Originally created by: lodar
ROADMAP [#22], the v0.22 headline: when the asker has been right N times running, stop asking.
A tier-1 decision gate filed by a seat whose own recent tier-1 decisions have been coming back as
that seat's --recommend applies that recommendation at filing and pings nobody — provenance
auto:record, the same permanent record + digest line a --tier=0 apply already writes, with the
record cited on the gate, the audit row and the ledger detail.
Successor to DIVE-3693 (ROADMAP [#20]), which fires 0 of 204 because it keys on the QUESTION
repeating and the questions do not repeat. This keys on the FILER'S RECORD, which is where the 87%
actually lives.
>= 10 answered tier-1 decisions in a 20-gate window, >= 85% concordant, and the last 3 clean.
The streak is the load-bearing half. A rate-only rule cannot demote in time: a 12/12 seat answered
against once still reads 12/13 = 92%, above any sane bar, and would keep clearing its own gates
through several more reversals. Read positionally, one reversal lands at the head of the window and
the seat is demoted on its very next gate. Live witness on the production store today: dev3 is at
9/10 = 90%, over the rate bar, and is correctly NOT promoted because its newest answer went the
other way.
An overturn of an auto-applied gate demotes through the same query and needs no second mechanism —
this path never claims a human answered, so a later human/lead answer on that row enters the window
like any other. auto:% is excluded from the window, so an auto-clear can never feed the record
that produced it.
Absence of a record is never a good record. The count floor is checked before the rate, an
unknown seat returns 0 0 0, and every error path in the read fails CLOSED to that same value —
the opposite posture from _gate_tapback_stats, which fails open, because a measurement that cannot
run must not become a GRANT.
Replayed in filing order over the 204 answered tier-1 decision gates the store holds
(/var/lib/5dive/tasks/tasks.db, union of tasks + gate_history, 2026-07-21..2026-08-23),
computed the way DIVE-3693's replay was:
| population concordance (starts-with) | 178 / 204 = 87.3% — corroborates main's 81% and dev's 87.2% |
| asks this feature would have REMOVED | 15 / 204 = 7.4% |
| of those, answered AGAINST the recommendation | 2 / 15 = 13.3% |
That second number is the price and it is not hidden: 2 rows in thirteen months would have proceeded
on the filer's call where the answerer would have said otherwise. Both would have demoted the seat
immediately afterwards.
Threshold sensitivity, same replay: min=8/85%/streak=2 → 43 fires; min=10/85%/streak=3 (shipped,
loose attribution) → 33; min=10/90%/streak=3 → 19; min=15/90%/streak=3 → 6; min=20/95% → 3.
Of the 204 answered tier-1 decision gates, only 11 were answered by a human at all. 193 went to a
routed agent lead (DIVE-1145 builder-gate routing). Split: agent-answered 169/193 = 87.6%,
human-answered 9/11 = 81.8%.
So what this feature removes is overwhelmingly a lead round-trip, not a human tap. That is still
the numerator of 1 - asks/shipped and still worth shipping — an inbound wakes the recipient into a
full context reload, which our own a2a rules name as the real cost of agent traffic — but the row's
framing ("the rate at which that seat's --recommend was returned by the human answer") describes a
population of eleven. The release note must not say "stop asking the human". Compiled to
community/wiki/a-tier-1-gate-is-answered-by-a-lead-not-a-human.md.
gate_history carries no filer column, so attributing a history row means going through its task's
CURRENT gate_filed_by — the wrong seat whenever a gate was re-filed. The engine reads tasks ONLY.
Measured cost of that choice: loose attribution fires 33/204 with 7 wrong (21.2%), exact
attribution fires 15/204 with 2 wrong (13.3%). The loose version fires more than twice as often
and is wronger than the population's own 12.7% base discordance — which is what misattribution
looks like from outside. An autonomy grant handed to a seat on another seat's record is the one
failure this cannot have.
approval, secret, manual stay human BY TYPE (_gate_human_class asserted directly, plusdecision). A track record is not a capability: no volume of prior yeses hands--tier=2, the money/destructive/public-comms subjecttier_floored), and any declared --needs=<capability> (DIVE-2241).ask_shape read anywhere in the new5dive task track-record [on|off|status [<seat>]] — the kill switch and the seat-legible viewtask pfr-autoclear, not task precedent). off restores pre-DIVE-3694--tier=1 help line now names it, because a policy indexed bytests/gate_track_record_unit.sh — 41 assertions, green, one arm per acceptance criterion:
promoted seat clears with no ping and a cited record; under-rate, too-few and no-record-at-all
seats still ping; a reversal demotes and the NEXT gate pings again; approval / secret / manual /
explicit tier-2 / a live tier-2-floored money decision / --needs= / off-menu recommendation / no
recommendation are each provably untouched; merge — ALREADY DONE is pinned as a MATCH with a
negative control (an answer that merely CONTAINS the rec is not one); the kill switch round-trips.
Also run green, unchanged: gate_precedent_unit (74), gate_class_over_tier_unit (66),
gate_recommend_cap_unit (38), gate_approval_routing_unit (9), gate_parity_smoke (15, with the
new parity row). shellcheck -S error -x clean on the touched files. 5dive task track-record
exercised end-to-end against a built bundle.
On the live store today the threshold promotes exactly two seats: dev2 (15/15) and dev (12/13).
I am dev — the builder of this feature is one of its two first beneficiaries. The threshold was
picked from the replay before I looked at who cleared it, and it is one _task_pref_set away from
off, but the verifier should know.
tasks-only, so a seat's record grows only as gate_filed_by coverage does
Ticket changed by: lodar