shadowgate report: opus-serve

Run opus-serve · mode serve · 520 decisions · tiers haiku → opus

Summary

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 2.7%, interval 0.54%-7.6%) the status would be INCONCLUSIVE; that is what a run without references reports.

The fast tier (haiku) answered 90% of 520 tasks without escalating. On those, it disagrees with opus on an estimated 2.7% (95% CI 0.5–7.6%, 127 audits, n_eff 111). 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
93.5%
95% CI 91.0–95.3%, 520 graded
Cost per task
$0.0020
Savings vs always opus
85.7%
95% CI 85.3–86.2%; always opus: $0.014/task
Audit overhead
169.2%
$1.80 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 binSkippedAuditedDisagreement (CI)Error vs references (CI)
[0.00, 0.50)00——
[0.50, 0.70)00——
[0.70, 0.80)00——
[0.80, 0.90)532714.8% (4.2–33.7%)20.8% (12.0–33.5%)
[0.90, 0.95)43220% (0.0–15.4%)9.3% (3.7–21.6%)
[0.95, 1.00]370781.3% (0.0–6.9%)2.7% (1.5–4.9%)
Skipped-case disagreement by confidence binBars show the disagreement per bin with 95% CI whiskers; the dashed line is the tolerance.0%10%20%30%40%[0.00, 0.50) n=0n=0[0.00, 0.50)[0.00, 0.50)[0.50, 0.70) n=0n=0[0.50, 0.70)[0.50, 0.70)[0.70, 0.80) n=0n=0[0.70, 0.80)[0.70, 0.80)[0.80, 0.90) 15% [4.2%, 34%] n=27n=27[0.80, 0.90)[0.80, 0.90)[0.90, 0.95) 0% [0%, 15%] n=22n=22[0.90, 0.95)[0.90, 0.95)[0.95, 1.00] 1.3% [0.032%, 6.9%] n=78n=78[0.95, 1.00][0.95, 1.00]tolerance 5%served confidencedisagreement
Skipped-case disagreement by confidence binBars show the disagreement per bin with 95% CI whiskers; the dashed line is the tolerance.0%10%20%30%40%0.00 n=0n=00.000.000.50 n=0n=00.500.500.70 n=0n=00.700.700.80 15% [4.2%, 34%] n=27n=270.800.800.90 0% [0%, 15%] n=22n=220.900.900.95 1.3% [0.032%, 6.9%] n=78n=780.950.95tolerance 5%served confidence (bin start)disagreement
Whiskers are 95% CIs; n is the number of audits per bin.

Tier usage and costs

Decisions served per tier
TierRoleServedShare
haikufast46689.6%
opusfinal, reference5410.4%
Audit coverage of skipped cases
QuantityCount
Skipped466
Audited127
Not selected for audit339
Audit pending0
Audit errors0
Undecided (judge could not decide)0
No audit record0
Graded against references466
Unweighted disagreement (biased)3.9% (1.7–8.9%)
Audit tier (opus) error vs references6.6% (2.7–12.9%)
Costs (USD)
QuantityValue
Serving cost (total)$1.06
Audit cost (total)$1.80
Cost per task$0.0020
Est. cost per task, always opus$0.014 ($0.014–$0.015)
Estimated savings85.7% (85.3–86.2%)
Audit overhead (audit / serving)169.2%
Decisions with unknown serving cost0
Decisions with unknown audit cost0

Threshold sweep

Offline threshold sweep over 520 eval-mode tasks (21 threshold settings); accuracy is measured against reference answers.

Accuracy vs cost per task21 simulated threshold settings (gray), the Pareto frontier (line), single-tier and oracle baselines, and the recommended point (star).91%92%93%94%95%96%$0.00050$0.0010$0.0020$0.0050$0.010cost per task (USD, log scale)accuracyoperating pointsPareto frontieralways haikualways opusoracle (lower bound)recommendedoracle (lower bound)always opusrecommendedalways haikuoperating pointsoperating pointsPareto frontierPareto frontieralways haikualways haikualways opusalways opusoracle (lower bound)oracle (lower bound)recommendedrecommended
Accuracy vs cost per task21 simulated threshold settings (gray), the Pareto frontier (line), single-tier and oracle baselines, and the recommended point (star).91%92%93%94%95%96%$0.00050$0.0010$0.0020$0.0050$0.010cost per task (USD, log scale)accuracyoperating pointsPareto frontieralways haikualways opusoracle (lower bound)recommendedoracle (lower bound)always opusrecommendedalways haikuoperating pointsoperating pointsPareto frontierPareto frontieralways haikualways haikualways opusalways opusoracle (lower bound)oracle (lower bound)recommendedrecommended
Gray: every simulated threshold setting. Line: Pareto frontier. Star: recommended thresholds.

Recommendation

Objective
max-savings
Thresholds
haiku ≥ 0.88
Selection accuracy (364 tasks)
94.0% (95% CI 91.0–96.0%)
Selection cost per task
$0.0036
Selection escalation rate
20.9%
Held-out accuracy (156 tasks)
94.2% (95% CI 89.4–96.9%)
Held-out cost per task
$0.0033
Held-out escalation rate
19.2%
Accuracy vs always opus
-1.3 pp (95% CI -3.2 to +0.0 pp)

cheapest thresholds with accuracy >= 0.9378 (best single tier 'opus' at 0.9478 minus 0.01) on the selection split; validated on 156 held-out decisions (seed 0)

Baselines on all 520 usable tasks
RouterAccuracy (CI)Cost / taskEscalationTier share
always haiku91.2% (88.4–93.3%)$0.000480%haiku 100%, opus 0%
always opus95.0% (92.8–96.6%)$0.0140%haiku 0%, opus 100%
oracle (lower bound)96.0% (93.9–97.3%)$0.00124.8%haiku 95%, opus 5%
recommended (all data)94.0% (91.7–95.8%)$0.003520.4%haiku 80%, opus 20%

The oracle serves from the cheapest correct tier without paying for the tiers it skips; no real router can reach it.

Pareto frontier
ThresholdsAccuracy (CI)Cost / taskEscalationSkipped error (CI)
haiku ≥ 091.2% (88.4–93.3%)$0.000480%8.8% (6.7–11.6%)
haiku ≥ 0.491.3% (88.6–93.5%)$0.000590.77%8.7% (6.6–11.5%)
haiku ≥ 0.691.5% (88.8–93.6%)$0.000661.3%8.6% (6.5–11.3%)
haiku ≥ 0.792.7% (90.1–94.6%)$0.00156.5%6.6% (4.7–9.1%)
haiku ≥ 0.7293.3% (90.8–95.1%)$0.00188.5%5.9% (4.1–8.4%)
haiku ≥ 0.893.5% (91.0–95.3%)$0.002010.4%5.4% (3.7–7.8%)
haiku ≥ 0.8593.7% (91.2–95.4%)$0.002916.2%4.4% (2.8–6.7%)
haiku ≥ 0.8894.0% (91.7–95.8%)$0.003520.4%3.4% (2.0–5.6%)recommended
haiku ≥ 0.9294.4% (92.1–96.1%)$0.004125.4%2.8% (1.6–5.0%)
haiku ≥ 0.9594.6% (92.3–96.2%)$0.004628.8%2.7% (1.5–4.9%)
haiku ≥ 0.9794.8% (92.6–96.4%)$0.005434.2%2.3% (1.2–4.5%)
haiku ≥ 0.9995.0% (92.8–96.6%)$0.01280.8%0% (0.0–3.7%)

Calibration

Confidence calibration per non-final tier
TierScoredECEBrierAUROCAURCMissing scores
haiku5200.0160.0690.8360.0260

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: haikuAccuracy per confidence bin with Wilson intervals; points on the diagonal are perfectly calibrated.0%20%40%60%80%100%0%50%100%mean confidenceaccuracyperfect calibrationobserved
Reliability diagram for haiku (520 scored tasks).
Risk-coverage: haikuError rate among the skipped cases as the threshold is lowered and the tier answers more cases alone.0%2%4%6%8%10%0%50%100%coverage (share answered by tier)risk (error among skipped)tolerance 5%risk
Risk-coverage curve for haiku.

Caveats

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.