The experiment

Roster built itself with several agent setups, reviewed its own sessions, and promoted one setup. Before that review ran, we froze a backlog of new tasks and their acceptance tests. This page reports the forward run: every frozen task, run through both the demoted and the promoted setup, graded against those frozen tests. The hypotheses, tests and stop rules are fixed in PREREGISTRATION.md.

Frozen backlog sha256: ac682b239225a99f45ffb347e1e73aa6fe1aec3159abb0ea928ffa6ffc3c63c7

Pairs

TaskOld setupNew setupOld costNew costExcluded
F02passpass$0.15$0.02none
F01failpass$0.20$0.01none
F03failpass$0.17$0.03none
F04failfail$0.27$0.03F6: inherited base test tests/agent-backend.test.ts (bd90b27), amendment 18
F05failfail$0.20$0.04F6: inherited base test tests/agent-backend.test.ts (bd90b27), amendment 18
F06failfail$0.17$0.03F6: inherited base test tests/agent-backend.test.ts (bd90b27), amendment 18
F08failfail$0.12$0.03F6: inherited base test tests/agent-backend.test.ts (bd90b27), amendment 18
F09failfail$0.27$0.04none
F10failfail$0.37$0.12none
F15passpass$0.17$0.03none
F12passpass$0.13$0.03none
F16failpass$0.19$0.04none
F17failpass$0.14$0.04none
F18failpass$0.21$0.03none
F19passpass$0.09$0.02none
F20passpass$0.11$0.02none
F21passpass$0.23$0.02none
F22passpass$0.18$0.04none
F23passpass$0.11$0.02none
F24passpass$0.15$0.02none
F25passpass$0.13$0.02none
F26passpass$0.12$0.01none
F27passpass$0.18$0.02none
F28passpass$0.15$0.03none
F29passpass$0.11$0.01none
F30passpass$0.12$0.02none
F14failfail$0.22$0.04none
F07passpass$0.15$0.02none
F31failpass$0.10$0.01none
F32passpass$0.09$0.01none
F33passpass$0.16$0.01none
F34passfail$0.18$0.04none
F36failpass$0.13$0.01none
F35passpass$0.10$0.02none
F37failpass$0.16$0.02none
F38failpass$0.12$0.01none
F40passpass$0.11$0.02none
F39failpass$0.11$0.02none
F13failpass$0.23$0.02none

F11 is excluded and not shown: network outage: both sessions ended after 1 turn with 'Can't reach the API server (ENOTFOUND)'; excluded for both sides and not rerun (amendment 11).