INCONCLUSIVEThe interval for skipped-case error vs references straddles tolerance 5%. More audits cannot help (every skipped case is already graded or audited); more tasks are needed to resolve it: about 13,247 more skipped cases if the rate holds. From the audit alone (weighted disagreement 0.80%, interval 0.01%-4.8%) the status would be OK; that is what a run without references reports.
The fast tier (haiku) answered 90% of 520 tasks without escalating. On those, it disagrees with sonnet on an estimated 0.80% (95% CI 0.0–4.8%, 122 audits, n_eff 110). Against reference answers, the skipped cases are wrong on 5.4% (95% CI 3.7–7.8%, 466 graded). That is about 25.0 wrong answers among the 466 skipped cases (17.1–36.3).
Answered by fast tier
89.6%
466 of 520 tasks
Escalation rate
10.4%
95% CI 8.0–13.3%
Served accuracy
92.7%
95% CI 90.1–94.6%, 520 graded
Cost per task
$0.0013
Savings vs always sonnet
82.6%
95% CI 82.2–83.0%; always sonnet: $0.0074/task
Audit overhead
132.9%
$0.89 audit spend
Expected wrong skipped cases
25.0
95% CI 17.1–36.3
Audit by confidence bin
Skipped cases by served confidence (95% CI in brackets)
Confidence bin
Skipped
Audited
Disagreement (CI)
Error vs references (CI)
[0.00, 0.50)
0
0
—
—
[0.50, 0.70)
0
0
—
—
[0.70, 0.80)
0
0
—
—
[0.80, 0.90)
53
19
5.3% (0.1–26.0%)
20.8% (12.0–33.5%)
[0.90, 0.95)
43
18
5.6% (0.1–27.3%)
9.3% (3.7–21.6%)
[0.95, 1.00]
370
85
0% (0.0–4.2%)
2.7% (1.5–4.9%)
Whiskers are 95% CIs; n is the number of audits per bin.
Tier usage and costs
Decisions served per tier
Tier
Role
Served
Share
haiku
fast
466
89.6%
sonnet
final, reference
54
10.4%
Audit coverage of skipped cases
Quantity
Count
Skipped
466
Audited
122
Not selected for audit
344
Audit pending
0
Audit errors
0
Undecided (judge could not decide)
0
No audit record
0
Graded against references
466
Unweighted disagreement (biased)
1.6% (0.5–5.8%)
Audit tier (sonnet) error vs references
3.6% (1.0–9.0%)
Costs (USD)
Quantity
Value
Serving cost (total)
$0.67
Audit cost (total)
$0.89
Cost per task
$0.0013
Est. cost per task, always sonnet
$0.0074 ($0.0072–$0.0076)
Estimated savings
82.6% (82.2–83.0%)
Audit overhead (audit / serving)
132.9%
Decisions with unknown serving cost
0
Decisions with unknown audit cost
0
Threshold sweep
Offline threshold sweep over 520 eval-mode tasks (21 threshold settings); accuracy is measured against reference answers.
cheapest thresholds with accuracy >= 0.9323 (best single tier 'sonnet' at 0.9423 minus 0.01) on the selection split; validated on 156 held-out decisions (seed 0)
Baselines on all 520 usable tasks
Router
Accuracy (CI)
Cost / task
Escalation
Tier share
always haiku
91.2% (88.4–93.3%)
$0.00048
0%
haiku 100%, sonnet 0%
always sonnet
94.2% (91.9–95.9%)
$0.0073
0%
haiku 0%, sonnet 100%
oracle (lower bound)
95.0% (92.8–96.6%)
$0.00076
3.8%
haiku 96%, sonnet 4%
recommended (all data)
93.5% (91.0–95.3%)
$0.0023
25.4%
haiku 75%, sonnet 25%
The oracle serves from the cheapest correct tier without paying for the tiers it skips; no real router can reach it.
Pareto frontier
Thresholds
Accuracy (CI)
Cost / task
Escalation
Skipped error (CI)
haiku ≥ 0
91.2% (88.4–93.3%)
$0.00048
0%
8.8% (6.7–11.6%)
haiku ≥ 0.4
91.3% (88.6–93.5%)
$0.00054
0.77%
8.7% (6.6–11.5%)
haiku ≥ 0.6
91.5% (88.8–93.6%)
$0.00058
1.3%
8.6% (6.5–11.3%)
haiku ≥ 0.7
92.3% (89.7–94.3%)
$0.00099
6.5%
6.6% (4.7–9.1%)
haiku ≥ 0.72
92.7% (90.1–94.6%)
$0.0011
8.5%
5.9% (4.1–8.4%)
haiku ≥ 0.85
92.9% (90.3–94.8%)
$0.0017
16.2%
4.4% (2.8–6.7%)
haiku ≥ 0.88
93.3% (90.8–95.1%)
$0.0020
20.4%
3.4% (2.0–5.6%)
haiku ≥ 0.92
93.5% (91.0–95.3%)
$0.0023
25.4%
2.8% (1.6–5.0%)
recommended
haiku ≥ 0.95
93.7% (91.2–95.4%)
$0.0026
28.8%
2.7% (1.5–4.9%)
haiku ≥ 0.97
93.8% (91.4–95.6%)
$0.0030
34.2%
2.3% (1.2–4.5%)
haiku ≥ 0.99
94.2% (91.9–95.9%)
$0.0065
80.8%
0% (0.0–3.7%)
Calibration
Confidence calibration per non-final tier
Tier
Scored
ECE
Brier
AUROC
AURC
Missing scores
haiku
520
0.016
0.069
0.836
0.026
0
ECE and Brier: lower is better. AUROC: how well the score separates right from wrong answers (0.5 = chance). AURC: area under the risk-coverage curve, lower is better.
Reliability diagram for haiku (520 scored tasks).Risk-coverage curve for haiku.
Caveats
est_savings treats serving cost per task as exact; its interval reflects only the uncertainty in the always-sonnet cost estimate, and excludes audit cost.
Methodology
Disagreement is the share of skipped cases (answered by a non-final tier without escalating) whose answer differs from the final (audit) tier's answer to the same task, as decided by the configured comparator or judge. In serve mode only a random sample of skipped cases is shadow-audited; each audited case is weighted by 1/π, the inverse of its known inclusion probability (a Hajek inverse-probability-weighted estimator), so the estimate describes all skipped cases rather than only the audited ones. In eval mode every tier answers every task, so every skipped case is observed (π = 1). The audit tier is a proxy, not ground truth: disagreement counts differences from its answers, not verified errors, unless reference answers exist, in which case the error rate against them is reported as well. Audits that are pending, failed or undecided are treated as nonresponse: within each inclusion-probability stratum, completed audits are up-weighted by the inverse of the stratum's response rate, which assumes audits are missing at random within their stratum; strata without any completed audit cannot be represented and are excluded from the estimate (a caveat says so). Intervals are Wilson score intervals for unweighted proportions and Korn-Graubard intervals (Clopper-Pearson at the effective sample size n_eff) for weighted ones; when a stratum holding at least 10% of the skipped cases has fewer than 10 completed audits, the status is never reported as ok. The status uses a fixed-sample interval: checking it repeatedly as audits accumulate and stopping at the first ok inflates the chance of a false ok, so fix the number of audits in advance or use a stricter confidence level for repeated checks. The savings interval reflects only the uncertainty in the cost of always using the final tier. Sweep accuracies are Wilson intervals; recommended thresholds are chosen on a selection split and re-evaluated on a held-out split.