Clearwood is building white-box monitoring infrastructure for enterprise AI agent deployments, running cheap classifier probes on a model's internal activations instead of an always-on LLM judge to deliver as-good-or-better safety monitoring at 50x to 10,000x lower compute cost. The wedge is a genuine cost/latency arbitrage: today enterprises choose between fast-but-shallow classic NLP monitoring or high-quality-but-expensive LLM-as-judge (2x+ latency, prohibitively costly at scale). Clearwood pulls activations mid-forward-pass and classifies them, working directly on open-weight models and via a proxy model for closed-weight ones. The claimed moat is not the probe technique itself (which comes from the AI-safety research community) but the tooling around it: an automated probe-development engine, a labeled failure-mode library, and simulator environments to re-tune probes per client, because probes are sensitive to distribution shifts and a centralized lab library will not generalize to enterprise deployments (Internal).
Validation is early but real. The founders reproduced and extended Apollo Research's deception-detection work via proxy models and built a code-reuse probe with a worst-case AUC of 0.82 across model families, with signal firing 3,000 to 4,000 tokens before rollout end (early-cutoff potential) (Internal). Commercially they have four active BD conversations — a pharma company automating Phase 3 adverse-event reporting (most advanced, wants regulatory attestation), a consumer kids-athletics app expanding agent scope, plus earlier-stage finance KYC/AML and HR bias-screening use cases — but no signed design partners yet (Internal). This is a first StoryHouse conversation on a ~$1M pre-seed with no lead and no set valuation; the deal is early, Fund II Enterprise, and the founders are the draw (Pomona CEO, prior Yofi/NoFraud exit on the co-founder). SH context: check size and ownership are undefined pending the deck and DD form (Internal).
Clearwood sits at the intersection of two of the fastest-growing enterprise-software categories: AI-agent observability (est. $0.4B in 2025 growing to ~$7.1B by 2035 at a 33.3% CAGR) and AI security (~$5B growing to $20B+ by 2028) (Web). The tailwind is the production deployment of autonomous agents, which forces buyers to fund monitoring, tracing, and governance to contain incident cost and satisfy regulators. The timing thesis is that white-box / mechanistic-interpretability techniques have lived in the AI-safety research community and have almost no enterprise footprint yet, so the category leader has not been minted (Internal).
| Player | Positioning | Funding / Stage | Edge vs. them |
|---|---|---|---|
| Clearwood | White-box activation probes for agent monitoring; open-weight direct, closed-weight via proxy; probe-dev engine + failure-mode library | Pre-seed, ~$1M | — |
| Goodfire | Interpretability primitives to monitor/steer models in production | Series B, $150M @ $1.25B (Feb 2026) | Clearwood is enterprise-deployment-first with a per-client customization engine vs. research-platform posture (Web) |
| White Circle | Real-time monitoring/control of deployed models | ~$11M raised (2026) | Clearwood's white-box probes claim far lower compute cost than black-box guardrail approaches (Web) |
| LLM-as-judge platforms | Always-on generative judge / eval (incl. forked "Control Tower") | Category incumbents | 50x–10,000x lower compute; no 2x latency penalty (Internal) |
| Black-box AI security | Input/output and chain-of-thought monitoring, zero-trust/sandboxing | Well-funded (Protect AI, Robust Intelligence) | Clearwood reads internal activations, catching signal black-box tools cannot (Internal) |
Moat: defensibility is claimed to come from the automated probe-development engine, a labeled failure-mode library, and per-client simulator environments that re-tune probes to each deployment's distribution — not the probe method itself, which is public research; this remains the central unproven question (Internal).
AI-security and observability is one of the most acquisitive corners of software right now: 2025 cybersecurity M&A hit ~$96B across 400 deals, and recent comps directly rhyme with Clearwood's category — Palo Alto Networks acquired Protect AI (~$500–700M) and Chronosphere ($3.35B for AI-era observability), Cisco bought Robust Intelligence (~$400M) and Galileo (AI observability, Apr 2026) (Web). Likely acquirers are the platform consolidators (Palo Alto, Cisco, Datadog, ServiceNow) plus model labs needing enterprise-grade monitoring. At a ~$1M pre-seed entry with no set price, the return math is undefined until terms exist, but the exit-comp density in the category is a genuine positive (Web).
Matthew Levinson reached out cold via LinkedIn Sales Navigator and walked through Clearwood: white-box activation-probe monitoring for enterprise AI agents, positioned as a cheaper, faster alternative to always-on LLM-judge monitoring. Early but technically credible; both founders first-timers.