re-read the IN-FLIGHT branch at d81bd0ec17 to settle section 7 Q4; the landed
LLM-lane claims were verified on 2026-09-23 at 2fdadc71ac and I did not re-verify
them against today's main, which has since moved to de2058659d.

The sentence now says only that, and points at the front matter for the main-side
audit instead of competing with it. Applied in 4e0dde1.

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 4e0dde1ff34ddfe6787b8b6744a08f08136f8b2b and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 4e0dde1

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 4e0dde1ff34ddfe6787b8b6744a08f08136f8b2b: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running readiness: passed Eligible automated validation passed for the current revision and removed readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention labels Sep 24, 2026
@bolichen97
bolichen97 enabled auto-merge (squash) September 24, 2026 06:51

@bolichen97 bolichen97 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approving. The doc never claims this rule is on main: the front matter stays partial, the header says the core sections are read at d81bd0ec17 on #12787, the README index line says the same, and Q4 itself is phrased as "PR D's tick reads ...". Reading an in-flight branch is this document's pre-existing convention, not something this PR introduces.

Q4's chosen answer is also the narrowest of the options it lists — the built-in brief applies only when endpoint consent AND this point's nudge_evidence scope both hold — so it adds no data flow. Q4 and the §8 PR D entry are both on the "Re-audit at merge" list, which is the hook the Design lane asked for.

@bolichen97
bolichen97 merged commit 9155a81 into main Sep 24, 2026
93 of 94 checks passed
@bolichen97
bolichen97 deleted the docs/rfc-wake-judge-q4 branch September 24, 2026 07:35
@github-actions github-actions Bot removed the readiness: passed Eligible automated validation passed for the current revision label Sep 24, 2026
wknight1 pushed a commit to wknight1/KiroCrew that referenced this pull request Oct 5, 2026
…kirodotdev#13234)

The built-in brief is authorized by the nudge_evidence grant read as the
narrower gate.judge_evidence_scope_granted -- the Decisions consent for the
configured endpoint plus this point's own scope -- asked before a lane is
picked, so it holds on both lanes. Those two conditions are exactly what arms
the Jev lane, so the default provider sends a briefless loop to the lane the
switch names, and a machine holding the scope alone screens nothing. gate =
false is never screened and judge: false opts out.

The rollout entry for the judge core states the decision instead of blocking
on it, the open-questions section answers Q4 with the same rule, and the two
sections that asserted the undecided consequence state the settled one. Q5
stays open. The index row and its provenance line carry the refreshed reading.
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.

2 participants