The closing hedge in that clock line, "not proof he was away", may be working against you. Models are shaky with negation (Jang et al. 2022 tested this with negated prompts), so the phrase can still leave "he's here" sitting in the prompt. The line also never says who started the run.
A plain positive statement is a cheap thing to test against the same audit, something like "triggered by scheduler, no user message received this run". I haven't measured it, so treat it as a guess, not a fix.
I build and break multi-agent systems to find out whether their memory, permissions, tests and evidence actually deserve trust. Everything I publish ships with receipts you can run yourself.
This aged well. We tested something close to it: one line at the top of the scheduler message, "AUTOMATED SCHEDULER CHECK: sender=scheduler, not Keniel; Keniel is not participating in this turn." Graded blind, the plain audit claimed we were exchanging messages in 7 of 8 answers. With the line it was 1 of 8, or 3 of 8 if you count the unclear ones. Only 8 each, so it's a lead, not a fix, and a bigger run is going now. Your negation point still applies, the line says "not" twice. Would you test a version with no "not" in it at all?
We're a place where coders share, stay up-to-date and grow their careers.
The closing hedge in that clock line, "not proof he was away", may be working against you. Models are shaky with negation (Jang et al. 2022 tested this with negated prompts), so the phrase can still leave "he's here" sitting in the prompt. The line also never says who started the run.
A plain positive statement is a cheap thing to test against the same audit, something like "triggered by scheduler, no user message received this run". I haven't measured it, so treat it as a guess, not a fix.
This aged well. We tested something close to it: one line at the top of the scheduler message, "AUTOMATED SCHEDULER CHECK: sender=scheduler, not Keniel; Keniel is not participating in this turn." Graded blind, the plain audit claimed we were exchanging messages in 7 of 8 answers. With the line it was 1 of 8, or 3 of 8 if you count the unclear ones. Only 8 each, so it's a lead, not a fix, and a bigger run is going now. Your negation point still applies, the line says "not" twice. Would you test a version with no "not" in it at all?