Primary proof
41.11% → 73.14%
Average good decision rate across the full GPT-5.5 suite—each point is match to the benchmark-defined optimal action.
Executive summary
GPT-5.5 makes better decisions with BIGHUB
BIGHUB takes GPT-5.5 from 41.11% to 73.14% average good decision rate across the full suite (match to the benchmark-defined optimal action). Same model, same frozen traces, same rubric; arms differ only by inclusion of the BIGHUB decision packet in the input.
Full suite
GPT-5.5 benchmark views
Seven benchmark views: the IT matrix split into incident and helpdesk, plus coldstart, incident large, incident large coldstart, refunds, and refunds large.
IT
Incident
71.95% → 91.67%
+19.72 pp
IT
Helpdesk
40.28% → 82.78%
+42.50 pp
Coldstart
Incident coldstart
71.39% → 85.56%
+14.17 pp
Cardinality scaling
Incident large
44.17% → 86.67%
+42.50 pp
High-cardinality coldstart
Incident large coldstart
44.45% → 75.55%
+31.11 pp
Transfer
Refunds
11.95% → 47.50%
+35.55 pp
Non-IT cardinality
Refunds large
3.61% → 42.22%
+38.61 pp
Proof
Baseline GPT-5.5 vs GPT-5.5 with BIGHUB
Same model family, same frozen benchmark traces, same rubric. Between arms, only whether the BIGHUB decision packet is included in the model input changes; the evaluation contract is otherwise fixed.
| Benchmark | Baseline GPT-5.5 | With BIGHUB | Uplift |
|---|---|---|---|
| IT incident | 71.95% | 91.67% | +19.72 pp |
| IT helpdesk | 40.28% | 82.78% | +42.50 pp |
| Incident coldstart | 71.39% | 85.56% | +14.17 pp |
| Incident large | 44.17% | 86.67% | +42.50 pp |
| Incident large coldstart | 44.45% | 75.55% | +31.11 pp |
| Refunds | 11.95% | 47.50% | +35.55 pp |
| Refunds large | 3.61% | 42.22% | +38.61 pp |
Good decision rate measures match to the benchmark-defined optimal action.
Why this matters
BIGHUB is not helping a weak baseline. It improves GPT-5.5 directly, even when the base model is already strong.
What the suite shows
The uplift is broad across IT, coldstart, large-cardinality, and transfer benchmarks. It is not a single benchmark trick.
Method
Frozen benchmark contract
The suite keeps the same basic design across all benchmark families: frozen datasets, deterministic benchmark rubrics, paired baseline vs packet evaluation, and three seeds per benchmark. Good decision rate measures match to the benchmark-defined optimal action under the same trace context; uplift is the gain in that rate when the packet arm is used.
This page focuses on a single question: once you already have a very strong frontier model like GPT-5.5, how much extra decision quality do you unlock by adding BIGHUB on top?
Common contract
- 3 seeds per benchmark view
- N = 120 traces per seed
- Baseline arm vs packet arm
- Oracle-aligned good decision rate
Benchmark families
- IT incident and helpdesk
- Coldstart structure-only probe
- Large-action incident probes
- Non-IT transfer and cardinality tests
Limits
What these numbers do and do not claim
These benchmarks measure alignment under a frozen authored contract, not guaranteed production business lift. The packet and the rubric share the same benchmark ontology by design; that is what makes the decision surfaces auditable, but it also means this is a framework-aligned evaluation rather than unconstrained production ground truth.
Authored outcomes
Outcome maps and scores are benchmark-authored, not learned from production logs.
Cost realism
API spend depends on current model pricing and does not include engineering overhead or downstream business cost.
Absolute vs relative quality
Higher uplift can come from a weaker baseline. That is why this page reports both uplift and absolute packet good rate.
Model volatility
Temperature is fixed, but API-side nondeterminism and model revision drift can still affect exact cell outcomes over time.