Research

H7 is not a story. It is a measurement instrument.

Pulsaride H7 is the operational downstream consumer of the CAE Research Kit — a two-year falsification programme that extracted a mathematical law from controlled adversarial destruction tests, then operationalised it as a kernel-native observation engine.

Sealed V1 validation — DARPA TC E5 (THEIA)

H7 v1 was evaluated on a real DARPA Transparent Computing Engagement 5 Linux user-workstation (THEIA family) against pre-registered acceptance thresholds (ADR-PROD-006, sealed before the run):

Pulsaride-H7 detected the known APT attack with TPR = 1.00 (lead time 69 s) while sustaining a measured False-Positive Rate of 0.58 % (Wilson 95 % CI [0.46 %, 0.73 %]) over n = 12 000 independent benign ticks across 4 disjoint baseline windows.
MetricMeasuredPre-registered threshold
TPR (attack detection)1.00 (1/1)1.00 (any-window)
Detection lead time69 s—
FPR (n = 12 000 benign ticks)0.58 % (69/12 000)< 5 %
FPR Wilson 95 % CI[0.46 %, 0.73 %]—

The operating envelope — stated, not hidden

CAE discipline requires the symmetric statement of limits. H7 v1 targets continuous Linux server workloads sustaining ≥ 2 000 sched_switch events per 100 ms tickover the calibration window. Below that volumetry the κ estimator's bias dominates the signal and H7 reports gate_alpha=false — the correct behaviour is to not deploy v1 there. Documented out-of-scope, by design:

  • Desktop / developer workstations under GUI load, idle hosts, edge devices, low-traffic VMs.
  • FreeBSD admin-heavy hosts (CADETS dataset: FPR up to 17 % — published as a failure, not hidden).
  • Network-rare-bursty hosts (TRACE dataset: structural κ inversion).
  • Single attack instance in the THEIA ground-truth window — the TPR confidence interval is wide.

The full run is sealed as a chain of custody — 9 Ed25519-signed manifests and 4 sealed ADRs with pre-registered thresholds — and ships with every pilot for offline, independent verification.

From controlled environment to production kernel

The CAE Research Kit README states that active engineering has shifted to the operational downstream consumer. That consumer is H7. The scientific protocol maps directly to the production runbook:

CAE Research KitPulsaride H7 (production)Operational meaning
Controlled environment E₀ / E₁Scoped cgroup (`/sys/fs/cgroup/.../docker-<ID>.scope`)The constraint field is the real Linux scheduler, not a simulation.
Adaptive agent AProduction workload + attackerThe entity interacting with system pressure.
Pressure redistribution (Wasserstein-1)eBPF `sched_switch` probesRaw metric captured by the sensor.
H₅ / H₆ (rejected)Classical EDR heuristicsMathematical proof that raw behavioural deviation (H₅) and observation-feedback artefacts (H₆) fail.
H₇-σ (inversion)Stealth paradoxA sophisticated attacker decreases apparent entropy to blend into background noise — threshold systems miss this.
H₇-κ (one-tick memory)κ metric (short vs long pressure distribution)Local temporal differentiation detector. An attacker cannot simulate perfect pressure-memory continuity tick-to-tick.
ADT (50 tests)Red Team runbook (Phase 10)Production validation that the κ signature survives noisy, hostile environments.

What H7 means

H7 does not stand for a marketing narrative. It denotes the H7 hypothesis family — from H₇-σ (inversion under constraint) to H₇-κ (reversal by single-tick memory) to κ-portability on a calibrated diffusive environment. Each stage was pre-registered, frozen, and published with a Zenodo DOI.

Commercial IOA products sell analyst experience distilled into heuristics. H7 sells the operational translation of a falsifiable metric (KAPPA_TRANSFERS) born from strict Adversarial Destruction Tests.

Calibration is not "learning the noise"

When you run h7ctl calibrate --duration 600, you are executing ADR-027 / ADR-030 in production. You compute μ_baseline — the null-Markov reference from which δ_σ and κ are derived in real time.

For a DORA auditor asking about gate_alpha: it is not an anomaly threshold. It implements the α = 0.005 threshold from the paired Wilcoxon statistical chain of ADR-027. When κ exceeds this gate, the probability that the pressure redistribution is due to chance is below 0.5%.

Why this is not Darktrace

Darktrace operates on non-public heuristics — the operational equivalent of our rejected H₅/H₆ hypotheses, which formally conflate feedback with structure. H7 operates on the H₇-κ signature, which survived 50 Adversarial Destruction Tests and whose portability was proven bit-for-bit on a calibrated diffusive environment (v0.5.0).

The evidence chain (not a pitch deck)

  1. The problem: H₅ and H₆ prove that measuring raw behavioural deviation or using adaptive supervisor models generates feedback artefacts. Classical behavioural detection is mathematically invalid under the CAE protocol.
  2. The discovery: H₇-σ revealed a paradox — under real constraint, adaptation makes the agent closer to the stationary baseline.
  3. The resolution: H₇-κ does not look at pressure deviation; it looks at local temporal differentiation (one-tick memory). A single clock tick breaks the inversion and reveals the adaptive agent.
  4. The product: Pulsaride H7 is the kernel-native observation engine that computes the κ signature on Linux workloads in real time, with cryptographic attestation.

Zenodo publications

Every .cal certificate issued by H7 carries a methodology_reference field pointing to the primary DOI. Independent verification requires no trust in Pulsaride infrastructure.

How it works →CAE Research Kit →