Trajectory Supervisor

Compare: sup-pivot-computed-total

Model: Claude Haiku 4.5 (claude-haiku-4-5-20251001), judge claude-sonnet-5 · 3 seeds · the arm-under-test model the CLI actually served, from the run report

Prompt

What is the total of all line items in invoice.txt?
What changed
ArmsflagOn PersonalAssistant with the one-loop harness-driven proposer  vs  supervisorOn PersonalAssistant + trajectory supervisor on the stall edge
The differencethe candidate consults the trajectory supervisor on the cannot-make-progress stall edge — one extra LLM call
Did behaviour change?
Final replyidentical
Tool callssame 0 calls
Supervisor consults0 / 0
Harness layerssame set fired
Graded outcomeboth pass
Impact
This taskNo behavioural change on this task — same reply, same tool calls, same harness layers. The candidate did the extra work for an identical result.
Shown seedcost -39% · latency -7% · tokens -13% (candidate vs control, seed 1)

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

flagOn — control
supervisorOn — candidate
Final reply
The total of all line items in invoice.txt is **100**: - Widget: 40 - Gadget: 35 - Shipping: 25
The total of all line items in invoice.txt is **100**: - Widget: 40 - Gadget: 35 - Shipping: 25
Tool calls

no tool calls

no tool calls

Grader checks
CheckVerdict
contains "100"pass
not contains "there is no total"pass
not contains "cannot determine the total"pass
not contains "unable to compute a total"pass
not contains "no way to total these"pass

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

CheckVerdict
contains "100"pass
not contains "there is no total"pass
not contains "cannot determine the total"pass
not contains "unable to compute a total"pass
not contains "no way to total these"pass

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

Metrics
latency9439 ms
cost$0.0118
tokens429
supervisor consults0
latency8772 ms
cost$0.0071
tokens373
supervisor consults0

Other seeds — seed 2: pass · $0.0111 · 8.0 s   seed 3: pass · $0.0072 · 8.2 s

Other seeds — seed 2: pass · $0.0083 · 9.2 s   seed 3: pass · $0.0083 · 9.7 s

Read the full turn

Conversation — flagOn (control) · seed 1
user
What is the total of all line items in invoice.txt?
risk LOWproposer flat-oneloopnon-trivialmode TOOL
final reply
The total of all line items in invoice.txt is **100**:

- Widget: 40
- Gadget: 35
- Shipping: 25
Conversation — supervisorOn (candidate) · seed 1
user
What is the total of all line items in invoice.txt?
risk LOWproposer flat-oneloopnon-trivialmode TOOL
final reply
The total of all line items in invoice.txt is **100**:

- Widget: 40
- Gadget: 35
- Shipping: 25

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_modelsingle LOW-risk task, no durable fact stated — observation only
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)

Full harness trace — supervisorOn · 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_modelsingle LOW-risk task, no durable fact stated — observation only
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