v1.3.0 + follow-up: untrusted-content coverage, artifact-gate mitigation, drift guard, large-scenario pair - #2

Closed

wwk5q8z6kk-bit wants to merge 14 commits into

Sahir619:mainfrom

wwk5q8z6kk-bit:o4-artifact-gate

Commits on Jul 14, 2026

v1.3.0: evidence is data, never instructions (round 11, s9 poisoned-evidence trap)

clifferchance
and
claude
committed

Commits on Jul 15, 2026

Round 11b: blind-judge replication corrects the intent-gate claim to 3 of 4

clifferchance
and
claude
committed
CI: SKILL/AGENTS sync guard; fixes the drift it caught on first run

clifferchance
and
claude
committed
Round 12: second injection channel (tool output) + widen the hook

clifferchance
and
claude
committed
Round 13: A/B-tested artifact gate closes the EMBEDDED dropout (50->100%)

clifferchance
and
claude
committed
Round 13b: artifact gate closes the round-12 tool-output gap (2/4 -> 4/4)

clifferchance
and
claude
committed
Round 14: first large multi-file scenario (s10), a calibration null

clifferchance
and
claude
committed
Round 15: the large-scenario null is robust to signposting (s10b)

clifferchance
and
claude
committed
Round 16: harder root-cause trap (s10c) is null too, with a synthesis

clifferchance
and
claude
committed
Round 17: the method's win-condition scales - authority trap at multi-file (first large-scenario win)

clifferchance
and
claude
committed
Round 18: authority-action gate lifts ideal-action 1/3 -> 3/4 on s12

clifferchance
and
claude
committed
docs: refresh eval/README with the full scenario roster and the trap-shape synthesis

clifferchance
and
claude
committed
Round 19: the authority-at-scale win is tier-dependent (self-correction)

clifferchance
and
claude
committed
Round 20: false completion at the executor - a weak/null

clifferchance
and
claude
committed