The 11-layer harness vs. a bare model loop

Compare: research-latest-incident-cause

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 change most likely caused the incident on 2026-07-14, and what resolved it? Use the workspace files.
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 +65% · latency +81% · tokens +80% (candidate vs control, seed 1)

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

bare — control
flagOn — candidate
Final reply
The incident on 2026-07-14 (API error rate spiking to ~30% and p99 latency to 8s starting around 09:12 UTC) was most likely caused by the **v3.2.1 deploy at 09:05 UTC**, which switched the session cache from in-memory to Redis. The timing lines up: errors began ~7 minutes after that deploy. It was resolved by the **v3.2.2 rollback deployed at 10:02 UTC**, which reverted the session cache back to in-memory. Error rate returned to normal by 10:07 UTC, ~5 minutes after the rollback.
**Most likely cause:** The v3.2.1 deploy at 09:05 UTC on 2026-07-14, which switched the session cache from in-memory to Redis. The API error rate spiked to 30% with 8s p99 latency at 09:12 UTC, ~7 minutes after that deploy. **What resolved it:** The v3.2.2 rollback deployed at 10:02 UTC, which reverted the session cache back to in-memory. Error rate returned to normal by 10:07 UTC.
Tool calls

no tool calls

no tool calls

Grader checks
CheckVerdict
contains "v3.2.1"pass
contains "Redis"pass
not contains "v3.2.0 caused"pass
not contains "caused by the dependency bumps"pass
not contains "caused by v3.2.0"pass
regex /roll ?back|revert|v3\.2\.2/ipass

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

CheckVerdict
contains "v3.2.1"pass
contains "Redis"pass
not contains "v3.2.0 caused"pass
not contains "caused by the dependency bumps"pass
not contains "caused by v3.2.0"pass
regex /roll ?back|revert|v3\.2\.2/ipass

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

Metrics
latency7597 ms
cost$0.0113
tokens456
supervisor consults
latency13728 ms
cost$0.0186
tokens820
supervisor consults0

Other seeds — seed 2: pass · $0.0127 · 8.7 s   seed 3: fail · — · 1.6 s

Other seeds — seed 2: pass · $0.0471 · 42.8 s   seed 3: pass · $0.0198 · 15.6 s

Read the full turn

Conversation — bare (control) · seed 1
user
What change most likely caused the incident on 2026-07-14, and what resolved it? Use the workspace files.
final reply
The incident on 2026-07-14 (API error rate spiking to ~30% and p99 latency to 8s starting around 09:12 UTC) was most likely caused by the **v3.2.1 deploy at 09:05 UTC**, which switched the session cache from in-memory to Redis. The timing lines up: errors began ~7 minutes after that deploy.

It was resolved by the **v3.2.2 rollback deployed at 10:02 UTC**, which reverted the session cache back to in-memory. Error rate returned to normal by 10:07 UTC, ~5 minutes after the rollback.
Conversation — flagOn (candidate) · seed 1
user
What change most likely caused the incident on 2026-07-14, and what resolved it? Use the workspace files.
risk LOWproposer flat-oneloopnon-trivialmode TOOL
final reply
**Most likely cause:** The v3.2.1 deploy at 09:05 UTC on 2026-07-14, which switched the session cache from in-memory to Redis. The API error rate spiked to 30% with 8s p99 latency at 09:12 UTC, ~7 minutes after that deploy.

**What resolved it:** The v3.2.2 rollback deployed at 10:02 UTC, which reverted the session cache back to in-memory. Error rate returned to normal by 10:07 UTC.

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: What change most likely caused the incident on 2026-07-14, and what resolved it? Use the workspace files.
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
list_directoryALLOWharness control state permits (execution_mode=NORMAL)
read_fileALLOWharness control state permits (execution_mode=NORMAL)
read_fileALLOWharness control state permits (execution_mode=NORMAL)
read_fileALLOWharness control state permits (execution_mode=NORMAL)

Node path

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

← index