Verifiable Evidence Intelligence

Analyze vast document universes.
Prove what you found.

Probans is an evidence engine that breaks a question into independent lines of inquiry across thousands to millions of files — on hardware you control — and compiles results an expert can verify without our software.

Request Private Access See the Pillars →
By Quasen · Available Q4 2026
The Category

We did not enter a market.
We defined a standard.

Verifiable Evidence Intelligence (VEI) is a category for domains where a conclusion is only useful if its evidentiary path can be inspected. Conventional AI asks a reviewer to trust a fluent answer. VEI asks them to inspect the proof: the sources, the absences, the contradictions, and the runtime that produced the result. The Probans engine, built at Quasen, is its first reference implementation.

Every query terminates in a Witness Pack (.vei) — a cryptographically signed bundle containing verbatim quotes anchored to source paragraphs, evidentiary references, integrity metadata, and a complete audit chain. Independently verifiable without our software — today, or by a stranger five years from now, on a different machine.

  • Evidentiary, Not GenerativeIt compiles what your documents already contain. Every claim must resolve to a source — and claims it cannot anchor are refused, not generated.
  • Parallel Inquiry, Not Single-Pass SummaryA question is decomposed into independent lines of inquiry before the evidence is compiled.
  • Operator-Controlled by ArchitectureSensitive material never leaves the perimeter you define. Architecture — not vendor policy — enforces the boundary.
  • Contradiction-AwareConflicting evidence is elevated, not smoothed into a convenient answer.
The Pillars

Three pillars. One proof.

Each pillar can be verified by a third party — on hardware we do not own, with software we do not control.

Verify the result

Open verifier.

The .vei format specification is published. A standalone reference verifier, probans-verify, is released under Apache-2.0. A third party reproduces verification on their own machine — without our engine, cloud, or infrastructure.

Verify the runtime

Pinned environment.

A frozen, content-addressed runtime pins every dependency. Verification replays to the exact result years later, on a different machine. Reproducibility is an artifact property, not a cloud promise.

Verify the airgap

Provable airgap.

The runtime instruments network primitives and records any external-call attempt as a signed audit event, sealed into the Witness Pack. Inference runs where you decide — self-hosted SGLang or vLLM, on-premise, private cloud, or inside a SCIF.

Run on an air-gapped SCIF. Produce a sealed .vei. An independent expert verifies it ten years from now — without our infrastructure, without our API, without us.

Built For

Wherever the burden of proof
falls on someone.

Probans is not primarily a legal tool. It is an evidence intelligence layer for any environment where large private corpora, sensitive information, and accountable conclusions meet.

i.
Government & Public Sector
Policy archives, procurement files, correspondence, public records, inspection materials, committee evidence, and administrative decisions — with findings that remain reviewable.
ii.
Defense & Restricted Environments
Operational files, logistics records, incident reports, after-action materials, supplier documentation — analyzed inside controlled or air-gapped infrastructure.
iii.
Energy & Critical Infrastructure
Engineering records, outage reports, maintenance logs, vendor evidence, safety documentation, regulatory correspondence, and risk registers across complex asset networks.
iv.
Banking & Financial Intelligence
AML investigations, KYC reviews, transaction reconciliation, sanctions screening, model-risk documentation — defensible to the supervisor and the audit committee.
v.
Healthcare & Life Sciences
Clinical record review, adverse-event correlation, SOP and trial documentation — under residency constraints that prohibit external transit of patient data.
vi.
Digital Forensics & Cybersecurity
Incident reconstruction, log correlation, communications analysis — with chain-of-custody preserved from ingest to expert report.
vii.
Legal, Disputes & Investigations
Litigation files, contract estates, regulatory dossiers, M&A diligence — assembled into chains that survive cross-examination.
viii.
Audit, Oversight & Compliance
Policy adherence reviews, control testing, whistleblower investigations — including the harder finding: that something is not there.
The Engine

Probans does not stop at an answer.
It compiles accountable evidence.

Probans decomposes a query into independent lines of inquiry, preserves provenance, detects conflicts, and constructs an irreducible chain of proof for every conclusion — across contracts and statutes, transaction logs and clinical records, system events and case files. It is built to expose what is known, what is contradicted, and what remains unproven.

01
Ingest
Millions of files processed locally — PDFs, scans, emails, logs, exports, records. Hashed and referenced at ingest, sealed against post-hoc modification. Provenance is established before reasoning begins.
02
Map
Extract entities, dates, references, versions, duplicates, and document families across the full corpus. Build a navigable evidence map — not a loose search index — spanning jurisdictions and domain vocabularies.
03
Compile
Each evidence card carries an explicit stance: support, refute, qualify, context, or absence. Every connection is anchored to its source paragraph and constrained by the compiled ledger. Where the ledger ends, so does the claim.
04
Verify
Export a Witness Pack with hashes, signatures, source references, audit metadata, and replayable evidence structure. Admissibility-ready for the regulator, board, expert, or independent counter-review — today, or in five years.
The Standard

Structural defenses.
Enforced at the evidence layer.

Cloud AI optimizes for conversational fluency. Probans optimizes for the weight a conclusion must carry. Reproducibility, contradiction-honesty, scoped negative proof, and audit-first execution are not interface features. They are enforced properties of the architecture.

Citation-Anchored or Refused

"I cannot answer this from the indexed documents." — Probans G3 refusal, in source

Every claim is anchored to a source span at decode time. Where the provider supplies native character-level citations, anchors come directly from the document; otherwise anchoring is enforced structurally by the G3 gate. Out-of-source means refusal — not disclaimer, not degraded mode, not apology. No silent pass.
Symbolic Verification
The system that generates is not the system that verifies. The faithfulness gate (G4) runs heuristic scoring, not an LLM. The contradiction gate (G5) is a pure function over a compiled ledger. Referential closure, Absence Certificate, and epistemic ablation are deterministic by construction. The thing that checks the AI is not AI.
Provable Airgap
The runtime instruments network primitives and records any external-call attempt as a structured, signed audit event. The sealed audit log travels with the result, so a reviewer can inspect whether the process stayed inside the declared boundary. Demonstrable on your machine.
Contradiction-First Gate
The G5 gate hard-codes a single rule: contradiction prevails over sufficiency. Where supporting evidence numerically dominates, the verdict remains contradicted, not closed. The engine surfaces inconsistencies, superseded clauses, and missing instruments. Investigation, not assistance.
Self-Verifying or Withheld
The .vei file is a typed container, not merely a naming convention. It validates its manifest, hashes, and signatures on export and on every replay — years later, offline, on an unrelated machine. If integrity fails, verification fails closed and the artifact is withheld. Not a promise. A rule.
Absence Certificate
Negative proof as a first-class artifact. "Three thousand files were searched for clause / transaction / event X. None was found. Here is the scope, here is the rationale." Composed for regulators, courts, and auditors — for whom absence, properly documented, is itself a finding. What is not there is also evidence.
Reproducibility Ladder
Four stages of reproducibility: retrieval trace, environment manifest, ledger-closed rendering, and bit-level deterministic kernels. A schema validator enforces that no claim escapes the compiled record. Stages we have not yet achieved are declared pending — never claimed. We declare what we have built, not what we wish to claim.
Vector Inversion Defense
Vector exports are quantized to int8 by default, reducing the surface for text-reconstruction attacks at marginal utility cost. Optional binary quantization removes the inversion surface in practice. We do not export raw FP16/FP32 vectors, which can leave recoverable text traces. We do not ship what can be read back.
Foundations

Named standards. Replaceable parts.

You cannot audit what is not named. Below are the load-bearing choices — open, replaceable, and verifiable on someone else's machine.

Identity & Integrity
Content addressing
BLAKE3-256
Every artifact identified by its hash.
Audit chain
BLAKE2b-512, prev-hash linked
Append-only. Deterministic replay.
Canonicalisation
Deterministic JSON
Sort-stable, separator-stable, no whitespace ambiguity.
Signing
Ed25519
eIDAS-2 QES on the deployment path.
Encryption & Keys
Symmetric encryption
AES-256-GCM
Key wrapping
X25519 (age)
Key derivation
Argon2id
OWASP-recommended baseline, tunable upward.
Root of trust
TPM2 + systemd-creds · DPAPI-NG · HSM/KMS
Linux, Windows, and fallback respectively.
Documents, Search & AI
Document parsing
pypdfium2 · Docling · PaddleOCR-VL · Mistral OCR · Tesseract
Office & email
python-docx · calamine · python-pptx · odfdo · python-oxmsg · olefile
MIME detection via Magika.
Vector & lexical search
Qdrant · Tantivy (BM25) · BGE-M3 sparse
Dense, lexical, and learned, side by side.
Embeddings & inference
mE5-large (local) · pluggable LLM endpoint
Offline-capable. Validated across 11+ vendor blocks across US, EU, and China — no single-vendor dependency.
Legal-Tech & Verification
Bates numbering
Native allocator with persistent catalog
Atomic commit. Relativity export wired.
Open verifier
probans-verify, Apache-2.0
Standalone CLI. Sample Witness Pack downloadable.
Witness Pack spec
Open under Apache-2.0
The engine remains proprietary; the verification path does not.
Host platforms
Linux · Windows · macOS
The engine and its open verifier run on all three. No dedicated hardware for review. Apple Silicon native.
Deployment & Security

Designed for environments
where data must not drift.

Probans is intended for organizations that cannot treat confidential files as disposable prompt material. Deployment models, audit logging, and human review are designed for controlled regimes such as GDPR, HIPAA, and sovereign-data mandates — with final decisions held by authorized reviewers.

On-Premise Run inside customer-controlled infrastructure and security boundaries.
Private Cloud Single-tenant deployment inside controlled accounts and jurisdictions.
Air-Gapped Option Operate in isolated networks without external connectivity where required.
Operator-Controlled Inference Engine-agnostic. Deploy against any OpenAI-compatible endpoint — self-hosted SGLang or vLLM, on-premise or inside a SCIF.
Audit Logging Append-only records of ingestion, retrieval, compilation, export, and review events.
Data Residency Documents, indexes, and evidence objects remain within designated environments.
Access Controls Role-based review, controlled workspaces, and attributable activity throughout.
Human Review Outputs are built for inspection, escalation, and responsible decision-making.
Conventional intelligence says: trust the answer.
Probans says: inspect the proof.
The Probans Standard