Craig Road Ventures LLC · Bellevue, Washington · Est. 2026
Intelligence you can defend.
Judgment stays with your firm.
Governed intelligence for decisions that must withstand scrutiny: Craig Road combines private-markets judgment, governed access to approved models, and evidence-linked knowledge systems — so a conclusion can be defended, and reconstructed when asked.
A conclusion your committee can check.
A fluent model paragraph and a defensible finding look identical on the page — until an investment committee, an LP, or counsel asks how the number was reached.
This is the work product: the claim, the excerpts that support it, a confidence score, the contradiction it survived, and the named reviewer holding it at the gate.
See the full sourced example and workflowIllustrative example — figures and names are fictional and redacted
Net revenue retention improved to 114% in FY2025, driven primarily by expansion within the top-quartile customer cohort.
- Source
- Quality of Earnings report v2, p.14 · ingested 2026-07-12
- Confidence
- 0.82 · Tier 2 — Uncontroverted
- Contradiction
- Management presentation cites 118% NRR. Variance traces to differing churn definitions; both shown to reviewer.
- Reviewer
- Held at gate by S. Breban 2026-07-14: confirm the churn definition the IC prefers.
A founder-led software company for institutions that must show their work.
Craig Road builds the Witness Stack — enterprise software that turns frontier-model output into evidence-linked analysis an investment committee can accept, with named human reviewers at every gate.
Commercially, it is a deliberate combination: design partnerships and governed pilots with funds, single-deal evaluations, and enterprise infrastructure licensing for firms that deploy the stack in their own environment. Every engagement is led directly by the principals.
Evidence-linked due diligence
Every material conclusion carries its supporting excerpts and source references.
Governed model access
Approved providers, allow-listed routes, and an append-only call ledger behind one proxy.
HyperWeave provenance
Sources, versioned claims, confidence, and contradictions live in one graph.
Reviewer-controlled outputs
Findings move only through named reviewer approvals. A person signs every gate.
Enterprise deployment
Hosted single-tenant or in your own cloud project, licensed to run inside your controls.
Founder & Principal
Terence Craig
Serial enterprise CTO and O'Reilly co-author. Former advisor at Toyota Ventures; former CIO at The Impact Seat.
Full profileCOO · UK commercial contact
Stephan Breban
35+ years in private markets. Former Watson Wyatt partner who built its alternatives research practice.
Full profilePublished
Privacy and Big Data
O'Reilly Media. Co-authored by Terence Craig — the privacy argument this system operationalizes.
On oreilly.comThe company
Craig Road Ventures LLC
Bellevue, Washington · Est. 2026. Founder-led: inquiries are answered by a principal, not a delivery layer.
Enterprise diligenceThe commercial progression is deliberate, and it starts with a conversation.
Initial executive conversation
Most engagements begin with a direct introduction to a principal. No forms stand between you and them.
Workflow & defensibility benchmark
A 60-minute working session: we walk your current diligence workflow on a recent deal and mark where conclusions lose provenance. Nothing sensitive changes hands.
One governed pilot or single-deal evaluation
One completed deal, your approved models, your reviewers at the gates. The corpus is fixed and the evidence links are live.
Side-by-side review
Your original memo next to the evidence-linked draft. You judge the difference with the comparison in hand.
Design partnership or enterprise license Optional path
Where appropriate: a design partnership across your pipeline, or an infrastructure license inside your own environment.
Validating Craig Road for someone else?
The Enterprise view covers architecture, deployment, data handling, and auditability — and what is implemented today versus roadmap. The Company view covers who stands behind the work.
The Witness Stack · For private-markets investment teams
Turn one completed deal into a defensible, evidence‑linked IC memo.
The Witness Stack keeps provenance intact from data room to committee: claims extracted from an approved corpus, linked to their sources, scored for confidence, checked for contradictions, and drafted into your memo format for named human review.
The same deal, with its provenance kept intact.
Before
Fluent drafts with no line back to the data room. A revised file lands mid-process and nobody knows which conclusions it touched. Reviewers re-derive the work by hand the night before committee.
After
Each claim is extracted from an approved corpus, linked to its sources, scored for confidence, and checked for contradictions — then drafted into your memo format for human review.
- 01Approved source corpusThe data room and agreed materials are fixed as the only evidence base.
- 02Structured extractionClaims, entities, and figures are pulled with their locations preserved.
- 03Evidence-linked claimsEach finding is bound to the excerpts and documents that support it.
- 04Contradictions, gaps & confidenceConflicts across sources are surfaced and every claim carries a score.
- 05 · The deliverableReviewer-ready IC memoA sourced draft in your format, gated by named human reviewers.
Illustrative example — figures and names are fictional and redacted
Net revenue retention improved to 114% in FY2025, driven primarily by expansion within the top-quartile customer cohort.
- Supporting evidence
- “Cohort-level NRR for FY2025 was 113.8%, up from 106.2% in FY2024, with expansion revenue concentrated in the top 24 accounts.” Two further excerpts corroborate the cohort concentration.
- Source
- Quality of Earnings report v2, p.14 · Data room /03-Financials/QoE-v2.pdf · ingested 2026-07-12
- Confidence
- 0.82 · Tier 2 — Uncontroverted
- Contradiction
- Management presentation (p.6) cites 118% NRR. Variance traces to differing churn definitions; the QoE figure treats downgrades as churn. Both definitions shown to reviewer.
- Reviewer note
- Held at gate by S. Breban 2026-07-14: confirm the churn definition the IC prefers before this section leaves draft.
A model can produce a conclusion in seconds. It cannot yet stand behind one. The Witness Stack is one governed path that can.
Due Diligence
Turns the approved corpus into evidence-linked analysis and reviewer-ready memo drafts — the working surface your team sees.
Governed LLM Proxy
Every model call passes one choke point: approved providers, redaction, routing, spend control, and an append-only call ledger.
Evidence & provenance — HyperWeave
Sources, versioned claims, confidence, and contradictions live in one graph, so when a document changes, the claims that depend on it are identifiable.
One connected system. The proxy governs how models are reached, HyperWeave records what the evidence supports, and Due Diligence assembles both into the memo. Change propagation across the graph is surfaced for reviewer action today; fully automatic cascade is on the roadmap and is presented to clients as such.
Faster review, gaps intact
Reviewers start from a sourced draft instead of re-deriving the work — and evidentiary gaps are stated in the draft, not hidden by fluency.
Traceable conclusions
Every material conclusion carries its excerpts and source references. Checking a claim means reading its evidence, not reconstructing it.
Explicit contradictions
When sources disagree, the disagreement is surfaced and written down — with both readings shown to the reviewer.
Named accountability
Findings move only through named reviewer approvals. The record shows who accepted what, and when.
Reconstruction on demand
When an LP, trustee, or counsel asks how a number was reached, the answer is a query against the audit trail — not a forensic project.
Knowledge that compounds
Sources, claims, and review history persist as institutional memory the next deal team can build on.
Implemented capability and roadmap are stated separately, here and in every engagement.
Available now
- Governed LLM Proxy with append-only call ledger
- Evidence-linked claims bound to sources
- Contradiction surfacing and confidence scoring
- Named reviewer gates and review history
- Memo drafts in your format
- Change propagation surfaced for reviewer action
Deployment-dependent
- Data retention and provider-use posture
- Hosted single-tenant on Google Cloud, or your own cloud project
- Self-hosted OpenAI-compatible endpoints
Engagement-specific
- Approved provider set and model routes
- Memo format and section order
- External-viewer access for counsel and third parties
Roadmap
- Fully automatic cascade of changes across the evidence graph. Today, change propagation is surfaced for reviewer action; clients are told exactly this.
Approved source corpus
The register of everything the analysis was allowed to rely on.
Evidence-linked claim set
Every material claim, bound to the excerpts and documents that support it.
Contradiction & gap register
Conflicts across sources and open evidentiary gaps, collected in one place.
Confidence & provenance record
Scores and source lineage for each claim, kept with the claim.
Reviewer-gated memo draft
A sourced IC memo draft in your house format, signed through named gates.
Audit record
Model calls, claim admissions, reviewer decisions, and retractions — appropriate to the engagement.
Built to sit inside your control environment, not beside it.
Deployment options
Hosted single-tenant on Google Cloud, or deployed into your own cloud project. Model routes include your existing provider agreements and OpenAI-compatible self-hosted endpoints.
Data handling
How your data is retained and whether providers may use it depends on the selected deployment, the providers you approve, and the retention and contractual configuration agreed for the engagement. We document the exact posture for your configuration before a pilot begins.
Access & identity
Authenticated access with role-based permissions and tenant-scoped isolation at the data layer. Read-only external-viewer access, with revocation, for counsel and other third parties when you choose to grant it.
Audit trail
Model calls, claim admissions, reviewer decisions, and retractions are written to an append-only ledger. Reconstructing any exchange is a query, not a forensic project.
Company · Bellevue, Washington
The people in the room do the work.
Two principals, no delivery layer: an enterprise CTO who writes the stack, and a private-markets researcher with three decades inside institutional investment processes.
Terence Craig
Founder & Principal · Product, architecture, delivery
Serial enterprise CTO and O'Reilly co-author (Privacy and Big Data). Former advisor at Toyota Ventures and former CIO at The Impact Seat. Hands-on with the stack — Rust, TypeScript, and LLM orchestration — which is why architectural commitments in a design partnership are made by the person who has to honor them.
Privacy, provenance, and data governance at enterprise scale
Multi-model orchestration and cost control in regulated settings
Stephan Breban
COO · UK commercial point of contact · Institutional private-markets research
35+ years in private markets and investment research. Founder and CEO of Giants' Shoulders Capital and Head of Research at Renewity; formerly a Watson Wyatt partner who built the alternatives research practice and led private-equity research. He sets the standard a memo has to meet before it leaves the building.
Manager selection, IC process, and LP reporting standards
Research operations across alternatives and private equity
A model can produce a conclusion in seconds. It cannot yet stand behind one.
A fluent paragraph and a defensible finding look identical on the page. The corpus keeps moving, and nothing tells you which conclusions a revised file just invalidated. And the record is the product: an IC, an LP, or counsel will ask how a number was reached — that answer has to exist before the question does. Craig Road exists to make provenance, governance, institutional memory, and human accountability properties of the system rather than habits of the analyst.
Start from what exists
Your architecture, your memo format, your review gates. We fit the governance you already run rather than asking you to adopt ours.
Name the uncertainty
A weak finding reported as weak is worth more than a confident one that fails in committee. The system is built to say so first.
Judgment stays yours
We produce evidence, exceptions, and review history. The investment decision, and the accountability for it, remain with your firm.
Privacy and Big Data
Published by O'Reilly Media, co-authored by Terence Craig. The privacy argument this system operationalizes: who holds personal data, what they may do with it, and what governance has to look like when the answer matters.
Conference notes and talks are collected on request.
Enterprise · Diligence & procurement
Built to sit inside your control environment, not beside it.
For the person asked to vet Craig Road: architecture, deployment posture, data handling, access control, and auditability — with implemented capability and roadmap stated separately, and named contacts for everything beyond this page.
One governed path, read top to bottom.
Every engagement runs the same sequence. The proxy is the only route to model providers; HyperWeave records what the evidence supports; nothing reaches committee without a named reviewer signing the gate.
- 01
Approved corpus
The data room and agreed materials, fixed as the only evidence base.
- 02
Due Diligence workflow
Extraction, evidence linking, contradiction and confidence analysis.
- 03
Governed LLM Proxy
One choke point: policy, redaction, routing, spend control, call ledger.
- 04
Approved model providers
Your existing provider agreements and OpenAI-compatible self-hosted endpoints.
- 05
HyperWeave provenance & audit layer
Sources, versioned claims, confidence, contradictions — one graph, append-only history.
- 06
Human reviewer gates
Findings move only through named approvals, recorded with the claim.
- 07
Reviewer-ready output
A sourced memo draft in your format, with its evidence and audit record attached.
Deployment
Hosted single-tenant on Google Cloud, or deployed into your own cloud project. Model routes include your existing provider agreements and OpenAI-compatible self-hosted endpoints where applicable.
Data handling
Retention and provider use depend on the selected deployment, the providers you approve, and the contractual configuration agreed for the engagement. The exact posture is documented for your configuration before a pilot begins.
Identity & access
- Authenticated access
- Role-based permissions
- Tenant-scoped isolation at the data layer
- Revocable read-only external-viewer access
Auditability
- Model-call ledger
- Claim admissions
- Reviewer decisions
- Retractions
- Append-only audit history
- Queryable reconstruction of any exchange
Implemented today
Governed proxy with append-only call ledger · evidence-linked claims · contradiction surfacing and confidence scoring · named reviewer gates · memo drafts in your format · change propagation surfaced for reviewer action.
Deployment-dependent
Retention and provider-use posture · identity integration · where the system runs (hosted single-tenant or your cloud project).
Engagement-specific
Approved provider set and routes · memo format and section order · external-viewer access for counsel and third parties.
Roadmap
Fully automatic cascade of changes across the evidence graph. Change propagation is surfaced for reviewer action today, and clients are told exactly this.
Request the enterprise brief.
In a live diligence process, and subject to the engagement, Craig Road provides the working materials on the right directly — from a principal, not a portal. Send NDAs and security questionnaires with your first note and they come back signed or answered.
Write to Stephan Breban, COOstephan@craigroad.com — UK commercial point of contact. Or terence@craigroad.com, Founder & Principal.
- 01Architecture overview
- 02Deployment posture
- 03Data-handling configuration
- 04Access-control model
- 05Audit & provenance model
- 06Pilot scope
- 07Commercial & support model
Contact
Continue the conversation.
Most engagements begin with a direct introduction. Write to a principal, or leave a note below — a principal replies, usually within two business days.