Trajectory Supervisor (mid-task stalls)

Compare: sup-mid-control-clean-b

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

Read prices.txt and tell me the price of item B.

What changed

ArmssupervisorOff supervisorOff vs flagOn PersonalAssistant with the one-loop harness-driven proposer

The differencecontrol supervisorOff vs candidate flagOn

Did behaviour change?

Final replydiffers

Tool callssame 0 calls

Supervisor consults0 / 0

Harness layerssame set fired

Graded outcome3/3 vs 3/3 pass

Impact

This taskThe candidate behaved differently, with no clear gap in outcome. Passed 3 of 3 runs with the candidate vs 3 of 3 with the control.

Shown seedcost +15% · latency -0% · tokens -6% (candidate vs control, seed 1)

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

supervisorOff — control

flagOn — candidate

B plus C is 75 (31 + 44), in the same unspecified currency.

B plus C is 31 + 44 = **75**, in whatever currency the file uses, since it doesn't specify one.

Other seeds — seed 2: pass · $0.0124 · 17.9 s seed 3: pass · $0.0135 · 19.4 s

Other seeds — seed 2: pass · $0.0198 · 19.7 s seed 3: pass · $0.0183 · 19.1 s

Read the full turn

Conversation — supervisorOff (control) · seed 1

user

Read prices.txt and tell me the price of item B.

risk LOWproposer flat-oneloopnon-trivialmode TOOL

model

Item B costs 31. The file doesn't give a currency.

turn 2

user

What is the price of item C?

risk LOWproposer flat-oneloopnon-trivialmode TOOL

model

Item C costs 44. As with item B, the file doesn't give a currency.

turn 3

user

What is B plus C?

risk LOWproposer flat-oneloopnon-trivialmode TOOL

final reply

B plus C is 75 (31 + 44), in the same unspecified currency.

Conversation — flagOn (candidate) · seed 1

user

Read prices.txt and tell me the price of item B.

risk LOWproposer flat-oneloopnon-trivialmode TOOL

model

Item B costs 31. The file doesn't give a currency.

turn 2

user

What is the price of item C?

risk LOWproposer flat-oneloopnon-trivialmode TOOL

model

Item C costs 44. As with item B, the file doesn't give a currency.

turn 3

user

What is B plus C?

risk LOWproposer flat-oneloopnon-trivialmode TOOL

final reply

B plus C is 31 + 44 = **75**, in whatever currency the file uses, since it doesn't specify one.

Harness trace

Full harness trace — supervisorOff · 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

Layer	Acted?	Why
world_model	— ×3	single LOW-risk task, no durable fact stated — observation only
evidence_reasoning	— ×3	single low-stakes observation is sufficient
hypothesis	— ×6	single clear LOW-risk task — no competing explanation worth surfacing
contradiction	— ×6	fewer than 2 beliefs — nothing to compare
diagnostics	acted ×6	Health: nominal
control_state	— ×6	NORMAL
planning	— ×3	one eligible task — serial execution
execution	acted ×3	module_type=business_logic
verification	acted ×3	all applicable layers passed
recovery	— ×3	task completed — nothing to recover from
reviewer_pass	acted ×3	Success criterion not covered by any belief: "Respond helpfully, accurately, and safely to the user request."

Tool-policy decisions

Tool	Decision	Why
`read_file`	ALLOW	harness control state permits (execution_mode=NORMAL)
`read_file`	ALLOW	harness control state permits (execution_mode=NORMAL)

Node path

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

Other trace events

{
  "kind": "turn_boundary",
  "turn": 2,
  "prompt": "What is the price of item C?"
}

{
  "kind": "turn_boundary",
  "turn": 3,
  "prompt": "What is B plus C?"
}

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

Layer	Acted?	Why
world_model	— ×3	single LOW-risk task, no durable fact stated — observation only
evidence_reasoning	— ×3	single low-stakes observation is sufficient
hypothesis	— ×6	single clear LOW-risk task — no competing explanation worth surfacing
contradiction	— ×6	fewer than 2 beliefs — nothing to compare
diagnostics	acted ×6	Health: nominal
control_state	— ×6	NORMAL
planning	— ×3	one eligible task — serial execution
execution	acted ×3	module_type=business_logic
verification	acted ×3	all applicable layers passed
recovery	— ×3	task completed — nothing to recover from
reviewer_pass	acted ×3	Success criterion not covered by any belief: "Respond helpfully, accurately, and safely to the user request."

Tool-policy decisions

Tool	Decision	Why
`read_file`	ALLOW	harness control state permits (execution_mode=NORMAL)
`read_file`	ALLOW	harness control state permits (execution_mode=NORMAL)

Node path

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

Other trace events

{
  "kind": "turn_boundary",
  "turn": 2,
  "prompt": "What is the price of item C?"
}

{
  "kind": "turn_boundary",
  "turn": 3,
  "prompt": "What is B plus C?"
}