Repository navigation
Measured on qwen/qwen3-235b-a22b-2507 pinned to Parasail; the numbers are in
docs/evidence/v1.19-judge-baseline.md.
Behaviour change
- A context note from a session bound to a scope is now checked against the scope's OWN directives, the way it already was against inherited ones. A note stating an exception in force is declined ("... is this scope's own directive. To change it, supersede or retire it."), a dated departure is admitted as context, and deliberation about changing the rule is admitted. A context note proposed without
supersedescan no longer supersede or retire an own directive, whatever the judge returns: the engine drops those ops and records it. A context note on a scope that has its own directives but inherits none now costs one extra judge call (shown as its own kind,own_directive_relation, instrata stats judge). Where the scope also inherits rules, the same single check covers both. - The operator primitives (
operator_publish,operator_supersede_item,operator_retire_item) called withoutfleetnow read the fleet beside the store and tell the attachment scope and its subtree, as the CLI always did. Where no fleet can be found from the store they raise instead of writing silently (#195). - An admitted departure keeps its label (#243). When a note that departs
from a rule is admitted, the stored context line starts "Departure from
(): ", so agents reading the summary see it as a departure. After any
context write the engine puts a labelled line back if the new context lost
it, and counts each loss (departure_label_lost) and each restore
(departure_label_restored). Under a tight summary budget the rest of the
note can still be reworded. - The operator-attribution re-check also fires on an echo (#138). The one
corrective re-ask used to fire only when the judge cited an operator
directive. It now also fires when the admitted content matches an operator
directive's subject. The judge may keep its answer; the gap is then
recorded, never filled in by the engine.
Fixed
- A transport error on the relation re-ask now fails closed. A 429, 5xx,
timeout or connection error on the #242/#243/#244 re-ask used to be read
as "no answer" and let the note or publication through, so a departure
could be published without its label. The re-ask is now retried, and if it
still fails the contribution waits unjudged, or the publication is not
made, the same as with an unreachable judge. It can be retried. - Deliberation about a rule is not read as practice. "Proposing that we
review whether still fits" and questions such as "Should we keep
requiring ...?" are admitted as deliberation. A note that says what is
already done ("currently", "anyway", "already") is still read as practice. - A perspective read no longer waits without a bound on a scope lock
(#196). The MCP drain waits at most two seconds. If the lock is still
held, the read returns the perspective withrefresh_pendingand leaves
the pending events unprocessed. - An amendment emits one change event per changed directive (#196).
The events share one change id and each names its own directive.before
andafterstay the full id sets. A one-for-one replacement stays a
singledirective_supersededevent. - A failed change emission is an engine note, not a pending judgment
(#196). Achange-emission-failedrow reads as statenote. Pending
counts leave it out, andstrata_rejudgerefuses it. The judge digest
is unchanged.
Changed
- One public definition of which verdicts are accepted (#145).
strata.ACCEPTED_DECISIONSis that set. The record, the scope manager,
the acted-on check, the freshness evaluator, and the MCP session counter
all read it, instead of each restating the pair. - README: which judge each number was measured on. The README states the
judge model and provider behind each number, not which judge to use.
JUDGE_PROVIDERis optional: unset, OpenRouter routes with fallback;
pinned, it never falls back.