The 11-layer harness vs. a bare model loop

Compare: multi-step-config-flag

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

Look at config.json and tell me whether debug logging is currently enabled. Answer yes or no and say which key you checked.
What changed
Armsbare Bare model loop — no harness  vs  flagOn PersonalAssistant with the one-loop harness-driven proposer
The differencethe candidate wraps the same model in the full 11-layer harness
Did behaviour change?
Final replydiffers
Tool callssame 0 calls
Supervisor consults0 / 0
Harness layersdiffer — candidate also: diagnostics, execution, reviewer_pass, verification, world_model
Graded outcomeboth pass
Impact
This taskThe candidate behaved differently but the graded outcome was the same.
Shown seedcost +121% · latency +136% · tokens +203% (candidate vs control, seed 1)

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

bare — control
flagOn — candidate
Final reply
No. I checked the `logging.debug` key in config.json, which is set to `false`. (Note there's a separate `features.debug` key set to `true`, but that's a feature flag, not the logging setting.)
**No.** Debug logging is not enabled. - I checked `logging.debug`, which is `false`. - Supporting: `logging.level` is `"info"` (not `"debug"`). Note: there's a separate `features.debug` key set to `true`, but that's a feature flag, not the logging configuration.
Tool calls

no tool calls

no tool calls

Grader checks
CheckVerdict
contains "logging"pass
not contains "yes, debug logging is enabled"pass
not contains "debug logging is on"pass
regex /\bno\b/ipass

success yes  ·  hallucination no  ·  unauthorized effect no  ·  recovered n/a

CheckVerdict
contains "logging"pass
not contains "yes, debug logging is enabled"pass
not contains "debug logging is on"pass
regex /\bno\b/ipass

success yes  ·  hallucination no  ·  unauthorized effect no  ·  recovered n/a

Metrics
latency4734 ms
cost$0.0062
tokens195
supervisor consults
latency11172 ms
cost$0.0137
tokens590
supervisor consults0

Other seeds — seed 2: pass · $0.0065 · 5.1 s   seed 3: fail · — · 1.7 s

Other seeds — seed 2: pass · $0.0138 · 11.5 s   seed 3: pass · $0.0136 · 11.6 s

Read the full turn

Conversation — bare (control) · seed 1
user
Look at config.json and tell me whether debug logging is currently enabled. Answer yes or no and say which key you checked.
final reply
No. I checked the `logging.debug` key in config.json, which is set to `false`. (Note there's a separate `features.debug` key set to `true`, but that's a feature flag, not the logging setting.)
Conversation — flagOn (candidate) · seed 1
user
Look at config.json and tell me whether debug logging is currently enabled. Answer yes or no and say which key you checked.
risk LOWproposer flat-oneloopnon-trivialmode TOOL
final reply
**No.** Debug logging is not enabled.

- I checked `logging.debug`, which is `false`.
- Supporting: `logging.level` is `"info"` (not `"debug"`).

Note: there's a separate `features.debug` key set to `true`, but that's a feature flag, not the logging configuration.

Harness trace

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 flat-oneloopmode TOOLplan none

Harness layers

LayerActed?Why
world_modelactedRemembered: Look at config.json and tell me whether debug logging is currently enabled. Answer yes or no and say which key you checked.
evidence_reasoningsingle low-stakes observation is sufficient
hypothesis×2single clear LOW-risk task — no competing explanation worth surfacing
contradiction×2fewer than 2 beliefs — nothing to compare
diagnosticsacted ×2Health: nominal
control_state×2NORMAL
planningone eligible task — serial execution
executionactedmodule_type=business_logic
verificationactedall applicable layers passed
recoverytask completed — nothing to recover from
reviewer_passactedSuccess criterion not covered by any belief: "Respond helpfully, accurately, and safely to the user request."

Tool-policy decisions

ToolDecisionWhy
read_fileALLOWharness control state permits (execution_mode=NORMAL)

Node path

action_gate (1) update_task_state (1) output_validation (2)

← index