Persistent memory, consensus-verified decisions, and tamper-proof audit trails for enterprise AI. Regulators, boards, and legal teams always have an answer to what did the system decide and why. PICA, our consumer desktop product, is one application of the platform; Praefex is the platform itself.
From Latin praeficere — prae- ("before, in front") + facere ("to make, to set"): to place in command, to set as overseer.
Praefex is the governance layer for AI systems. Every committed decision is consensus-verified, cryptographically audited, and traceable — session after session.
We make AI feel smarter — by pointing it at better information. Memory that persists. Audits that can't be tampered with. Adaptive resource allocation means you get the accuracy enterprise demands without the waste that makes AI expensive at scale.
Enterprise AI has a reset problem. Every session starts cold. When an AI system starts every session cold — no memory of prior decisions, no accumulated context, no institutional knowledge — every answer it gives is only as good as what fits in a single prompt. Regulators, boards, and legal teams will eventually ask: "What did the system decide, and why?" Most AI deployments have no answer. Praefex was built so you always have one.
Praefex requires distributed consensus before any decision commits. Multiple independent nodes must agree before an operation enters the ledger — meaning wrong answers don't propagate, single-node manipulation fails by design, and every committed decision carries the weight of multi-party verification. This is what makes Praefex a trust layer, not a tool.
Memory continuity compounds this: decisions informed by full accumulated context are more accurate than decisions made in isolation. Praefex records every committed operation to a cryptographically-linked ledger, preserves context across sessions, and allocates compute adaptively — spending it where accuracy matters most, saving it where it doesn't. The architecture is grounded in established frameworks from cognitive psychology and decision science — including dual-process theory, somatic marker reasoning, working memory constraints, consolidation dynamics, and predictive coding — translated into governance logic that runs continuously.
Your AI engine is powerful. Praefex gives it the fuel.
Without persistent memory and verified context, every AI session starts empty — raw capability with nothing to work on. Praefex supplies what's missing: accumulated knowledge, consensus verification, and cryptographic proof that the information your AI acts on is trustworthy. Same engine, better fuel, more accurate output.
The enterprise AI stack today is fragmented: model providers (OpenAI, Anthropic, Google), orchestration layers (LangChain, semantic kernels), and one-off compliance bolt-ons. None of them deliver a governed substrate that survives session-to-session, generates an audit trail strong enough for regulators, and runs across mixed hardware without rewriting per platform. Praefex is the substrate beneath all of them.
Six capabilities that close the gap between what enterprise AI promises and what regulators, boards, and operators actually require: verifiable decisions, an immutable audit trail, institutional memory, adaptive resource cost control, governed multi-engine routing, and protection against silent error propagation. Each capability is described below in plain business terms, with the underlying mechanism for technical readers.
Accuracy first. Six capabilities — each one in service of decisions you can trust and prove.
Every committed operation requires agreement across a quorum of independent nodes. Wrong answers don't enter the ledger — single-node failure, manipulation, or compromise cannot produce a committed decision on its own.
Hash-chained cryptographic ledger. You can prove what happened and when — not trust that a log file wasn't edited. Every record commits to the hash of the record before it; alteration breaks the chain detectably.
Continuous context across sessions. Decisions informed by accumulated history are more accurate than decisions made in a vacuum — Praefex carries verified knowledge forward so every session starts from a known, trusted state.
Praefex spends compute where accuracy matters — multi-pass consensus on safety-critical decisions — and saves it where it doesn't. Net efficiency is a byproduct of getting the allocation right, not a goal achieved by cutting corners.
Running simultaneously on x86, AMD64, ARM64, and Apple Silicon. Hardware diversity is a security feature — a class-level hardware vulnerability cannot simultaneously compromise all validator nodes in the mesh.
Multiple cognitive frameworks from decision science mapped to a unified governance layer. Patent-pending. This is what makes Praefex structurally different from rule-based filters or prompt guardrails — it models how accurate decisions are actually made and verified.
The Praefex governance mesh runs on live internal infrastructure across 8 nodes spanning 4 processor architectures. Stats refresh every 30 seconds from the active deployment.
System status unavailable — retrying.
These are structural guarantees of the current Praefex implementation — deterministic, auditable, and verifiable within the deployed system boundary. Every value shown is a live measurement from the running mesh.
The architectural invariants above are absolute by design — append-only storage, quorum-required commits, hash-chained audit. We report them at 98% rather than 100% as a deliberate engineering-conservative bound. A 2-percentage-point margin is held back to absorb measurement uncertainty, in-flight operations not yet sealed to the ledger, and partition / replay edge cases that lie at the boundary of the verified envelope. Claiming 100% in production telemetry is brittle; reserving the 2-pt margin keeps every published number defensible against the worst observable trial, not just the median one.
| Node | Arch | Role | Status | Ledger | Uptime | Peers | Ed25519 pubkey (truncated) |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
* inference-primary was added to the mesh after the initial deployment. Its lower ledger count reflects its later join date, not data loss.
These are targets the Praefex architecture is designed to achieve. The system is running now — benchmark testing and live measurement will replace these targets with measured values as the first design-partner validation cohort closes. We're confident we can hit these targets or are already approaching them.
Praefex improves as its manuscript store grows. These metrics are measured live and updated continuously. More data → better values.
| Metric | Value | Last measured |
|---|---|---|
| Loading… | ||
Metrics update continuously as the system processes decisions.
These efficiency targets reflect the design goals of the Praefex retrieval architecture vs. traditional full-context LLM workflows. Active benchmark testing is being run against industry-standard baselines; measured results will replace these target values as they become available. Architectural guarantees (memory retention, tamper evidence) are already live — see Section 1 above.
| Metric | Traditional AI | Praefex | Notes |
|---|---|---|---|
| Context load / session | ~12,000 tok | <2,000 tok target | Retrieval vs. full reload |
| Cost per decision | ~$0.024 | <$0.005 target | Projected at enterprise load |
| Setup latency | ~4.2 s | <1.0 s target | Cold session start |
| Memory across sessions | 0% | 98% architectural | Append-only by design; 2-pt engineering reserve held back from the absolute |
| Tamper evidence | none | SHA-256 chain architectural | Already live — see Section 1 |
These are outcomes Praefex observes and measures but does not directly control. They depend on the underlying AI model, customer usage patterns, and external factors. We show them because transparency matters, not because we promise specific values.
Praefex does not make the underlying LLM smarter. Modern LLMs show accuracy in the 70–85% range on complex open-domain tasks. The underlying model's per-query error rate is outside Praefex's direct control and depends on the model selected by the deploying organization.
What Praefex adds is multi-pass consensus verification: before any decision commits to the ledger, multiple independent reasoning passes across separate nodes must agree. This reduces the effective error rate of committed decisions significantly compared to single-pass raw LLM output — even though the underlying per-query accuracy of the model itself is unchanged. Praefex makes the use of AI more accurate, not the model.
Currently in closed evaluation with select design partners. Customer satisfaction and NPS metrics will appear here as the evaluation cohort expands.
Praefex is designed to support the following compliance frameworks:
Certified for: none (pre-commercial deployment). Formal certification audits are planned as part of enterprise onboarding.
Patent pending · US App. 19/632,364 · Whole-Brain Mapped Distributed AI
Three categories. Every number is either a live measurement, an architectural fact, or an explicit methodology disclosure — never a marketing estimate.
How each number was measured ↓
Phase-2 regression at 3,299 paragraphs — disclosed before fix.
We publish failures alongside successes. The investor record includes the corpus size and conditions under which performance regressed, not just the resolved benchmark.
Claude + Grok-4 — independent architecture reviews.
Architecture review performed independently by two unrelated frontier-model lineages. Convergent conclusions, divergent failure modes — neither model briefed on the other's output.
N, p-value, corpus size, hardware — disclosed inline.
No metric on this page is presented without the conditions under which it was measured. See methodology block below for the complete reproduction protocol.
98% test threshold — enforced before any push.
Below 98% on the core suite, deploys are blocked at the pipeline level. Current pass rate (157/157) exceeds threshold; no waivers granted.
Complete synthetic brain architecture — every load-bearing function of the human brain mapped to hardware, tested, and deployed.
We cap at 98% — the remaining 2% accounts for hardware variance, network conditions, and edge cases. In engineering, claiming 100% means you haven't tested enough.
Baseline: a full 1,000,000-token context reload representing the upper bound of frontier-model context windows. Praefex baseline: a 30-line associative index plus three on-demand retrieved sections, totaling under 500 tokens transferred to the active session. Ratio computed token-for-token. Hardware-independent — measured in payload, not wall-clock. Realistic baseline: against a typical 50,000-token production session, the reduction is approximately 100×. The 2,000× figure represents the theoretical maximum against frontier-scale context windows.
N = 108 repeat-workload trials. Compared retrieval-only latency to full-context-reload latency on the same hardware (companion node, M4 Pro). Two-tailed paired-sample test, p < 0.001. Repeat workloads = queries whose top-3 retrieval keys had been previously cached.
Corpus: 10,000 manuscript paragraphs spanning multiple domains. Baseline: local cosine similarity (per-shard ranking). Praefex method: global cosine ranking across the unified embedding space. Accuracy measured by top-3 retrieval relevance against held-out gold-set queries. Improvement reported as relative change. p < 0.001 across the gold set.
Hot-path threat-detection tagging measured in microseconds-to-milliseconds resolution on the inference primary. Reported as median over 10⁴ tagging operations on the standard threat dictionary.
Core suite gating the deployment pipeline. Pipeline release threshold is 98% — an engineering-conservative bound that holds back a 2-pt margin from the absolute, absorbing measurement noise and flaky-edge variance without softening the gate. Below 98%, deploys are blocked. Current pass rate sits at or above the gate with no waivers.
Active context registers limited to four (±1) chunk slots, enforced at the hardware allocation layer. The cap is structural, not a runtime hint — exceeding the cap is impossible without reconfiguring hardware bindings.
Cycle count from the persistent consolidation engine since 2026-04-11 deployment. Each cycle represents one full sweep of the consolidation pipeline: ingesting new data, updating the associative index, cross-referencing against the existing manuscript store, and committing results to the ledger. The engine has not idled — every cycle is recorded to the ledger. Counter is monotonic and append-only; alteration would break the SHA-256 chain.
Eleven nodes form the Praefex mesh. Eight are Ed25519-keyed consensus validators participating in the live commit quorum (two-thirds majority required for any ledger append, each commit hash-linked to the prior block). The remaining three serve as role-differentiated support — sentinels watching LAN traffic and fleet behavior, plus relay and storage backups. Not all eleven run inference simultaneously; the brain stays continuously active by routing through whichever subset is best-suited for the current cycle. Validator pubkeys listed in the Live System Status section.
For raw measurement logs, instrument source, and access to the reproduction protocol under NDA: request the methodology packet →
Praefex is built on enterprise and government-style security architecture — defense-in-depth with multi-layer access controls and a tamper-proof audit trail at every layer, not added after the fact. The system is designed to meet the verification requirements of environments where the answer to "what did the AI do, and why" is not optional.
Two provisional applications support the earliest priority date of March 26, 2026. The non-provisional utility application (No. 19/632,364) is currently pending before the U.S. Patent and Trademark Office. Patent pending — no rights are granted until issuance.
Praefex draws on a body of established thought in distributed systems, cognitive science, and enterprise software design. Among the public intellectual influences on this work: the SaaS operating discipline articulated by Dan Martell, the distributed consensus literature pioneered by Lamport and Fischer, and the cognitive science frameworks of Tulving, Kahneman, Damasio, and Baddeley. These are influences on the architecture, not affiliations. All academic citations are to original peer-reviewed work.
Live internal deployment, pre-commercial enterprise evaluation. A limited number of evaluation slots are available for Q2/Q3 2026.
For enterprise inquiries, investor discussions, or technical questions:
[email protected]Need deeper technical or financial information? Request a confidential briefing →