Virion.ai

Evidence Systems

Evidence Intelligence — Intelligence must be traceable, temporal, and composable.

Problem. Most AI systems can retrieve text. Few can answer where a claim came from, when it was true, or what evidence should override it. In regulated and document-heavy work, that gap is where trust collapses.

Solution. We design evidence substrates — not chatbots, not wrappers. Source objects, temporal facts, provenance, and governed retrieval as architecture, so AI outputs can be audited as easily as they are produced.

Market. Regulated operations, due-diligence workflows, internal knowledge platforms, and intelligence products built on partitioned or sensitive data.

Edge. Lineage, partitioning, and time are first-class concerns in the substrate — not retrofitted after a model picks an answer. Teams inherit a foundation where every claim resolves to its source, its timestamp, and its governance rules by default.

Portable Work Products

Portable Work Products — The work product should carry its own context, provenance, and handoff.

Problem. AI work products usually arrive stripped — a chat transcript, a PDF, a JSON blob. The reasoning, context, and handoff path stay locked inside the model that produced them, and the next intelligence in the chain has to start over. We built our way out of that — first with Vault, the substrate where intelligence rests, then with Capsule, the format intelligence travels in. That's why Capsule exists: we got tired of strapping different solutions together to ship the same kind of work twice.

Solution. A capsule packages mission, reasoning history, state, artifacts, provenance, and the next-step handoff in a single signed, hash-chained file — so the work can travel between people, systems, and models without losing itself. We authored the protocol and open-sourced it. Intelligence in motion shouldn't be gated; the field evolves through open cooperation in technological innovation, not through proprietary lock-in on the carrier.

Market. Multi-party AI workflows, field and offline operations, regulated handoffs, and teams treating the artifact itself as the surface of value.

Edge. The capsule is the product, not the chat that produced it. That makes it durable across model swaps, ownership changes, and the long tail of intelligence-to-intelligence handoff. Open by design — so the protocol can outlive any single vendor, including us.

Agentic Harnesses

Model Harnessing — The model is the engine. The harness is the work.

Problem. The industry is still asking which model is best. That question expired. Every frontier model has a shape — strengths, failure modes, reasoning patterns, latency profiles — and the work gets done or doesn't depending on the harness wrapped around it. The same model, under different harnesses, produces fundamentally different intelligence. Most teams haven't started building harnesses yet; they're still swapping models and hoping the next one solves it.

Solution. We started where most teams are — building a routing layer (Rolodex) for failover, attribution, and provider neutrality. We kept it. But the real work is upstream of routing: custom agentic harnesses that change what a model can actually express. Multi-agent loops, evidence-grounded reasoning, tool-use with rollback, rooms instead of threads — each harness shape unlocks a different kind of intelligence from the same underlying model. We treat harness design as an experimental discipline, not a feature checkbox.

Market. AI platform teams, applied research groups, and product teams who have outgrown "pick a model" and need a substrate for harness experimentation across providers.

Edge. Routing is solved infrastructure. Harnesses are open territory. We hold an honest position — no team has the universal harness yet, including us — and we ship the substrate that makes the experimentation legible. Cost attribution, failover, and provider neutrality come from Rolodex underneath, so harness experiments cost time, not architecture.

Supply Chain Decisioning

Supply Chain Gate — Trust should be computed before installation.

Problem. Most supply-chain tooling activates after code is already running in your environment. By that point the compromise path has already crossed the threshold the tool was meant to defend. The industry has spent a decade building runtime detection for problems that should have been pre-install decisions.

Solution. We move the trust decision earlier. Package metadata, dependency signals, and ecosystem behavior get scored at the gate — before install, before CI, before a developer's machine inherits the risk. Sherlock is the gate. A broader firewall architecture (Aidenied) is the surface it eventually composes into — but the gate ships first, by design.

Market. Security-conscious engineering organizations, regulated software teams, and developer platforms that need lightweight pre-install controls before the next dependency postmortem gets written.

Edge. The gate keeps its own dependency tree at zero. A trust tool should not recreate the problem it exists to solve — and that restraint is the design statement.

Human Workspaces

Comprehension Layer — Agents produce. Humans decide. The workspace is for the human.

Problem. AI now produces a thousand times more work than any human can read in a day. The bottleneck moved — from generation to comprehension — and most interfaces haven't noticed. They keep adding agents, threads, and outputs to the same single-stream chat surface, and the human at the end of it falls further behind every week.

Solution. We build the workspace for the human, not for the agent. Rooms instead of threads, persistent memory, routed models underneath, evidence-grounded outputs from Vault, portable handoffs through Capsule — all of it converging on a single question: can the human interpret what the system produced, and decide. That's what the workspace exists for. The agents are upstream production. The room is where production becomes legible enough to act on.

Market. Operators, analysts, decision-makers, and teams drowning in AI output that no human pipeline can absorb. Internal AI platforms where the gap between "we deployed agents" and "we made better decisions" has become visible.

Edge. The industry built agent-centric interfaces. We built a human-centric one. The workspace stands on routed harnesses, evidence substrates, and portable artifacts so the human at the end isn't asked to do what the substrate should have done — they're asked to do what only a human can do: decide.

Evaluation Loops

Fitness Engine — You cannot improve what you have not committed to measuring.

Problem. Most AI programs improve invisibly. Someone tunes a prompt on a Tuesday, the output gets better, no one writes down why. Three months later the person leaves and nobody can reconstruct which decisions held and which were luck. The tuning cycle stays informal — and informal is invisible. The deeper failure is upstream: teams build evals against proxies before they have committed to what they are actually trying to measure, and end up confidently improving the wrong thing.

Solution. We work two layers. Upstream, goal articulation as the load-bearing artifact — what is this code for, and what would "closer to that goal" actually look like as a number you can defend. Downstream, the loop that ratchets toward it: score, diagnose, propose, execute, decide, all inside an auditable ledger. Improvement becomes a property of the system rather than the property of whoever happened to be tuning prompts that week.

Market. Agent teams, AI platform groups, and operational AI programs that need evaluation as system behavior — not as a notebook one engineer maintains.

Edge. We make goal articulation part of the engagement, not an assumption. The loop is downstream of a target you can defend. Same architectural family as Karpathy's autoresearch — anything you can score, you can ratchet toward — applied beyond ML training into operational AI workflows. The fitness ledger is itself a capsule, so improvement decisions travel with the system that made them.

Vault

Evidence Substrate —

Status. Anchor implementation for Evidence Systems

Engagement fit. Best as the substrate under an applied engagement, not as a self-serve product. Vault is commercial — the substrate stays accountable.

Capsule

Portable Intelligence —

Status. Anchor implementation for Portable Work Products · Open protocol, v0.6 with v1 roadmap

Engagement fit. The protocol is open and self-serve. Implementations, capsule suite design, and integration into existing workflows are commercial engagement work.

Rolodex

Routing Substrate —

Status. Anchor implementation for Agentic Harnesses

Engagement fit. Rolodex is commercial substrate. Most engagements pair it with harness design work, where the real differentiation lives.

EveryAI.Chat

Comprehension Surface —

Status. Anchor implementation for Human Workspaces · Open-trajectory as it matures

Engagement fit. Co-designed per customer. The shape of the comprehension layer depends on the work being comprehended. This is consultative architecture, not a SaaS deployment.

Sherlock

Supply Chain Gate —

Status. Alpha — stage 3 of 6, beta opens at stage 6

Engagement fit. Not yet pipelined. Conversation-only at this stage — security teams interested in the architecture or the trajectory are welcome to reach out.

Evolve

Fitness Engine —

Status. Anchor implementation for Evaluation Loops

Engagement fit. Commissioned per customer. Goal articulation is part of the engagement, not an assumption.

Adryos

Applied Arbitrage · Asset World Ecosystem —

Status. Alpha — Artpraisal module accepted by client · Two additional modules in development

Engagement fit. Ecosystem partnership. Conversations welcome with operators in adjacent asset-world domains (collectibles, real assets, structured provenance) where the same pattern would apply.

ReadySet

Applied Arbitrage · SMB Lending Readiness —

Status. In active development · readyset.now

Engagement fit. Partnership-driven. Best paired with operators in the lending or advisory ecosystem who bring domain authority and distribution.

Fix.Now

Applied Arbitrage · Home Improvement —

Status. Approaching launch · Real-world applied AI solution

Engagement fit. Partnership and operator conversations welcome. The engine is portable; the distribution channel is the variable.

ComplianceQ

Applied Arbitrage · RegTech —

Status. Approaching launch · Real-world applied AI solution

Engagement fit. Partnership and pilot conversations open. Regulated buyers, by definition, do not move fast — but they buy substrate, not features.

PromptScript

Markdown-Native Forms —

Status. Portfolio experiment

Engagement fit. Research-stage. Conversations welcome with teams thinking about LLM-native form replacements at scale.

chatsdk

AI Router/Harness Frontend —

Status. Frontend library · Open-trajectory as it matures

Engagement fit. Currently embedded in Virion implementations. Open-source release planned as the API surface stabilizes.