Hive primitives are benchmarked head-to-head against published SOTA, scored on real datasets, and every result is cryptographically signed. No marketing screenshots — receipts.
Protected or Pending by Hive ColonyIP
Corpus-tested across four enterprise CASB and agent workload types. Numbers are measured, not modeled. Invariant recall clears the 99.5% bar required for court-admissible receipts. Amber = preliminary; signed receipt on every run.
| Workload | v1 Baseline | v5 Registry+Neural | Multiplier |
|---|---|---|---|
| Enterprise CASB / policy prompts | 1.04x | 9.15x | 8.8x lift |
| Agent context windows | — | 8.29x | — |
| Verbose / filler-heavy prompts | 1.14x | 5.09x | 4.5x lift |
| RAG repetitive context | 1.58x | 3.41x | 2.2x lift |
| Overall mean | ~1.2x | 5.49x | 4.6x lift |
Corpus results are from production workload testing. Not a head-to-head SOTA benchmark — see signed benchmark cards below for peer-reviewed adversary comparisons. smshPQMax product page →
Every record below is a signed Ed25519 receipt from hivemorph. Click any card for the full methodology.
Inference primitives — production workloads with paying customers
Other primitives — trust infrastructure
Voice primitives — STT/TTS compression & tamper detection
Four non-negotiable rules applied uniformly across every primitive, every adversary, every dataset.
We do not compare against straw men. Adversaries are the highest-citation published baseline for each task: LLMLingua-2 for compression, NIST FIPS-204 for signatures, Llama-Guard for safety, self-consistency CoT for reasoning, DSPy for prompt compilation, Constitutional AI for factuality.
Hive v2 primitives are ensembles that include the SOTA adversary itself as one candidate, plus 3–4 Hive-specific strategies. A quality oracle picks the per-input winner. By construction, the ensemble cannot lose to the adversary alone.
We commit to the dataset, sample size, metric, and decision criteria before running the benchmark. Pre-registration is published at github.com/srotzin/xcalibur-evaluation.
Every result line, every paired t-statistic, every Cohen's d is committed in a signed Ed25519 receipt.
Receipts are queryable at /v1/trust/benchmarks on hivemorph.onrender.com.
Tamper any field — signature breaks. No editable marketing slides.
Every benchmark record carries one of three statuses. Status is computed from the data, not editorially assigned.
Every record is public. You do not need an account or an API key.
curl -sS https://hivemorph.onrender.com/v1/trust/benchmarks/{record_id}
Every field is signed. Re-derive the signature against the record's pubkey_hex.
If you can verify it, the record is authentic.
Hive primitives are Patent Pending. Provisional patents filed. The methodology, benchmarks, and receipts are open. The cryptographic primitives are protected.
Hive ColonyIP →