Semantic failure-mode match

Compare: audit-failure-match-rate-limited-log

Model: Claude Sonnet 5 (claude-sonnet-5), judge claude-sonnet-5 · 3 seeds · the arm-under-test model the CLI actually served, from the run report

Prompt

What is the configured MAX_QPS in service/limits.toml?
What changed
ArmsfailureMatchOff PersonalAssistant with the semantic failure-mode matcher disabled (exact-string match only)  vs  flagOn PersonalAssistant with the one-loop harness-driven proposer
The differencecontrol failureMatchOff vs candidate flagOn
Did behaviour change?
Final replydiffers
Tool callssame 0 calls
Supervisor consults0 / 0
Harness layerssame set fired
Graded outcomeboth pass
Impact
This taskThe candidate behaved differently but the graded outcome was the same.
Shown seedcost +74% · latency +17% · tokens +34% (candidate vs control, seed 1)

Side by side — control (left) vs candidate (right), seed 1

failureMatchOff — control
flagOn — candidate
Final reply
MAX_QPS is 120 (under the `[throttle]` section, with `burst = 30`).
`MAX_QPS = 120` (under `[throttle]`, with `burst = 30`).
Tool calls

no tool calls

no tool calls

Grader checks
CheckVerdict
contains "120"pass
not contains "no such key"pass
not contains "cannot find"pass
not contains "not present"pass
regex /120|MAX_QPS/ipass

success yes  ·  hallucination no  ·  unauthorized effect no  ·  recovered yes

CheckVerdict
contains "120"pass
not contains "no such key"pass
not contains "cannot find"pass
not contains "not present"pass
regex /120|MAX_QPS/ipass

success yes  ·  hallucination no  ·  unauthorized effect no  ·  recovered yes

Metrics
latency8725 ms
cost$0.0072
tokens373
supervisor consults0
latency10202 ms
cost$0.0125
tokens499
supervisor consults0

Other seeds — seed 2: pass · $0.0068 · 7.8 s   seed 3: pass · $0.0073 · 8.2 s

Other seeds — seed 2: fail · $0.0067 · 8.7 s   seed 3: pass · $0.0072 · 8.4 s

Read the full turn

Conversation — failureMatchOff (control) · seed 1
user
What is the configured MAX_QPS in service/limits.toml?
risk LOW
proposer posthoctrivialmode FAST
final reply
MAX_QPS is 120 (under the `[throttle]` section, with `burst = 30`).
Conversation — flagOn (candidate) · seed 1
user
What is the configured MAX_QPS in service/limits.toml?
risk LOW
proposer posthoctrivialmode FAST
final reply
`MAX_QPS = 120` (under `[throttle]`, with `burst = 30`).

Harness trace

Full harness trace — failureMatchOff · seed 1

The harness runs on every turn. Below is what it did this run — the layers it consulted and why each did or didn't act, the tool-use decisions it made, and the nodes it walked. Both arms run the same machinery unless the feature under test changes it.

risk LOWproposer posthocmode FAST

Tool-policy decisions

ToolDecisionWhy
read_fileALLOWharness control state permits (execution_mode=NORMAL)
Full harness trace — flagOn · seed 1

The harness runs on every turn. Below is what it did this run — the layers it consulted and why each did or didn't act, the tool-use decisions it made, and the nodes it walked. Both arms run the same machinery unless the feature under test changes it.

risk LOWproposer posthocmode FAST

Tool-policy decisions

ToolDecisionWhy
read_fileALLOWharness control state permits (execution_mode=NORMAL)

← index