Score shown is the deflated final; the base (company-quality) is in the facts row. Base 5.5+ = a genuinely good company; final 7+ = worth a real look given US + small raise. None of these is a conviction “back up the truck.” The move is one or two small first checks into the best-fit, best-executing names — and, per the calibration discipline, an adversarial re-verify on any single name before a check, exactly like the Winter Innovations pass. The strongest thesis-fit pick is Auxilium; the cleanest shovel is SBQuantum; the hottest-category is ZeroMark.
Tier A — the strongest, most actionable
Auxilium Biotechnologies
The best real-execution proxy for the bioelectric/regeneration thesis. It 3D-prints implantable nerve-regeneration conduits (NeuroSpan Bridge) for traumatic peripheral-nerve injury — and developed the microfabrication aboard the ISS, with NASA as a named backer. Real device, real government relationship, preclinical, tiny team on a small raise. Actionable: last money Sep 2025; small team suggests a near-term raise. Watch: long medical-device arc, capital-intensive over time.
SBQuantum
The fundable, North-American answer to QZabre — diamond quantum magnetometers, the tool every altermagnet/spintronics/nav lab buys. Won a NIST magnetic-navigation challenge. It's the pick-and-shovel that rides whichever exotic-magnetism horse wins. Actionable: just closed a fresh $4M seed, so deploying — an extension or A is the window. Watch: Canada (geo 0.75, small penalty), instrument-vendor ceiling.
ZeroMark
The Ukraine-Brave1 thesis, US and investable: hardware + autonomy that lets an infantryman shoot down a drone automatically. Real category pull (counter-UAS is the hottest defense niche) and a16z American Dynamism already in at seed. Actionable: seed closed early 2025; a Series A in 2026 is the likely window. Watch: a16z pricing means you're a follower, not a lead; category is heating fast (crowding).
InstantDB
The tools-for-thought / local-first thesis in its most fundable form: a real-time, offline-capable, sync-first database that makes user-owned, moldable software easy to build. Genuine developer traction. Its cousin ElectricSQL ($7.9M, Spark Capital) is the other credible name here. Actionable: post-seed, likely raising as it scales. Watch: infra-tool crowd (Rocicorp/Zero, Liveblocks, PowerSync); differentiation is on execution, not moat.
Enkrypt AI
The investable, execution-stage cousin of the AI-interpretability scene: automated multimodal red-teaming + guardrails + compliance for regulated enterprises (BFSI/healthcare), Gartner “Cool Vendor,” strong Yale-PhD founders. Actionable: Feb-2024 seed, due for an A — timely. Watch: the space is consolidating via M&A (Lakera→Check Point, Prompt→SentinelOne, Robust Intelligence→Cisco) — that's why I deflated the base: real product, thin moat, big-platform pressure.
Tier B — notable, with caveats
| Company | Thesis | Raised / HQ | Score | The caveat |
|---|---|---|---|---|
| Haize Labs | AI red-teaming | $250K / NY | 7.0 | Best brand in the scene + Anthropic/HF pilots, but Harmonic flags a ~$100M valuation on a tiny raise (expensive) and labs are commoditizing red-team. |
| ElectricSQL | Local-first sync | $7.9M / US | 7.0 | Category leader in local-first Postgres sync; well-capitalized, so less of an early entry than InstantDB. |
| Silimate (YC S23) | AI-for-chip-design | $2.5M / US | 6.75 | Real AI-EDA product predicting chip PPA, but a crowded lane (Astrus, ChipAgents=$134M, Cadence/Synopsys own it). |
| EZKL | ZK proofs for ML (ZKML) | $3.8M / US | 6.75 | The actual programmable-crypto core — verifiable AI inference via zero-knowledge. Real tech; crypto-adjacent demand is the question. |
| Steg.AI | Provenance / watermarking | $5M / US | 6.75 | Deep-learning content provenance (NSF, Paladin). Real, but Series A and provenance is getting crowded/regulated. |
| Protopia AI | Private AI / data protection | $8.75M / US | 6.75 | Privacy-preserving data transformation for enterprise AI; solid, but mid-raise Series A caps the EV bonus. |
| NeuroBionics | Bioelectronic / neural | $9.25M / US | 6.5 | Flexible neural-interface fibers, Future Ventures — real, but heavier raise and long clinical arc. |
| Piramidal (YC W24) | EEG foundation model | $8M / SF | 6.5 | A foundation model for brain (EEG) data — the neuro-data angle; early, unproven commercial pull. |
| Mesodyne | Thermophotovoltaic power | $5M / US | 6.75 | Light-to-power portable energy, ARPA-E + The Engine. Real, but hardware/energy is slow and capital-heavy. |
| Living Ink | Living materials | $4.9M / US | 6.75 | Algae-based carbon-negative pigments with real offtake (Fashion for Good); materials scale is slow/capital-heavy. |
| Rowan Scientific | Computational chemistry | $2.4M / Boston | 6.5 | AI computational-chemistry platform (AI Grant, Pillar) — a picks-and-shovels SaaS for the materials/chem scenes. |
| Ares Industries | Defense (low-cost missiles) | $3.6M / US | 6.5 | YC low-cost cruise missiles; very early (pre-seed), execution + procurement risk high. |
Dropped — real, but already past our filter
Over the $15M raise hard-filter or otherwise priced out of a first-check fund: Truepic (provenance, $39M B), Archetype AI (physical-AI foundation model, $48M), Ndea (Chollet's ARC/program-synthesis lab, $30.5M), Symbolica (category-theory ML, $33M), ChipAgents (AI chip design, $134M), Extropic & Femtosense (thermodynamic/edge compute, ~$14M — right at the line, warming fast), Bexorg (brain data, $23M), Novoron/Keel Labs/SiTration (~$15–18M), Diode ($15M). Basis Research (the probabilistic-world-models group) is a nonprofit/FRO — not investable, but the single best relationship target in the active-inference scene.
Pick one or two to actually engage, don't spray. My call: Auxilium (best thesis-fit with real gov-validated execution) and SBQuantum (the clean shovel, freshly capitalized so a warm relationship now positions the next round). ZeroMark if you want defense exposure and can accept following a16z. Before any check, run the adversarial re-verify on that single name — the same 5-agent pass we ran on Winter Innovations — because a formula 7.0 is a starting point, not a decision.