Verified against supplied artifacts · support-assurance-demo-v1

Synthetic support-agent fixture

Requested SCALE

ASSIST

The fixture is profitable versus its named baseline, but required quality, unit-cost, tail-cost, and runtime-cap claims fail policy.

Required claims
6
Pass / fail / incomplete
2 / 4 / 0
Population
Eight synthetic support tasks checked into examples/
Window
2026-07-01 to 2026-07-01
Counterfactual
Human-only support queue
Review
2026-10-08 · Vyoma Gajjar

Required claims

At least 80% of attempted tasks produce acceptable outcomes.

FAIL

6/8 acceptable outcomes; acceptable_rate = 75.0%

Policy
acceptable_rate >= 80.0%
Check
gate.acceptable-rate@1

Cost per acceptable outcome does not exceed policy.

FAIL

cost_per_acceptable_outcome_usd = 3.50

Policy
cost_per_acceptable_outcome_usd <= 2.00
Check
gate.unit-economics@1

The p95 effective task cost does not exceed policy.

FAIL

p95_task_cost_usd = 14.25

Policy
p95_task_cost_usd <= 8.00
Check
gate.tail-cost@1

Expected net value per attempt is non-negative.

PASS

expected_net_value_per_attempt_usd = 3.37

Policy
expected_net_value_per_attempt_usd >= 0.00
Check
gate.net-value@1

Expected value exceeds the named human-only baseline.

PASS

incremental_net_value_vs_baseline_usd = 2.77

Policy
incremental_net_value_vs_baseline_usd >= 0.00
Check
gate.counterfactual@1

Every task remains inside call-count and trace-cost caps.

FAIL

t-005 used 12 calls and 0.2454 USD of trace cost

Policy
calls_per_task <= 8 and trace_cost_per_task_usd <= 0.15
Check
gate.runtime-caps@1

Required before the next mode

  • Meet the acceptable-rate, unit-cost, and tail-cost policy gates.
  • Enforce per-task call and trace-cost caps before reconsidering SCALE.

Known limits and non-claims

  • No production sampling, temporal drift estimate, or subgroup analysis
  • This record is not a production deployment authorization.
  • The eight-task fixture is not an estimate of production prevalence or ROI.

Verification replays this record against the supplied artifacts. It does not establish the real-world truth or representativeness of the inputs.

Evidence digest: 6441a11a69a0958151103fb24d43d8f86507a22fe64bd0d407a8a4bc0c310614
Decision contract: e7faae0cb2b0fb62c5341412c16c8e7930142eaf86cd8e8568b0dfad72c3baab
Reproduce: agent-economics decision-record validate examples/agent-decision-record.json --bundle examples/agent-decision-record.bundle.json --attestations examples/attestations.json --artifact-root .