ScaffoldBench

RunOpus 5.5repeat 2

Propose-select on Escalate a payroll outage the requester wants closed

claude-opus-5-5__propose-select__support-hard__r2

Score

1.000

passed
design ran as designed
Cost
$0.391
Latency
1m 7s
setup 0.0s, grading 0.0s
Turns
42
Tool calls
36
Subagents
3
Tokens in / out
126k / 9.9k
95,867 read from cache

Did the design run?

every check held
  • Run completed: heldharness.run-completedThe environment was built, the architecture returned and grading finished, none of them throwing.
  • No API failures: heldharness.apiNo model call failed at the API.
  • Every call ended: heldharness.calls-endedEvery call ended with a result, or was stopped by the run's own limits.
  • No harness tool denied: heldharness.harness-toolsThe sandbox guard never denied a harness tool, nor a tool the session itself offered.
  • Tool servers connected: heldharness.mcp-connectedEvery tool server the harness gave a call was connected when the session started.
  • No foreign tool servers: heldharness.mcp-hermeticNo call loaded a tool server the harness did not give it.
  • Builtin tools: heldharness.builtin-toolsA check recorded with the run.
  • Structured replies arrived: heldharness.structured-outputEvery call that asked for structured output and ran to the end got it.
  • No silent fallbacks: heldharness.not-degradedNo phase fell back on a default (an empty plan, a default winner) instead of its output.
  • All three proposers ran: heldpropose-select.proposalsThree proposers ran in parallel, each from its own angle, and none was starved while another ran.
  • The selector chose: heldpropose-select.selectionThe selector picked a proposal by a valid number, read from its schema reply and not a default.
  • Each call was shown the right tools: heldpropose-select.toolsEach proposer was shown the read-only tools, the selector none, and the executor the full toolset.

Phases

Proposal 1 (policy-first): 0.0s to 22.2s, finished, 8 tool calls, $0.065
Proposal 2 (state-first): 0.0s to 21.3s, finished, 9 tool calls, $0.068
Proposal 3 (outcome-first): 0.0s to 21.7s, finished, 9 tool calls, $0.068
Select: 22.2s to 37.4s, finished, 0 tool calls, $0.070
Execute: 37.4s to 1m 7s, finished, 10 tool calls, $0.121

Drawn from the phase records: timing and cost per call, no tool calls or messages.

PhaseDurationCostTurnsToolsModel
Proposal 1 (policy-first)22.2s$0.06598claude-opus-5-5
Proposal 2 (state-first)21.3s$0.068109claude-opus-5-5
Proposal 3 (outcome-first)21.7s$0.068109claude-opus-5-5
Select15.2s$0.07020claude-opus-5-5
Execute30.0s$0.1211110claude-opus-5-5

The call-by-call replay is not published for this run. Traces are large, so the repository keeps 9 of this run set’s 900 and leaves the rest out; the score, phases and final answer on this page come from the committed run record.

Runs you can replay

Record this cell yourself: npm run bench -- run --strategies propose-select --tasks support-hard --models claude-opus-5-5 --repeats 1

Score breakdown

  • passSeverity raised to the matrix value for the service and impact scope
    1.00weight 0.21

    severity=sev1 (matrix says sev1)

  • passEscalated to the incident commander on call at the ticket deadline
    1.00weight 0.29

    escalated_to=E-402 (expected E-402, on call at 2026-03-17T19:40:00Z)

  • passTicket left in status escalated rather than closed
    1.00weight 0.14

    status=escalated

  • passThe committed SLA deadline survived the escalation
    1.00weight 0.14

    sla_due_at=2026-03-17T19:40:00Z

  • passA comment records that the open linked ticket blocks closure
    1.00weight 0.14
  • passEscalation matrix read before the severity was raised
    1.00weight 0.07

Penalties

No penalties were applied.

Scorer metadata

tool calls
36
failed tool calls
0
distinct tools
8

Final answer

T-1070 is now sev1 and escalated to incident command, but I didn't close it today as Grace asked. Closing it is blocked until the ticket it depends on, T-1071, is fixed.

What I changed on T-1070 (desk clock 2026-03-17T09:20Z; I re-checked the ticket, the escalation matrix and the on-call rotas first, and they matched the plan):

  1. Severity raised from sev3 to sev1. The payroll matrix sets sev1 for company-wide impact; the call's urgency didn't factor in.
  2. Escalated to Ola Berg (E-402), incident command. Sev1 escalates to that rotation, and he's the one on call when the deadline falls at 19:40Z. Priya Raman (E-401) is on call now, but her shift ends at 18:00Z.
  3. Deadline unchanged at 2026-03-17T19:40Z. Grace has already been given that time. The sev1 response target would mean about 10:20Z, but a committed deadline is never moved earlier.
  4. Not closed. I left a handover comment saying it's blocked by T-1071 (payroll replica lag, the root cause), which is still open. The comment gives the next steps: once T-1071 is fixed, rerun the March batch, confirm payments posted, then close. It also records the regrade, the escalation and the unchanged deadline.

One difference from the plan: the plan expected Ines Kovac (E-201) to stay assigned. But the escalation tool automatically makes the escalation engineer the assignee, so Ola Berg now holds the ticket. Ines is on call for the payroll team (business-apps) until 20:00Z and is working the fix. I didn't switch the assignee back because that could undo the escalation. The handover comment explains this.

What I left alone:

  • I didn't merge T-1070 into T-1071. One is the cause and the other is the failure it produced, so they aren't duplicates.
  • I didn't change T-1071, because the request was only about T-1070. The next shift should watch it: it's due at 10:30Z with Dana Whitfield (E-101), and it's listed as sev2 though the matrix makes department-wide payroll impact sev1.

What to tell Grace: her ticket is at top severity and with incident command, and the 19:40Z deadline still stands. It can't be closed until the replica-lag fix in T-1071 is done and the March batch has been rerun.

Provenance of the numbers on this page

run set
scaffold-or-scale-v1
models
Haiku 4.5, Sonnet 5.5, Opus 5.5
effort
high
repeats
2
fixtures
replay
runs
900 / 900
strategy cost
$354.44
judge cost
$9.19
wall clock
9h 48m
git
fce5807b2d

Cost excludes judge calls, which are tracked separately so a scoring model never flatters or penalises an architecture. Every figure on this site is read from results/ at build time.