Daxos Research: 50 Overlooked Companies, 4 October 2026

1. BoxLite, boxlite.ai. COMPANY 8.0, FOR DAXOS 8.0, IN. #

What it is. A microVM runtime for AI agents: a real isolated Linux kernel for every box, built on libkrun rather than Firecracker, shipped two ways, as an open-source library you can embed on a laptop and as a managed cloud. The org tagline is "the compute substrate for AI agents, light enough to embed on your laptop, elastic enough to power an agentic cloud." Harmonic: founded 2025-01, Menlo Park, headcount 2, $0 raised across 0 rounds. This is picks-and-shovels in its purest form, a runtime other people build on top of, and it sits in the lane the thesis calls the loudest asymmetry in the whole dataset.

The founders. Dorian Zheng, founder and CEO, sole. Infrastructure at Weixin and WeChat across two stints, then PhoenixAI (formerly CelerData) and ByteDance. The scale experience is real and a decade deep. Forty-two GitHub contributors appear in the repo, roughly six of them sustained, and the org maintains its own libkrunfw fork, which means the team works below the API line rather than wrapping someone else's runtime.

The special thing. Yingjun Wu, founder and CEO of RisingWave Labs, has fourteen commits in the boxlite repo. A sitting database-infrastructure CEO writing code into another startup's runtime is peer validation of a kind that cannot be manufactured, and it is worth more than any logo wall.

The measured traction. 80,367 PyPI downloads last month and 1,349 the previous day. 2,416 GitHub stars, 191 forks, 42 contributors, last push 2026-10-03. Harmonic web traffic grew from 124 to 6,853 year on year. Live pricing is published.

The red flags. 2,416 stars against four Hacker News attempts totalling eleven lifetime points, so the distribution is narrow. Harmonic lists Impa Ventures as an investor against $0 and 0 rounds, which cannot both be true. Forty-two contributors and six sustained committers at "headcount 2, nothing raised" is unexplained and is the first question. Entirely PRC employment history on a security-adjacent US infrastructure product, which is a personal-history fact rather than a capital-source fact but will surface in diligence. The HQ is weakly sourced: Harmonic says Menlo Park, the founder's GitHub says Singapore.

The score and its math. Base 6.00 + EV 1.00 ($0 raised) + Geo 1.00 (US on the record) + Chain 0.00 = COMPANY 8.00. Note honestly that two of those eight points are the $0-raise and US bonuses; the base is defended at 6.00 on a working artifact with third-party adoption.

The price that makes sense. $250k to $500k on a SAFE at roughly an $8M to $12M cap buys a structural position, and Daxos would plausibly be the first or only named institution, as it was at Hirechain. If Impa has already taken an uncommitted allocation at a lower number, match that price rather than pay up. Above a $20M cap this stops being a Daxos deal, because the ownership stops mattering.

Three questions for the first call. What funds six sustained committers at zero raised, and are any of them employees? Where is the operating entity incorporated and where do you live? Has Impa priced anything, and is there an existing SAFE? If the first answer reveals undisclosed capital or an undisclosed parent, this drops to WATCH on the call.

Contact. dorian@boxlite.ai, pattern-inferred and unproven because the domain is a Google-hosted catch-all. The real routes are GitHub issues on boxlite-ai/boxlite, the Discord linked from the site, and the blog at dorianzheng.github.io.

2. Vaults.fyi (Wallfacer Labs), vaults.fyi. COMPANY 8.08, FOR DAXOS 7.5, MEET. #

What it is. One API that lets a wallet or a fintech embed DeFi yield without building the plumbing. Coverage, per the company's own documentation index read 2026-10-04: over 1,000 yield strategies, over 80 protocols, over 20 networks including Ethereum, Base, Arbitrum, Optimism and Solana. The important detail is that it is not a data company. The API returns ready-to-sign transaction calldata with no intermediary contract, alongside market data and portfolio positions. Read plus write. That distinction decides the whole case, because the pure-data version of this business has already been tried and buried: the nearest pure-data competitor raised $18.7M and is recorded OUT_OF_BUSINESS.

The founders. Ryan Rodenbaugh, CEO and co-founder; Tyler Wallace, co-founder, product and business operations. Both US-educated and US-employed throughout. Nine staff, up from seven a year ago. There is no technical co-founder on an API company, which is the single largest hole in the case and fails test T6.

The special thing. Three years and nine people with zero recorded outside capital, in a category where the direct comparables raised $9M to $24M. If the bootstrapping is deliberate, this is the Steel.dev shape exactly. If a round failed to close, it is the opposite. That one question is the deal.

The measured traction. The npm package @vaultsfyi/sdk ran 30,049 downloads last month and 44,210 in August 2026, holding between 28k and 44k every month since April 2026, checked against api.npmjs.org. Arbitrum Earn is live at portal.arbitrum.io/earn. Named customers with published case studies: Kraken Wallet, Gauntlet, Maple Finance, Jumper, Arbitrum.

The red flags. No technical co-founder. Every customer attribution reachable traces back to vaults.fyi's own documentation rather than to the customer's site, so the logo wall is unconfirmed from the other side. No public pricing page, since vaults.fyi/pricing returns 404, and no verified revenue. Harmonic records no HQ at all. The $0 at three years is unexplained. The policy advisor is simultaneously General Counsel at two companies and Executive Director of "The Three Arrows Family Foundation", which could not be resolved.

The score and its math. Base 6.08 + EV 1.00 + Geo 1.00 + Chain 0.00 = COMPANY 8.08. It passes ten or eleven of the twelve house tests, failing T6 and partially T11.

The price that makes sense. No round is announced, so entry is a SAFE. Against Enso at $9.2M and Superform at $13.9M for a similar stage, a $15M to $25M post is the defensible zone for $250k. Above $30M the ownership stops mattering and this becomes a pass on stage.

Three questions for the first call. Why zero capital in three years, and was a raise attempted? Who owns engineering, and does a technical co-founder exist in substance rather than in title? What is the paid customer count and the revenue, and will Kraken, Gauntlet and Maple confirm from their side? Daxos has something real to add here, which is unusual: the customer set is wallets and exchanges, and the book holds Xverse, Botanix, pump.fun and Trendex.

Contact. ryan@wallfacer.io, medium confidence and the address on the Harmonic person record. Note that ryan@vaults.fyi is invalid because the vaults.fyi domain has no MX record. Backups: rrodenbaugh@gmail.com, linkedin.com/in/ryanrodenbaugh.

3. Microsandbox, microsandbox.dev. COMPANY 8.25, FOR DAXOS 7.0, MEET. #

What it is. Every workload gets its own kernel. Instead of a container boundary, model-written code runs in a microVM built on libkrun, with per-sandbox network egress policy, credential placeholder substitution, disk snapshots and resume, and SDKs in Rust, TypeScript, Python, Go and Ruby. Three deployment paths: locally with one command and no daemon or account, on their managed cloud, or on your own metal. There is an MCP server, a Kubernetes integration, a benchmarks repo and a published eleven-way competitor comparison. This is the substrate underneath browser-agent infrastructure, which is to say the layer below Steel.dev.

The founders. Stephen Akinyemi (GitHub appcypher), co-founder and CEO; Tochukwu Nkemdilim (GitHub toksdotdev), co-founder and CTO. Akinyemi has no computer science degree: he read Mass Communication at the University of Lagos, was a contributing engineer at Wasmer in 2018 and a senior Rust and WebAssembly engineer at Fission in 2022, and has ten years of public artifacts under one handle, 110 repos and 832 followers since 2016. Nkemdilim was a Cloudflare systems engineer for the two years to 2026-03 and did Mozilla's "Increasing Rust's Reach" programme in 2018. The CEO sits in Nigeria and the CTO in Toronto, so San Francisco is the YC batch address rather than where the team lives.

The special thing. The artifact, and the order of events. 8,563 stars and 65 outside contributors on an Apache-2.0 Rust microVM runtime, built by two engineers who are themselves the top two committers at 522 and 352 commits. The repo was created 2024-10-03, fifteen months before the company. The company exists because the project already worked, not the other way round, and that is the hardest thing on this list to fake.

The measured traction. The npm package microsandbox ran 741,389 downloads in the 30 days to 2026-10-03, against 700,716 in September, 451,934 in August and 220,886 in July. Maintainership verified: appcypher and toksdotdev, repository git+github.com/superradcompany/microsandbox. GitHub: 8,563 stars, 467 forks, 65 contributors, 107 open issues, last push 2026-10-04, weekly commits running 12 to 84 over twelve weeks. Hacker News front page 2025-05-30 at 402 points and 186 comments.

The red flags. Nothing is monetised and the business model is the whole risk: the runtime is Apache-2.0 and free to self-host, and the managed cloud is behind "request cloud access". 741,389 npm downloads against 2,482 monthly website visits and 594 Twitter followers should be read sceptically, because package-manager counts for an infrastructure SDK are inflated by CI runners and registry mirrors. 107 open issues against three people is a real support-load failure mode. The CEO's prior company, Zerocore AI, ran 2024-05 to 2026-05 and sits in Harmonic at $0 and 1 headcount, so it wound down. "Teams building with microsandbox: OpenWork, Keystroke and many more" is adoption language, not a paying customer. The YC company page returns 404, so F26 membership rests on their own badge plus Harmonic's record.

The score and its math. Base 6.25 + EV 1.00 ($500k raised falls in the $0 to $1M tier) + Geo 1.00 + Chain 0.00 = COMPANY 8.25, the joint-highest company score in the set.

The price that makes sense. The company is right and the price is the question. A YC Fall 2026 company with 8,563 stars will not raise at the $2M Harmonic's dilution model implies, and a post-demo-day seed for this profile prices where $250k buys roughly one percent. That is the Outpost and Fieldstone test, and it is the only reason this is not an IN. Go only at a cap where $250k is a named line.

Three questions for the first call. What is the round and the cap, and is it still open? Does anyone pay, specifically OpenWork, Keystroke or Starsling? What happened to Zerocore AI? What Daxos can add is concrete: the fund owns Steel.dev, which sits directly on top of what microsandbox sells, and an introduction between the two is worth more than the cheque.

Contact. Not email. The site publishes a direct booking link to the CTO at calendar.notion.so/meet/toksdotdev/microsandbox, plus a Discord at discord.gg/T95Y3XnEAK. LinkedIn: linkedin.com/in/appcypher, linkedin.com/in/toksdotdev.

4. Lightpanda, lightpanda.io. COMPANY 7.92, FOR DAXOS 7.0, MEET. #

What it is. A headless browser that parses HTML and executes JavaScript but never renders pixels. Every other headless browser is Chrome with the screen turned off, carrying a rendering engine it never uses. Lightpanda threw that away and wrote a new engine in Zig. The payoff is resource cost: the published crawler benchmark, 25 workers over 933 links, is 4.81 seconds and 123MB against headless Chrome at 46.70 seconds and 2.0GB. At agent scale that is the difference between a viable unit cost and an unviable one. Version 1.0 shipped 2026-10-02 with 10,000 commits and 1.7M passing Web Platform Test subtests, which retires the beta objection the sourcing note carried.

The founders. Francis Bouvier, CEO; Pierre Tachoire, CTO; Katie Hallett, also recorded as Katie Brown, co-founder and COO. Eleven staff, up from eight a year ago. Paris.

The special thing. Almost nobody in this lane owns an engine. Everyone else rents Chromium. Owning the engine is why the benchmark is 10x faster and 16x lighter, and it is not a thing a competitor copies in a quarter.

The measured traction. github.com/lightpanda-io/browser: 35,983 stars, 1,697 forks, 67 contributors, last push 2026-10-04. npm @lightpanda/browser 20,329 downloads last month, up from 6,487 in April 2026. A Show HN at 319 points and 137 comments. Harmonic web traffic 54,961 a month.

The red flags. The top contributor by a wide margin is not a founder and does not list the company as his employer: 4,259 commits against the CEO's 475. The AGPL-3.0 licence blocks closed-source embedding and no commercial exception was found on sale, which is a direct problem for the agent companies that would be the natural buyers. The company has never published its round size and Harmonic's $286k is the only figure on record. The blog names founders of Mistral, Hugging Face and Dust as angels and none appear in Harmonic's investor array. No French legal entity was found in the government SIREN registry.

The score and its math. Base 6.67 + EV 1.00 ($286k) + Geo 0.25 (France, EU) + Chain 0.00 = COMPANY 7.92. Geo is the only reason this sits below Microsandbox and Cua; the thesis warns that a mechanical geo application under-rates exactly the non-US teams the fund has bought three of.

The price that makes sense. At $286,000 recorded total a $250k cheque is a real position rather than a rounding error, so T9 passes today. But ISAI led the pre-seed and will have pro-rata, and the 1.0 release is precisely the event that triggers a proper seed. If that seed is already assembled at a size where $250k is the twentieth line, say "pass on stage, not quality" and move on. Daxos adds modest value to a Paris deep-tech team, since the fund's network is crypto and US consumer.

Three questions for the first call. Is there a commercial licence for AGPL-incompatible embedding, and what does it cost? What is the round the 1.0 release is triggering, and at what size? Who is the 4,259-commit top contributor, and is he employed? Move quickly: the 1.0 release means the pricing window is closing.

Contact. francis@lightpanda.io, medium-high confidence, confirmed by two Harmonic records and the CTO's matching pierre@lightpanda.io. linkedin.com/in/francisbouvier.

5. Pantheon AI, pantheonai.co. COMPANY 7.75, FOR DAXOS 7.0, MEET. #

What it is. Automation of the front end of commercial insurance underwriting: it reads incoming broker emails and attachments, pulls structured policy data out of them, organises the submission, and feeds rating and quoting workflows. The website is not a website. pantheonai.co returns sixty characters of text and a React shell, which reads as a red flag until you open the bundle. The 2.75MB application bundle contains Leaflet, which in this context is geographic exposure accumulation; TinyMCE 8 with a paid Tiny Cloud licence key, for document and referral-template editing; NAICS industry classification strings such as "Commodity Contracts Brokers and Dealers"; an analytics view called "Monthly Submissions By Brokerage"; and the strings "Start Instanda Quote" and "Update Instanda Quote". That last pair matters, because INSTANDA is a real UK policy administration platform with $93M raised and 239 staff, and a two-way quote integration against it is something you build for a customer, not for a demo. The domain root serves the product itself. There is no marketing site at all.

The founders. Greg Karabinos, co-founder, FCAS-credentialed actuary with four years as Principal Data Scientist at Liberty Mutual; John Karabinos, his brother, four and a half years engineering at Dropbox. Senior team: Andrew Hannon (VP Engineering, ex-Dropbox Staff), Sean Stephens (ex-Dropbox Senior), Greg Flood (career commercial underwriter out of CNA and Relm), Daniel Bryck (Director of Operations). Seven people.

The special thing. Seven people, two and a half years, zero outside capital, and a shipped enterprise product. The combination is rare and this team explains it: a credentialed actuary who knows what the output has to be, plus three Dropbox engineers who know how to build it.

The measured traction. No revenue figure exists publicly. Third-party-observable signals, all 2026-10-04: a live two-way INSTANDA quote integration found in the bundle; two separate public product-feedback boards, pantheon-ai.supahub.com and a customer-branded pantheon-ai-uip.supahub.com, which is the strongest single hint of a real customer; LinkedIn followers 158, up from 50 a year ago; web traffic 1,340 a month against 146 a year ago.

The red flags. Not one customer is named anywhere. No revenue, pricing or ACV is published. INSTANDA, the platform Pantheon integrates with, has launched its own AI-powered underwriting workbench, so the integration partner is also a likely displacer. The two engineering leads' prior venture together, Valt, did not survive. Federato at $192M and hyperexponential at $92M are far better capitalised, so T10 fails on Federato.

The score and its math. Base 5.75 + EV 1.00 + Geo 1.00 + Chain 0.00 = COMPANY 7.75. Ten of twelve house tests pass; T7 fails for want of an independent number and T10 fails on Federato.

The price that makes sense. There is no round, so this is a SAFE priced by the founders. A bootstrapped seven-person team with real revenue will know what it is worth; one without revenue may take a friendly cap. The defensible band for $250k is a $10M to $18M post. Above $25M the ownership stops mattering.

Three questions for the first call. What is revenue, and who are the customers by name? Why no capital in thirty months, and was a raise ever attempted? What does INSTANDA MAX mean for a product that sits on top of INSTANDA? Good answers move this to IN quickly; "we are pre-revenue and the founders have been funding it" moves it to WATCH. Be honest that Daxos brings money and little else here, because nothing in the book touches commercial carriers, MGAs or wholesale brokers.

Contact. gkarabinos@pantheonai.co and john@pantheonai.co, both verified live on 2026-10-04 with RCPT 250 on a non-catch-all domain, so both mailboxes genuinely exist. High confidence, and the best contact data in the set.

6. Viridian Space Corporation, viridianspace.com. COMPANY 7.25, FOR DAXOS 7.0, MEET. #

What it is. Air-breathing electric propulsion: a plasma thruster that ingests the rarefied atmosphere at 200 to 300km and uses it as propellant, so a satellite can hold very low Earth orbit indefinitely instead of decaying. VLEO is worth holding because imaging resolution and link budget both improve sharply as you drop. The stated long-term ambition is to operate its own VLEO satellites, but all ten federal awards are for propulsion, and the propulsion component is the real company.

The founders. Rostislav "Slava" Spektor, CEO, and Matthew Solomon Feldman, CTO. Two Princeton electric-propulsion PhDs who ran the Electric Propulsion and Plasma Science lab at The Aerospace Corporation for a combined 22 years. Kris Rydberg joined as Chief Business Officer around 2026-03. Seven on Harmonic, eight on LinkedIn. Delaware corp 6155488 incorporated 2021-08-10, CAGE 953D5, SAM.gov active through 2027-06-08.

The special thing. A granted US patent on the core component, US 12,286,943 B2, a turbomolecular air-scoop granted 2025-04-29, assigned to the company with both founders named as inventors, plus $5.34M of federal money against $2.5M of equity. That ratio, 2.1x non-dilutive to dilutive, is the Cosmic Shielding template executed better.

The measured traction. $5,336,046 across ten federal awards, verified line by line on USAspending and SBIR.gov: three Phase IIs totalling $4,153,631 (NSF 2527072 at $1,218,494; USAF FA2541-25-C-B102 at $1,687,447; USAF FA8649-25-P-0355 at $1,247,690), six Phase I and STTR-I awards totalling $607,561, and DARPA HR001126CE026 at $349,855. Phase I to Phase II conversion is three of five. A five-year AFRL CRADA was announced 2026-02-04. Web traffic 4,613 a month, LinkedIn 2,676.

The red flags. Nothing has been built that runs on air. The NSF Phase II abstract is future tense, and the company's own careers page advertises a role to "build the new version of the thruster that can run on air and test it in a third-party vacuum chamber", so the inlet and the thruster have never been integrated and they do not own a chamber that can test the integrated article. The direct competitor Kreios Space has roughly EUR 10M and a satellite bus under contract with Kongsberg NanoAvionics for a first flight in early 2027, so the rival flies first. A third-party claim that Viridian "demonstrated the first prototype of their air-scoop and an oxygen-compatible Hall thruster" traces to no primary source and has not been corrected. Spektor's Aerospace Corporation staff bio is still live in the present tense. The CBO is absent from the company's own team page after six months and his start follows the DARPA Embedded Entrepreneur award by weeks, so he is probably a DARPA-funded placement rather than a permanent equity hire. The website team page has shown four staff, unchanged, since October 2024. Zero commercial customers, zero launch manifest, nothing flown. CourtListener and California FTB status were both blocked and could not be cleared.

The score and its math. Base 5.50 + EV 0.75 ($2M, the $1M to $3M tier) + Geo 1.00 + Chain 0.00 = COMPANY 7.25. Federal award money is contract revenue and is correctly excluded from the raise under the AP14 rule.

The price that makes sense. Find out the November 2025 round's size, instrument and valuation before anything else. At up to roughly a $20M to $25M post, $250k to $500k buys a real position at eight staff and clears T9. Above about $30M, pass on stage and not on quality. A Gaingels SPV already raised $517,032 to invest here in November 2025, so a cheque of this size is precedented on this cap table.

Three questions for the first call. When does the integrated inlet-and-thruster article fire on air or a simulated VLEO gas mix, in which named chamber, and what thrust and Isp will you publish? Is there a launch slot or a bus partner? What was the November 2025 round's price and instrument? This is the company thesis asymmetry E asks for, a sub-$5M space supplier on the Cosmic Shielding template with better founders and a cleaner record than the one the fund already owns.

Contact. slava@viridianspace.com, high confidence, printed by the company in the author footnote of its own conference paper IEPC-2024-244 and listed as business contact on SBIR award FA8649-24-P-0909. CTO: matthew@viridianspace.com.

7. eunomia-bpf, eunomia.dev. COMPANY 7.08, FOR DAXOS 7.0, MEET. #

What it is. eBPF-based runtime infrastructure sitting underneath AI agents, doing three things at the system layer: observability, policy enforcement and safety. The foundation is bpftime, a userspace eBPF runtime covering uprobes, USDT, syscalls, XDP and GPU paths, with LLVM JIT and AOT compilation. On top of it sit AgentSight, zero-instrumentation observability that correlates eBPF syscall tracing with TLS traffic to show what an agent actually did to the filesystem, the network and the process tree; and ActPlane, policy-as-code enforcement over process trees, files, network effects, data-flow labels and workflow gates. Everything is MIT licensed. AgentSight already instruments Claude Code and Codex.

The founders. Yusheng Zheng, founder, a currently-enrolled second-year PhD student at UC Santa Cruz. The second contributor is GitHub "Officeyutong" with 116 commits, almost certainly Tong Yu. Headcount 2 on all three Harmonic measures.

The special thing. bpftime was published at OSDI 2025, one of the top two operating-systems conferences in the world: "Extending Applications Safely and Efficiently", Zheng first author with Dan Williams of Virginia Tech and Andi Quinn of UCSC, USENIX OSDI 25, Boston, pages 557 to 574, verified at usenix.org. This is the hardest-to-fake credential in the entire set of sixty-four.

The measured traction. bpftime 1,583 stars, 186 forks, 40 contributors, last push 2026-10-03. AgentSight 724 stars, 109 forks, last push 2026-10-04. The org holds 153 public repos and 772 followers, and the site claims over 9,950 stars org-wide. Harmonic web traffic 10,380, the highest in its batch. Zero revenue, zero named customers.

The red flags. There is probably no company. No pricing page, no contact page, no enterprise page, all 404, and no incorporated entity was found anywhere. Harmonic's own description sets the commercial posture as "scoped advisory work, design-partner pilots, and implementation projects", which is consulting rather than product. The founder is mid-PhD. Everything is MIT licensed with no open-core split, so there is nothing to sell today. Geography is contested: Harmonic says San Francisco, the GitHub org says United Kingdom and the founder's profile says London. IP ownership is genuinely unclear given ISCAS and four universities among the co-authors.

The score and its math. Base 5.08 + EV 1.00 + Geo 1.00 + Chain 0.00 = COMPANY 7.08. The base is the lowest in the top ten and the honest reading of it: this is a research project, not yet a company.

The price that makes sense. $250k as the entire first round, structured as a SAFE after incorporation. There is nothing to wire money into today, which is why it is a MEET.

Three questions for the first call. Do you intend to incorporate in the US, and would you defer or go part-time on the PhD? Who owns the IP, given UCSC, ISCAS and the co-authors? Where is the open-core boundary, and would you name two design partners? Conditions before money: a US entity with clean IP assignment, the founder full time or on a credible leave of absence, a named open-core boundary, and two design partners by name. None is exotic, and all four are the ordinary work of turning a strong research project into a company, which is the exact stage at which this fund claims its edge. Of everything in this set, this is the one to call first: it is the layer the thesis names as "never searched", built by someone who published it at the top venue in the field and has not yet been found by anyone.

Contact. team@eunomia.dev, confirmed, published in the GitHub organisation profile with Cloudflare email routing, so it is deliverable. Also linkedin.com/in/yunwei37, GitHub @yunwei37, yunwei37.com. High confidence.

8. Convallax, convallax.com. COMPANY 7.0, FOR DAXOS 7.0, MEET. #

What it is. European cash-settled calls and puts on the YES price of Polymarket binary markets. Collateral is USDC, option positions are ERC-1155 tokens, settlement is on Polygon. Price discovery runs through a request-for-quote relay: a taker opens a quote request, market makers compete over Server-Sent Events and REST, the relay signs an EIP-712 order, and the taker calls fill() to mint the option, lock collateral and pay premium atomically. Expiry uses a TWAP resolution price with an EIP-712 attestation. There is a smart-margin engine and a documented path to become an independent market maker with an API key. This is infrastructure rather than an app, which is the distinction the house has already said it prefers in prediction markets, and the thesis names "prediction market infrastructure rather than prediction market apps" as an uncovered adjacent lane.

The founders. Kushal Mungee and Rahul Rai. Mungee's entire record is three internships: quant trading in 2024, equity trading in 2023, software engineering in 2022. Rai is the reason to take the call, and he is absent from both the Harmonic company record and the Convallax website.

The special thing. Rahul Rai co-founded Gamma Point Capital in 2021 and sold it to BlockTower Capital for a reported $35M in June 2021, at twenty-four or twenty-five, then ran BlockTower's market-neutral book. The acquisition is confirmable in SEC EDGAR rather than only in press: CIK 0001756971 carries former names "Shyn Capital Fund (US) I, LP" and "Gamma Point Capital Fund I (US), LP", now files as "BlockTower Gamma Point Fund LP" with Matthew Goetz and Ari Paul as executive officers, and shows $22,346,871 sold on the Form D of 2021-08-02. A quant who has actually run an options book is the right person to build an options venue.

The measured traction. Zero real-money usage, and this must be said first. Four smart contracts are deployed and carry bytecode on Polygon Amoy testnet, chain 80002, verified by eth_getCode. The same addresses return 0x on Polygon mainnet, so nothing has ever settled with real money. The live API at api.convallax.com/v1 returns six active markets and 1,616 registered option series across roughly 120 Polymarket markets. A Sherlock collaborative audit report dated 2026-07-02 exists and downloads at 895,259 bytes.

The red flags. The product is testnet only and the company does not say so on its marketing site: "Trade now" leads to a Polygon Amoy deployment with testnet USDC. The site has no team page, so the one genuinely credentialed founder is invisible on the company's own property. Rai's LinkedIn headline reads "Forbes 30U30" when the list is Forbes India 30 Under 30, which flatters by shorthand. His record shows no role between leaving Gamma Point in 2023-04 and Convallax in 2026-03, an unexplained three-year gap. Harmonic's web traffic of 14,010 is identical at 14, 30, 90, 180 and 365 days, so it is one stale measurement rather than evidence of traffic. Options on Polymarket YES tokens is CFTC-shaped activity and the Terms contain no US-person restriction, only a sanctions representation. LinkedIn followers: 22. Twitter is down from 3,469 to 3,247.

The score and its math. Base 5.00 + EV 1.00 ($500k accelerator) + Geo 1.00 + Chain 0.00 = COMPANY 7.00. Two of the seven points are bonuses; the base says "good idea, right person, nothing live".

The price that makes sense. Harmonic's $3M post is its own 20-percent-dilution model and not a reported figure, so the real number has to come from the founders. $100k to $250k at anything near a $3M to $8M post buys a real position.

Three questions for the first call. Who writes the contracts, given one founder is three internships deep and the other is invisible on the site? When does mainnet ship, and with which named market maker? What happened between April 2023 and March 2026? Daxos has something to add: Trendex, worm.wtf and Wager give it prediction-market relationships, and the thing this company needs most is market makers. Alliance is already in, and Alliance sits on the pump.fun cap table, so this is inside the referral graph that actually produces the fund's positions.

Contact. kushal@convallax.com, held by Harmonic as company primary_email, medium confidence on a catch-all domain. support@convallax.com is published in the documentation FAQ. rahul@gammapoint.capital sits on Rai's person record but belongs to the prior fund.

9. Lexe, lexe.app. COMPANY 7.17, FOR DAXOS 6.5, MEET. #

What it is. Lexe runs a Lightning node for you that you still fully own. The node executes inside an Intel SGX enclave on Lexe's hosting, so it stays online and can receive and settle payments while your phone is off, and Lexe cannot reach the keys. Pricing is 0.5 percent per payment with no subscription, plus a daily interest rate on inbound liquidity you actually use. What the sourcing note missed is that the product has quietly split in two. Since December 2025 Lexe has shipped lexe-sdk in Rust, lexe-sidecar-sdk, a Python package and a CLI, and the homepage now says "Point an AI at Lexe's SDK docs to integrate Lightning payments within a couple of minutes." The consumer wallet is the demo. The thing with growing numbers is a managed-Lightning-node API sold to developers, which moves the company out of a category the house has already buried and into picks-and-shovels.

The founders. Max Fang, CEO, and Philip Kannegaard Hayes, CTO, both from 2022-04. Hayes spent 2018 to 2021 at Facebook as a Diem and Libra Core developer, which is production Rust consensus engineering. They co-founded Blockchain at Berkeley together in 2015, so this is a ten-year working partnership. Three staff, flat for a year, plus one advisor.

The special thing. The trust claim is third-party verifiable, which almost nothing in this lane is. Lexe builds the node reproducibly and remotely attests the enclave measurement to your phone when the app connects, so a stranger can rebuild the published source, derive the same measurement, and check that the thing holding the keys is the code they read. The README names the stack: LDK for Lightning, BDK for on-chain, Fortanix EDP for SGX. Release commits on 2026-10-02 are literally titled "release (1/2): node-v0.10.6 (Reproducible commit)."

The measured traction. Two products, two completely different pictures. Consumer is dead: three iOS ratings since the App Store release on 2026-03-23, "100+ Downloads" on Google Play, web traffic 660 a month. Developer is the live one and the numbers are independent: crates.io lexe 4,225 total downloads with 3,881 recent, created 2026-02-26, now v0.1.25; PyPI lexe-sdk 871 downloads last month, 371 last week, 134 last day, across 23 releases. Twitter doubled from 345 to 700 followers. The $998,500 raise is confirmed to the dollar by SEC Form D, CIK 0002084253: offering $998,500, sold $998,500, remaining $0.

The red flags. Nothing dishonest was found, and several checks came back actively clean, so the flags are category and execution. The house has already issued a category-death verdict on this exact category: at Tuyo 4.50 it wrote that self-custodial wallets and consumer crypto money apps have "ZERO standalone venture outcomes", and Lexe is a self-custodial wallet. Four years in with one sub-$1M round, three staff and three App Store ratings is the velocity flag the house raised at Lonestar. The Form D was filed 2025-11-21 against a first sale of 2025-05-21, six months past the fifteen-day deadline. The homepage has softened "Intel SGX" out of its visible body text, where it now says only "secure hardware"; the phrase survives in the og:description tag and the README, so this is marketing softening rather than contradiction, but it is the right direction to watch given Intel's deprecation of SGX on client parts. The newsletter form is broken with Mailchimp error text leaking into the page. Fulgur Ventures is the gatekeeper of this entire lane, is the common investor across three companies in this set, and is nowhere in the Daxos referral graph.

The score and its math. Base 5.17 + EV 1.00 ($998,500, just inside the $0 to $1M tier) + Geo 1.00 + Chain 0.00 = COMPANY 7.17.

The price that makes sense. There is probably nothing open: the Form D shows totalRemaining $0, Harmonic's currently_raising is null, and seventeen months have passed on three heads, so a raise is likely near but unconfirmed. That is the main fit deduction. If one opens the entry is unusually good, because a $998,500 SAFE across 26 investors averages $38,000 a head, so $250k would be the single largest position in the company. $250k at or under an $8M post. Above $12M, pass on price.

Three questions for the first call. Is anybody paying for the SDK, and who? What Lightning volume ran through Lexe nodes this month? Is a round open, and is Fulgur leading it? If the answer is a named integrator with volume, this is a cheque. If the answer is the App Store, it is a pass on a dead category. What Daxos adds is concrete rather than a logo: Xverse is a Bitcoin wallet and Lexe sells Lightning node infrastructure to Bitcoin wallets, and Bound and Saturn both need Lightning rails.

Contact. max@lexe.app, medium-high confidence; Harmonic carries it as the only exec_email and the Form D names Maxwell Fang as Executive Officer and Director. Backups: hello@lexe.app published on the GitHub org profile, maxfangx@gmail.com, and philiphayes9@gmail.com from the CTO's own signed commits.

10. Cua, cua.ai. COMPANY 8.25, FOR DAXOS 6.0, MEET. #

What it is. Steel.dev gives an agent a browser. Cua gives it the rest of the machine: an open-source container runtime that boots a full operating system, macOS, Linux or Windows, inside a lightweight VM an agent can drive. The positioning has moved since the sourcing note. The site now leads with Fleets: "Scale computer fleets for every agent. Run computer-use agent training, eval and data-generation workloads in parallel on Linux, Windows, macOS and Android machines. Fork from snapshots, reproduce failures, and turn agent activity into training data." That is a deliberate move from developer tool toward the reinforcement-learning environment market, which is where the 2026 budget actually sits. They also publish Cua-Bench, whose headline result is that the best frontier agent clears just six of twenty-five expert KiCad tasks.

The founders. Francesco Bonacci, founder; Dillon Dupont, CTO; Robert Wendt, Chief Infrastructure Officer; Alessandro Puppo, co-founder, departed 2025-07. Eight staff. San Francisco. YC.

The special thing. Full macOS virtual machines for agents at 97 percent native CPU speed on Apple Silicon. Apple's licence terms require macOS to run on Apple hardware, so the Linux-container-first competitors cannot simply copy this. Separately, Robert Wendt was founding engineer at Brev.dev, which NVIDIA acquired, so the sandbox-infrastructure experience on this team is bought and paid for.

The measured traction. The best-evidenced developer adoption in the set. github.com/trycua/cua: 28,033 stars, 1,988 forks, over 100 contributors, last push 2026-10-04. PyPI cua-computer 90,992 downloads last month and cua-agent 27,017. npm @trycua/cua-driver 6,381,755 downloads last month at a steady 113k to 144k per day. Five separate Hacker News front pages.

The red flags. A co-founder joined and left inside seven months and is now running a stealth company. Two of the three remaining technical leaders carry a parallel current venture, Dupont at Viralmind LLC and Wendt at Flakery. 1,017 open issues against 28,033 stars is a support load eight people cannot carry. The pricing page fails to load, so the commercial product could not be observed at all. The YC batch label is inconsistent across Harmonic, the Launch HN and the YC page, X25 against P25. The MIT licence means no licensing revenue.

The score and its math. Base 6.25 + EV 1.00 + Geo 1.00 + Chain 0.00 = COMPANY 8.25, joint-highest in the set, and FOR DAXOS 6.0, held down entirely by stage and price rather than by quality. That eight-tenths-of-a-point gap between the two numbers is the clearest illustration in this report of why the ranking rule is fit first.

The price that makes sense. $250k on the seed, only if Daxos is named rather than stacked. A YC company with 28,000 stars, 6.38M monthly npm pulls and a Series-B-funded competitive set does not stay at a $500k pre-seed; the next round will be a real seed, probably oversubscribed, and Daxos has no YC edge. The house has written this sentence before, at Fieldstone Bio and Outpost: the rating stays, the deal fails on stage. Trendex is the counterexample in the book, a YC company Daxos did get into, so the channel is not impossible.

Three questions for the first call. Is a round open and at what price? Who services 1,017 open issues? What does Fleets actually bill today, and why did Puppo leave after seven months? If the round is already at a cap where $250k is the twentieth line, pass on stage and say so plainly.

Contact. francesco@trycua.com, medium confidence on a catch-all domain. Harmonic also holds f@trycua.com and francesco.bonacci@outlook.com. linkedin.com/in/francesco-bonacci-70428a121.

L1. Bitcoin-native infrastructure #

Funnel. 3,327 raw rows from five similarity runs and 66 keyword queries, 2,462 unique, 10 dropped on the exclusion file (Botanix, Bound, Bubblemaps, Saturn, Xverse, Zoth, Artifact, MintGreen, Symbolic Capital, Tuyo), 874 scored candidates after a Bitcoin-plus-lane keyword pass, the $15M cap, a 60-headcount cap and a dead-site check, 15 written up in the lane plus the eight that reached the ranked fifty.

Best two. Lexe (#9, fit 6.5), the only company in the lane whose security claim a stranger can verify by rebuilding the source, and whose real business turned out to be a managed-Lightning-node API rather than the consumer wallet. Second (#11, fit 6.0), a genuine Bitcoin protocol maintainer shipping the Rust implementation of Ark, the first new Bitcoin payments primitive since Lightning, live on mainnet since 2026-06-09.

Where Daxos is strong and where it is blind. This is the fund's most worked lane and it shows: ten exclusion hits, five portfolio positions, and every one of those five at the application tier. The memo states the overweight plainly: the book is heavy in L2s (Botanix), wallets (Xverse), ordinals (Ord.io), BTC stablecoins (Bound) and BTC yield (Saturn), so the value left is one or two layers down, in protocol research, payments primitives, non-custodial credit mechanics and mining software. Three specific blind spots are named with evidence. Insurance and custody counterparty: AnchorWatch is "the kind of thing a fund holding Xverse and Bound has no counterparty for today", and institutional custody is the one piece of the Bitcoin stack the fund has no exposure to at all. Blockspace plumbing: Rebar Labs sits between Ord.io's order flow and the miners, and almost nobody else is in Bitcoin MEV. And the BTC credit secondary market: every BTC lender the fund sees eventually needs one and nobody is building it. Against that, the lane is where five correlated exposures to one asset's cycle sit, with Ord.io at one headcount and Botanix flagged OUT_OF_BUSINESS by Harmonic.

Attention: less at the top, more underneath. Do not buy another wallet or another BTC-collateralised dollar; Bima was dropped from serious consideration precisely because it duplicates Bound. Buy the layer the five positions all depend on. The one structural warning the lane carries is that Harmonic's funding_rounds array came back empty for every top-15 company, so $0 here means "not recorded" rather than "nothing raised", and five of fifteen had no public GitHub org at all.

L2. Onchain trading, perps, prediction and market structure #

Funnel. 1,866 unique URNs from four similarity runs and 86 keyword queries, all 1,866 fetched with zero errors. 38 dropped for no domain, 85 over the $15M cap, 4 publicly listed, 12 over 250 headcount, leaving 1,726 past hard filters, 661 genuinely in lane, 414 shortlisted, around 120 hand-reviewed, 22 deep-verified, 15 written up.

Best two. Convallax (#8, fit 7.0), options on prediction-market outcomes sold as infrastructure rather than as an app, with Alliance already on the cap table. Neutral Trade (#19, fit 5.5), which has the best-measured traction of anything in the set at $24,757,006 of DefiLlama TVL, three real audits, and a risk engine that automatically pulled funds out of the Drift exploit before loss while the Drift-architecture vaults took roughly $3.67M.

Where Daxos is strong and where it is blind. This lane produced the starkest single finding in the whole exercise: one company out of 1,866 hit the exclusion file. The fund holds four positions here (Pear, Vest, Trendex, Wager), understands order flow and perp mechanics better than it understands anything else, and has barely searched the lane at all. The thesis already said this at asymmetry D, that all five logged calls in the lane happened in 2026 and none before, and that the fund is buying from relationships rather than from search. The exclusion file is not the problem: it is the file that holds vest.markets, trendex.vip, raze.finance, seer.trade and nemesis.trade, all correctly caught. The lane simply has not been worked. The matching blind spot is the fund's own stated preference, picks-and-shovels, which it does not apply here: of 661 relevant companies, the memo found exactly one selling the engine rather than the app.

Attention: more, and it is the widest unmet universe in the exercise. With the caveat that the pool is wide rather than exhaustive, because a short query like "matching engine" hits the fifty-result cap. Two lane-level warnings must travel with any work here. First, volume is unmeasurable: DefiLlama's derivatives endpoint is paywalled, so every traction figure in the lane is locked value rather than traded volume, and for a perp venue volume is the number that matters. Second, the category has a visible death pattern: Ranger Labs (the best team in the entire pool on paper, now a 404 with a suspended Vercel deployment), Silver Koi, Sour Finance, Sports Value Market and Propr are all dormant or gone with real backing behind them.

L3. Tokenization, RWA, vaults and yield infrastructure #

Funnel. 1,373 unique after dedupe from three similarity runs and 45 keyword queries. 78 dropped over $15M, 67 out of business or acquired or Series D-plus, 8 over 250 headcount, 10 with no resolvable domain, 6 on the exclusion file, leaving 1,204 survivors and 905 genuinely in lane, 25 verified, 15 reported. Query overlap was low (only 95 of 1,373 appeared in two or more queries), so hit count was a weak ranking signal and the ranking used a lane score plus a Daxos-shape score.

Best two. Vaults.fyi (#2, fit 7.5), the read-and-write layer underneath the whole vault market, which is the layer beneath the fund's own LAGOON position rather than a competitor to it. Philidor Labs (#18, fit 5.5), a working credit rating agency for DeFi vaults whose scores reconcile against DefiLlama to within rounding, filling the gap that the vault-curator market LAGOON operates in has no rating agency at all.

Where Daxos is strong and where it is blind. The fund's LAGOON position is the anchor and the lane keeps using it. Arkonix is filed by DefiLlama in the identical category as LAGOON at $159,652 of TVL against LAGOON's $158.9M, which is the memo's phrase for "the same position a thousand times earlier". The clearest blind spot is cheque size rather than taste: the lane's genuinely best assets are out of range. Anemoy earns fees on $1.57bn and has no use for $250k. USIG's bench absorbs the cheque without noticing. Birch Hill at $9M is crowded by exactly the funds that compete for this lane. So the lane rewards being early on infrastructure nobody rates yet, not competing at the institutional end. One instruction worth carrying: Vaults.fyi, Philidor and stablewatch cannibalise one another, so back at most one of the three.

Attention: selective rather than more. Three names are worth a call this week and the rest of the lane is priced past the entry point. The lane also carries the most specific regulatory warnings in the exercise: secondary liquidity in fund interests is broker-dealer and ATS territory, residential lien tokenization does not escape state mortgage licensing or foreclosure law, and tokenized equity as collateral imports the wrapper risk of whichever equity token is pledged. Anonymous operators running permissionless vaults is flagged as the precise structure that has produced losses elsewhere, which is a direct warning on Sierra Protocol and Philidor.

L4. Consumer crypto apps, launchpads and social-fi #

Funnel. The largest pool in the exercise: 5,368 unique URNs from five similarity runs and 120 keyword queries, all enriched via batchGet in about 90 API calls. Six exclusion hits, of which three are real lane candidates the fund has already met (Pumpcade, Dream, ATHL). Then 68 with no domain, 107 over $15M, 167 dead or exited or Series B-plus, 11 over 120 headcount, 421 service businesses, 1,219 with no crypto signal, 455 off-lane, and 2,441 removed by a basic real-company quality gate, roughly 45 percent of the pool. 473 survived, top 170 hand-reviewed, 15 written up, and only two reached the ranked fifty.

Best two. Melee Markets (#15, fit 5.5), a parimutuel market maker with a minimum-return floor, which the memo calls the pump.fun shape pointed at outcomes instead of tokens and therefore the clearest read on what Daxos already owns. Ratio (#48, fit 2.5), which ranks only because Alliance is on the cap table, and whose two headline numbers (10,000 users, $12M notional) appear nowhere public.

Where Daxos is strong and where it is blind. This is the lane where the fund is genuinely strong and the sourcing came back thin. Five positions, the relationships that produce them (Alliance sits beside Daxos on both pump.fun and worm.wtf), and a top-of-list that is either already priced (Noise at $7.1M with Paradigm, Melee at $3.5M), possibly a zombie, probably not venture-fundable (Hey has 29,354 GitHub stars and no revenue model; Damus is likely grant-funded), or unidentifiable, with six of the lane's fifteen having no resolved founder. Scores top out at 7.5 against 8.5 in L1, L2 and L3. One genuine finding: creator monetisation onchain is declared dead as a category, with every candidate dormant or writing about itself in the past tense.

Attention: less. The fund already knows this lane and the lane has less in it than the fund's own positions imply. Two integrity notes belong in the record. Smithii, the best bootstrapped profile in the entire sweep at 14 people, zero raised since 2022 and 121,800 monthly visits, was cut on an adverse sweep: Trustpilot reports of charging before verifying deployment and user allegations of private-key solicitation via wallet authentication links. And the memo refused to surface a token-creation tool with credible private-key-solicitation allegations in front of the partner, which is the right call.

L5. Agent and developer infrastructure #

Funnel. 1,740 raw records pooled from two similarity runs on Steel.dev and Textile.dev plus 44 keyword queries, 1,725 unique root domains, 1 dropped on the exclusion file, 1,514 after hard filters (62 over the $15M cap, 102 founded pre-2018, 30 at Series B-plus or exited, 14 acquired or dead), then 93 after a quality gate that required real company signal, and 15 written up. Nine of those reached the ranked fifty, the most of any lane.

Best two. BoxLite (#1, fit 8.0), the only IN in the exercise and the closest thing in the set to the shape Steel.dev was bought at. Microsandbox (#3, fit 7.0), the joint-highest company score at 8.25, where the project predates the company by fifteen months and the two founders are the repo's top two committers.

Where Daxos is strong and where it is blind. The one exclusion hit out of 1,725 was steel.dev, the portfolio company itself. The memo's sentence is the finding: "Daxos owns a browser-infrastructure position and has held, as far as the call record shows, zero conversations with anything above it, below it or beside it." The thesis independently called this the loudest asymmetry in the dataset, because the house ledger rates this family at the very top of everything it has ever scored (Sponge 8.0, ATXP 7.5, Maven 7.5, MCPay 6.5 for fit) and the fund took zero calls in the lane across nine months of 2026. The structural version is worse than the sourcing version: Steel.dev sits in a stack with engines below it (Lightpanda), isolation substrates beneath that (Microsandbox, Superserve, BoxLite, eunomia-bpf), training environments above (Foundry), memory beside it (Zep, Caura), governance (Guardion, Archodex) and a direct thesis-negator in Parse, which turns any website into an API and so attacks the layer the fund owns. Owning one position in a stack and talking to nobody else in it means carrying the risk of the whole stack while seeing one layer of it.

Attention: more than any other lane, and the fund should start here. Nine of the fifteen lane names are at zero recorded capital, which is the price Steel.dev was bought at. The constraint is sourcing geography rather than deal availability: Lightpanda is Paris, Caura is Tel Aviv, Microsandbox's founders are in Lagos and Toronto, eunomia is contested between San Francisco and London, all outside the New York and Miami circuit the fund sources through. One hard methodological warning travels with this lane: Harmonic keyword recall is a floor rather than a census. Twenty-six nameable companies in this lane, including Mem0, Letta, Langfuse, Modal, Firecrawl, Exa and Composio, never appeared in a 1,740-company pool built from 44 queries, and five of the final fifteen arrived only through a named-company pass. The real lane is larger than 1,725 domains.

L6. Consumer AI apps #

Funnel. 4,544 unique Harmonic companies from five similarity runs and 121 keyword queries, plus a parallel App Store sweep of 49 iTunes terms yielding 3,376 unique apps, because Harmonic's web-traffic field is a poor proxy for a mobile-first consumer app. 4,489 after the $15M filter, 622 founded 2023 or later, 609 at or under $5M, 336 apps released 2023-plus with at least 1,500 ratings, 15 written up, six in the ranked fifty. Two of the four given seeds were known bad resolutions: Haloo resolved to a Toronto trademark-AI company at $7.19M Series A, and Pinned Social to an Essex marketing consultancy with one head.

Best two. Studley AI (#22, fit 5.0), four Dalhousie undergraduates who took an AI study app past a million Android installs with live paid subscriptions on zero outside capital. Swerve (#26, fit 4.0), two Matic Robots perception engineers who reached 41,643 App Store ratings at 4.7 in ten and a half months on a $500k YC cheque.

Where Daxos is strong and where it is blind. Exactly one of 4,544 companies hit the exclusion file, but the memo is careful that this is a property of the file rather than proof of neglect, since the file is mostly non-consumer domains. The real finding is the opposite of a sourcing gap and it is uncomfortable. The exclusion file did fire on the App Store side, and it caught SellRaze: 40,849 iOS ratings at 4.81, YC F25, $500k across two rounds, twelve staff, and the direct upgrade to the Moe position the fund did buy and then rated 3.75 after diligence found LLM cost at 142 percent of revenue and a fabricated resume. Daxos has already spoken to SellRaze and did not buy it. As the memo puts it, the resale sub-lane is not empty, it just was not won by the company Daxos bought, and that is a conversion question rather than a sourcing question. Somebody should read the Fireflies record for why it did not convert.

Attention: less, and redirected from sourcing to conversion. The thesis already graded this the weakest lane in the book, four positions with no demonstrated edge and one completed autopsy, and the sweep agrees and sharpens it. The method's own limit is the lane's dominant risk: rating counts measure installs, not retention, so every company on this list could have the Moe problem and this method would not see it, and six of the lane's fifteen are inference-heavy chat or companion products where the negative-margin failure is live. Five of fifteen have no identifiable founder, and no cheque should move on any of them before a name exists.

L7. Space, defense and frontier hardware #

Funnel. 3,733 unique companies from similarity runs on all three portfolio positions plus 93 keyword queries, 3,730 after the exclusion file dropped three, 1,072 after the over-$15M, startup-only, live-website filters, 724 after a lane-relevance and vintage filter, around 200 read in full by hand, 41 pulled for founder and award verification, 15 written up, four in the ranked fifty.

Best two. Viridian Space (#6, fit 7.0), two Princeton propulsion PhDs with a granted patent and $5.34M of federally verified money against $2.5M of equity. Coastal Assembly (#21, fit 5.0), ten years of measured sand accretion in the Maldives out of the MIT Self-Assembly Lab, wrapped in a ten-month-old company.

Where Daxos is strong and where it is blind. This is the lane where the fund has done the most analytical work (roughly forty rated companies, the deepest sector coverage in its history) and converted almost none of it, and it is also the lane where the exclusion file failed. Only three of 3,733 domains matched, and two companies unambiguously already in the house set surfaced in the shortlist and had to be dropped by hand: BESXAR, a canonical 7.5 benchmark anchor, and Aethero, which sits on the BORO-SM list where all partners voted in. The file does hold sophia.space, cosmicshielding.com, kinddesigns.com, satlyt.ai, katomed.ai, wootznano.com, teroai.co, kragon.space, delivr.ai and auron.solutions, so it is partly loaded, and the memo's warning is the right one: the gap is partial rather than total, which makes it more dangerous, because a reader would reasonably assume the file is authoritative. Every other lane run against this file should assume more known names leaked through. On the investment question, the thesis asymmetry is price rather than novelty, and the lane delivers: Rocket Propulsion Systems has turned roughly $5.4M of federal money into hot-fired hardware with zero equity dilution, and NuWatts, which did not get a deep card, sells passive thermal management for orbital compute and therefore sits directly in the critical path of the fund's own Sophia Space position at $0 raised and three people.

Attention: unchanged in volume, redirected in kind. The lane is already well covered relative to L5, L6 and L8, so the marginal value is in pricing rather than discovery. Odyssey at $0 is useful mainly as a cheap comparable that lets the firm price its own Sophia thesis at $20.89M, and Orbital at a $5M pre-seed with sixteen investors including a16z is the valuation comparable for the same position. Two categories are structurally unsuitable for the cheque and should be skipped: microreactors, where NRC licensing routinely outlasts a seed fund's horizon, and anything needing eight or nine figures before first revenue. The lane's evidence quality is also the weakest in the exercise: no federal award was found for ten of fifteen, founder names failed to resolve for four of fifteen, and the adverse sweeps are vacuous by construction, since CourtListener on thirteen names and Hacker News on twelve returned nothing about any of them, which is absence of coverage rather than a clean bill of health.

L8. Neurotech, bio devices and vertical AI #

Funnel. 3,234 pooled unique companies from two similarity runs and 78 keyword queries plus eight any-of bundles. 65 dropped for no domain, zero dropped on the exclusion file, 153 over $15M, 38 public or acquired or dead or over 400 headcount, 2,978 survived to scoring, top 250 hand-read, 15 written up, eight in the ranked fifty.

Best two. Pantheon AI (#5, fit 7.0), a bootstrapped seven-person bench of one FCAS actuary and three ex-Dropbox engineers shipping commercial underwriting automation with a live two-way INSTANDA integration and no outside money in thirty months. Infera-Neuro (#14, fit 5.5), an EPFL silicon spinout decoding neural signals on the implant itself, where the professor who invented the chips is the Chief Scientific Officer and the CTO already sold a silicon company to Sony.

Where Daxos is strong and where it is blind. This is the cleanest blind spot in the exercise and the number says it: 3,234 companies pooled, 328 recorded calls since June 2023, and the intersection is the empty set. Not one company this lane surfaced has ever been spoken to. Set that against the thesis finding at asymmetry F, that three of the five highest-rated companies in house history are neurotech, all at 8.75 (NeuroBionics out of the MIT Anikeeva lab, Auricle out of Stanford Biodesign and YC, and Elastro out of UCSD, Harvard and Sandia), and the fund holds one neuro position at $3M. The fund's own rating engine has told it three times what the best company it has ever seen looks like, and it did not buy. What bio deal flow exists runs through the Cane Angels group at the University of Miami medical school, which is a seat at someone else's table rather than proprietary sourcing. One correction belongs here and it is awkward: Elastro, one of those three 8.75 ratings, was re-rated by this exercise to 5.33 company and 2.0 fit on $2,325,000 verified by Form D, "clinical stage" claimed with no trial registration, and the CEO's record showing a 2025-10-01 end date. The anchor should move.

Attention: more, on the strongest evidence of any lane. Geography is the friction rather than quality: Australia twice, Switzerland, the UK and Canada. Three lane-level hazards must travel with any work. SBIR.gov is unusable, since its API returned 403 for every query, so no SBIR sweep ran at all. NIH RePORTER worked against a control and returned zero awards for every hardware candidate checked, which is weak evidence rather than strong. And Harmonic returned an empty investors array and empty funding-round detail for every single company in the shortlist, including the funded ones, so for Elastro a $2.33M seed closed 2025-08-15 is visible but not who wrote it. That is a systematic gap in the dataset, and it means nobody can tell who is already on any cap table in this lane. Finally, flexible high-density arrays are the single most crowded sub-segment in neurotech (Precision, Paradromics, Neuralink) and spinal cord stimulation is a litigated category with active US product-liability suits as of September 2026, which taints it.

5.1 No outside capital, with traction somebody else measured #

Twenty-six of the fifty show $0 on the Harmonic record. That is not the same as bootstrapped, and the distinction matters enough to split the list in three. The first group is the real thing: no capital on the record, no contradicting signal, and a traction number produced by someone other than the company.

Below the ranked fifty, four more belong in this section on the evidence: Nomadtable (nomadtable.app, $0, 20,612 iOS ratings, a claimed 75,000 MAU and a first $100,000 month), Sevren (sevren.com.au, $0, university-sponsored through UNSW's Tyree Institute, a prototype at 234 channels per square centimetre detecting down to 8 microvolts and a second generation at 1,024 channels in 0.3 square centimetres), Columnar (columnarbio.com, $0, one person in Seattle, spinal cord stimulation from structural biology, which the L8 memo calls the maximum-Daxos shape and also the rawest thing it found), and Doefin (doefin.com, $0, four people in Mauritius building Bitcoin hashrate derivatives with a CEO out of Nomura, Merrill and JPMorgan).

5.2 Zero on the record, and probably not bootstrapped #

This group is the single most important data-quality warning in the report, because the house formula pays a full +1.00 for a $0 raise. Harmonic flags funding_attribute_null_status = EXISTS_BUT_UNDISCLOSED on Caura AI, Fairgate, TariffSense, USIG Digital, Zult and Glimpse, which means Harmonic itself believes money was raised and was not disclosed. Separately: Second shows $0 against a Bitcoin Magazine claim of $5.1M from an unnamed private investor, with both the single-backer structure and the missing name unanswered. Fairgate runs 19 staff on $0 recorded. Firehorse runs 23 staff on $0 recorded, which is simply unexplained. Infera-Neuro shows $0 as funding_total while carrying a GRANT round dated 2026-09-08. Coastal Assembly shows $0 while 021T Capital lists it in its portfolio and says it helped form the company. Satora, Ratio and Glimpse each carry a logged round with no amount attached, and Glimpse also names Concentric as an investor against $0, which is internally inconsistent. Daimon shows $0 on the Daimon record while the operating company, Prospinity Inc, carries $2.1M from Xfund and Slow Ventures, and the card correctly uses $2.1M. And Creed shows $500,000 on Harmonic against $4,700,000 actual, stated on the record by a co-founder, so Harmonic understates that one by roughly nine times.

Every one of those is a reason to treat the +1.00 EV bonus as provisional until a founder confirms the cap table on a call.

5.3 The genuinely unusual founder stories #

Each with the evidence that made it checkable, because an unusual story that cannot be verified is a flag rather than an asset.

Peer validation nobody can buy. Yingjun Wu, the sitting founder and CEO of RisingWave Labs, has fourteen commits in the BoxLite repository. A database-infrastructure CEO writing code into another startup's runtime is the rarest signal in the set.

Published at the top venue in the field, by a student. Yusheng Zheng of eunomia-bpf is a currently-enrolled second-year PhD student at UC Santa Cruz and the first author of "Extending Applications Safely and Efficiently" at USENIX OSDI 2025, pages 557 to 574, with Dan Williams of Virginia Tech and Andi Quinn of UCSC. OSDI is one of the top two operating-systems conferences in the world and the paper is verified at usenix.org. There is no company yet.

No computer science degree, running a YC infrastructure company. Stephen Akinyemi of Microsandbox read Mass Communication at the University of Lagos. He was a contributing engineer at Wasmer in 2018 and a senior Rust and WebAssembly engineer at Fission in 2022, and has ten years of continuous public artifacts under one handle, 110 repos and 832 followers since 2016. He is also the top committer in his own repo at 522 commits. His co-founder Tochukwu Nkemdilim was a Cloudflare systems engineer and came through Mozilla's "Increasing Rust's Reach" in 2018. Both are from Lagos; the CEO sits in Nigeria and the CTO in Toronto.

Sneaker bots at twelve, now the legal version. Alex Forman of Parse started writing sneaker checkout bots at twelve and dropped out of UC Berkeley Engineering. Parse is the generalised version of exactly that skill, turning any website into an API by reverse-engineering its network layer. His Hacker News front page, 105 points and 27 comments, came from a stranger submitting his site rather than from him.

Sold a hedge fund at twenty-five, confirmable in EDGAR. Rahul Rai of Convallax co-founded Gamma Point Capital in 2021 and sold it to BlockTower Capital for a reported $35M in June 2021, then ran BlockTower's market-neutral book. The acquisition is confirmable in SEC EDGAR rather than only in press: CIK 0001756971 carries the former names "Shyn Capital Fund (US) I, LP" and "Gamma Point Capital Fund I (US), LP", now files as "BlockTower Gamma Point Fund LP" with Matthew Goetz and Ari Paul as executive officers, and shows $22,346,871 sold on the Form D of 2021-08-02. The unusual part cuts both ways: his record shows no role at all between April 2023 and March 2026.

A founder who shipped in 1985 and IPO'd in 1993. Yanki Margalit of Caura AI founded Aladdin Knowledge Systems in 1985 on $10,000, took it to a NASDAQ IPO in 1993, exited to SafeNet in 2009, and chaired SpaceIL for the Google Lunar X Prize. He is now building agent memory, and he holds four concurrent current roles including a partnership at a VC fund, which is the flag attached to the story.

A B-2 crew chief and a SpaceX propulsion lead. At Rocket Propulsion Systems, Max Ismailov holds a Purdue propulsion PhD with four years at Blue Origin, three at SpaceX, propulsion lead at Stratolaunch and Director of Propulsion at Firefly. His co-founder Ankur Saxena was a USAF B-2 crew chief from 2011 to 2015 before Notre Dame and NYU Stern. Saxena stepped down as COO in November 2025, which leaves one senior person.

A company born in a tank in its founder's own MIT lab. Skylar Tibbits of Coastal Assembly founded the MIT Self-Assembly Lab in 2013, co-founded Rapid Liquid Print, and is editor in chief of the 3D Printing and Additive Manufacturing Journal. He has run Growing Islands in the Maldives with his co-founder Sarah Dole's company Invena since 2017. His CTO, Walter Zesk, holds an MIT PhD in Computation and Architecture. The flag is the same fact: Tibbits holds six concurrent current roles including a full MIT associate professorship.

Four undergraduates and a million installs. Studley AI is Emran Hassan, Yahia Aly, Samuel Kuti and Kayhan Talebzadeh, four Dalhousie University students in Halifax, who shipped a consumer app past a million Android installs with live paid subscriptions on zero outside capital, and pushed a binary update the same day the card was written.

A twenty-one-year-old dropout who pivoted an infoproduct business into a gym app. Mauro Botanes of Symmetry is 21 or 22, a university dropout who ships the product himself, and built it in a public forty-day challenge. The April 2024 Wayback capture of symmetry.club is a paid one-day personal-branding event in Madrid, which is what the company was before. The asset is a distribution machine: roughly 500 creator accounts producing about 80,000 videos at zero paid ad spend. Three founders aged 20 to 23 took it to 361,000 store ratings and 33 staff, and roughly half the revenue reportedly goes to the creators.

The people who wrote the atomic swap, on their third attempt at the same idea. Philipp Hoenisch of Satora holds a TU Wien PhD and was research lead at CoBloX and COMIT. He and Lucas Soriano Del Pino wrote comit-network/xmr-btc-swap, the reference Bitcoin-Monero atomic swap at 710 stars and 183 forks, which is the implementation the Monero community actually forked; Hoenisch is its ninth contributor at 81 commits and Soriano its eighth at 86. They then spent two years on 10101, which no longer operates. Satora is attempt three.

A Bitcoin protocol maintainer rather than a crypto founder. Steven Roose of Second has 180 commits to rust-bitcoin/rust-bitcoin, 8 to bitcoin/bips and 8 to bitcoin/bitcoin, a GitHub account since June 2011, 224 public repos and 495 followers, and authored hal at 216 stars.

The man who designed Rootstock. Sergio Demian Lerner of Fairgate designed Rootstock, published the Patoshi-pattern research on early Bitcoin mining, and has twelve published BitVMX papers. The L1 memo calls him one of maybe five people who can credibly build dispute-resolution systems on Bitcoin. His co-founder Jonatan Altszul founded Core Security, which shipped Core Impact. The flag is that Lerner still carries a concurrent current RootstockLabs role and his GitHub profile still reads "IOV Labs".

A surgeon attacking the complication he watches happen. Taiyo Kuroda of Helioverse is a practising cardiovascular surgeon with an MD and PhD from Toyama who spent 2021 to 2024 as a research fellow at the Cleveland Clinic Lerner Research Institute and authored first-report papers on the Cleveland Clinic continuous-flow total artificial heart. He is building the driveline-free implantable pump to remove the driveline infections he sees. He has no engineers.

Philosophy and literature credited for the method. Hyonmyong "Hoch" Cho of Firehorse is a CIO at Prime Data Centers and co-founded Forum Partners past $2.5B AUM, and his claims are checkable in his own filings: he filed SC 13G on Mitek Systems in 2011 and 2012, and was elected a Destiny Media Technologies director on 2026-02-27. The L8 memo notes the patent-pending "Data Lathe" method is credited to a literature and philosophy background, which is at least an honest origin story. The method is not findable in any public patent database, and the site says he "co-founded Forum Partners" where his own LinkedIn says Managing Director, Head of Execution.

A SpaceX hardware engineer selling against the incumbent using the incumbent's own software. Simon Hakimian of Cerelog is an ex-SpaceX Hardware Development Engineer and a Purdue EE, working solo, who ships a $390 ADS1299 EEG board against OpenBCI's $1,249 Cyton. The board ships with a custom fork of the OpenBCI GUI plus BrainFlow, so it drops into existing workflows through Lab Streaming Layer. That is the distribution channel, not the product. He wrote 258 of 258 commits and has merged zero outside pull requests.

A career VC with no customs experience, sitting on the busiest border crossing. Tina Safaei of TariffSense has an EPFL and MIT materials science background, a University of Toronto PhD in Electrical and Computer Engineering, and five years in venture capital at Plaza Ventures with board observer and board seats, so she has sat on both sides of a term sheet. She has zero customs, trade compliance, brokerage or logistics experience, and she is attacking a market whose whole difficulty is accumulated regulatory domain knowledge. She operates from Windsor, Ontario, the busiest US-Canada crossing, where the brokers actually work. Her Harmonic primary email is still tina@plaza.ventures.

A Ledger engineer and a vault curator, alone in Tbilisi. Guilhem Pequillat of AlphaShell came through École Normale Supérieure Paris-Saclay, did flight-mechanics numerical simulation at ArianeGroup, spent 2021 to 2024 as a Software Engineer at Ledger, and 2024 to 2026 as Lead Engineer at MEV Capital, a live DeFi vault curator, leaving to build the execution layer for the trade he had been curating. Key handling plus basis-trade operations is exactly the two-part stack his product needs. He is one person, his LinkedIn says "Co-founder" with no co-founder identifiable anywhere, and his own binding Terms of Service name the entity "under identification code [ID code], with its registered address at [registered address]", both placeholders unfilled in production.

A Diem core developer and a ten-year partnership. Philip Hayes of Lexe spent 2018 to 2021 at Facebook as a Diem and Libra Core developer, which is production Rust consensus engineering. He and Max Fang co-founded Blockchain at Berkeley together in 2015, so this is a ten-year working partnership.

A professor who can actually make the thing. At Sevren (below the cut), the founders are Scientia Professor Nigel Lovell, who heads UNSW's School of Biomedical Engineering, directs the Tyree Institute of Health Engineering and is a past President of IEEE EMBS, and Professor Francois Ladouceur. Josiah Firth on the team is a process engineer at ANFF NSW, the national fabrication facility, which is what makes "we can fabricate this" a verifiable claim rather than a hope.

And three that are unusual in the wrong direction, recorded so they are not lost. At Melagen Labs (below the cut, and otherwise the closest second Cosmic Shielding in the pool at $620,000), the CEO Muhammad Hunain has no work history before 2023 and education running through Flushing High School and LaGuardia Community College, while the CTO Dr James Secrest holds a UCLA PhD and a Doctor of Chiropractic and is concurrently Operations Manager at Lustra USA. At NuWatts, whose insight about radiator mass in vacuum is excellent, founder James Ross holds a Liverpool PhD but his career is community-college chemistry instruction from 2009, with no aerospace thermal engineering role. And at Parallax Defense, Yelda Bartlett is an attorney and former political campaign manager whose company was formed the month her Berkeley Hacking for Defense team "Triton Shield" ended.

Gap 1. The fund rates agent infrastructure highest, owns its best position there, and has stopped looking #

The evidence. The house ledger rates this family at the top of everything it has ever scored: Sponge 8.0, ATXP 7.5, Maven 7.5, MCPay 5.5 company but 6.5 for fit, four of the ten highest fit scores in house history. The fund holds Steel.dev, which the thesis calls arguably the best-shaped position in the whole book: founded 2024, $0 raised at entry, Daxos named alongside Eterna, and Eniac leading a $4M seed afterwards. And the fund took zero calls in the lane across the nine months of 2026, against four in 2025. The sourcing run confirms it from the other side: of 1,725 pooled domains in the lane, exactly one hit the 303-entry exclusion file, and that one was steel.dev itself.

Why it is worse than a sourcing gap. Steel.dev sits in a stack. There is an engine layer below it, isolation substrates below that, training environments above, memory beside it, governance beside that, and at least one company whose product undermines the layer the fund owns. Holding one position in a stack while talking to nobody else in it means carrying the risk of the whole stack and seeing one layer of it. Parse is the sharpest version: it turns any website into an API by reverse-engineering the network layer, which is the bear case on browser automation, and the fund would learn more from that call than from any other in this report.

The named companies that would close it. BoxLite (#1, fit 8.0, the only IN) is the Steel.dev entry shape repeated: two people, $0, a working runtime with 80,367 monthly PyPI downloads. eunomia-bpf (#7) is the observability and policy layer underneath, published at OSDI, with no company yet, which is the earliest the fund could possibly be. Microsandbox (#3) and Superserve (#20) are the isolation substrates, one with 741,389 npm downloads a month and the other with the only live priced self-serve product in the set. Lightpanda (#4) is the engine layer, and the only team in the lane that owns an engine rather than renting Chromium. Cua (#10) is the adjacent primitive at the same stage in the same city. Caura AI (#16) is memory, with 300 agents in production at a Nasdaq-listed fintech. Zep AI (#38) is memory with the adoption and none of the distribution. And Parse (#24) is the competitive-intelligence call.

Gap 2. The fund's rating engine has told it three times what the best company it has ever seen looks like, and it has never bought one #

The evidence. Three of the five highest-rated companies in house history are neurotech, all at 8.75: NeuroBionics out of the MIT Anikeeva lab, Auricle out of Stanford Biodesign and YC, and Elastro out of UCSD, Harvard and Sandia. The book holds one neuro position, Theta Neurotech at $3M and three staff, which is the exact shape. The 2026 bio and medtech call traffic, eight calls, runs almost entirely through the Cane Angels group at the University of Miami medical school, which is a seat at someone else's table rather than proprietary sourcing. And the sourcing run produced the cleanest finding in the exercise: 3,234 companies pooled in the lane, 328 recorded calls since June 2023, and the intersection is the empty set. Not one company this lane surfaced has ever been spoken to.

One correction the fund should absorb first. Elastro, one of those three 8.75 anchors, was re-rated by this exercise to 5.33 company and 2.0 fit, on $2,325,000 verified by SEC Form D accession 0002076219-25-000001, "clinical stage" asserted with no trial registration found, and the CEO's own record showing a 2025-10-01 end date. If the ledger keeps 8.75 for Elastro, the anchor distorts every neurotech rating made against it.

The named companies that would close it. Infera-Neuro (#14, fit 5.5) is the strongest: an EPFL silicon spinout with the inventing professor as CSO, a CTO who already sold a silicon company to Sony, and genuinely silicon-proven peer-reviewed results, a 256-channel closed-loop neuromodulation SoC at 453 microwatts in IEEE JSSC 2022 and a 512-channel record-and-decode die doing 31-class handwriting decode at sub-900 microwatts in JSSC 2024. SymKardia (#23) is a four-person company that is the registered lead sponsor of an FDA-regulated human trial, NCT07501819, with five verified Boston Scientific and LivaNova cardiac-device veterans and sixty-one LinkedIn followers, which the L8 memo correctly calls a pricing error. Cerelog (#28) is the only company in the lane with a number anyone can count independently. Sevren and Columnar, both below the ranked cut, are the pre-institutional lab-spinout and solo-founder shapes the book has shown it will write. Carry two category warnings: flexible high-density arrays are the most crowded sub-segment in neurotech, and spinal cord stimulation is a litigated category with active US product-liability suits as of September 2026.

Gap 3. The fund is a customer of private-market plumbing and has only ever met the consumer end of fintech #

The evidence. Nine non-crypto fintech counterparties, fourteen calls, accelerating from one in 2024 to four in 2025 to nine in 2026, and zero positions. Every one of those nine was a consumer money app: 0DTE, Cero, Flip Labs, Kem, Money&Co., Parrot Finance, SOAR, SecondLane and UNest. Two were later rated and killed, Cero at 2.0 with "do not meet" and Kem at 6.0 on a fatal 3.48 percent DAU-over-MAU. That is the sub-lane with the worst outcome record in the entire ledger, and the fund keeps meeting it. Meanwhile the fund runs its own operations on NAV Fund Services, Houlihan Capital, Setter Capital, Marex and FalconX, so it has lived the pain of fund-operations and private-markets infrastructure, and it met SecondLane exactly once.

The named companies that would close it. Vaults.fyi (#2, fit 7.5) is the clearest: the read-and-write API layer underneath the vault market, which is the layer beneath the fund's own LAGOON position rather than a competitor to it, and the fund's holdings in Xverse, Botanix, pump.fun and Trendex are warm introductions to exactly the wallets and exchanges it sells to. Philidor Labs (#18) fills the specific hole that the vault-curator market LAGOON operates in has no rating agency; its output reconciles against DefiLlama to within rounding, and the issuer-pays revenue model is the single thing to argue about. USIG Digital (#39) has the heaviest wholesale-collateral bench in the set, including the man who built OUSG at Ondo and a Deutsche Bank Global Head of Strategic Liquidity, and met the SEC Crypto Task Force on 2026-01-12. Pantheon AI (#5, fit 7.0) and Firehorse (#31) are the same gap one layer out into insurance and loan-servicing plumbing: a named form checked against a named rulebook, which is the compliance-obligation wedge the house rates well and has never sourced. Fission (#35) and Anemoy (#37) are the institutional end, and both are instructive about why this gap stays open: Anemoy is the best company in the set and is probably not purchasable at all.

Gap 4. The deepest analytical work in house history, almost no conversion, and the bar is price rather than novelty #

The evidence. Space and defense is the ledger's deepest sector at roughly forty rated companies, with Thermopylae Aerospace and Aethero on the BORO-SM hot list where all partners voted in. The fund holds three positions. All five space calls happened in 2026, none before, and none got a second meeting. Two of the three positions, Kind Designs at $26.5M and Sophia Space at $20.89M, would fail the fund's own $15M hard filter today, so the lane the fund likes most is the lane it can no longer afford to enter at the point it likes. The thesis instruction is specific: bring sub-$5M component and subsystem suppliers on the Cosmic Shielding template, where the customer is a prime rather than a consumer.

The named companies that would close it. Viridian Space (#6, fit 7.0) is that company with better founders and a cleaner record than the one the fund already owns: two Princeton propulsion PhDs out of a 22-year combined run at The Aerospace Corporation's plasma lab, a granted patent on the core component, and $5,336,046 of federal money verified line by line against $2.5M of equity, a 2.1x non-dilutive-to-dilutive ratio. Rocket Propulsion Systems (#27) is the extreme version: roughly $5.48M of federal money turned into hot-fired methalox hardware and a 2027 Space Force delivery with zero equity dilution, though it is still a three-person Washington LLC after 7.6 years with no sellable instrument. Coastal Assembly (#21) is the Kind Designs thesis restated by a ten-month-old company with ten years of measured field work behind it. Melagen Labs, below the cut, is the closest thing in the pool to a second Cosmic Shielding at $620,000 against Cosmic's $4.61M and in New York, and the team evidence is the reason it is below the cut rather than above it. And NuWatts, which did not get a deep card, sells passive thermal management for orbital compute at $0 raised with three people, which puts it directly in the critical path of the fund's own Sophia Space position: in vacuum there is no convection, radiator mass scales with heat load, and the power spent moving heat becomes more heat to reject. Owning a small position in your own portfolio's physical bottleneck is cheap insurance.

A fifth, shorter gap: infrastructure rather than apps in prediction markets #

The fund holds Worm, Trendex and Wager, and it has met seven prediction counterparties, every one of them a consumer app (Arena, Converge, Dare, Pumpcade, SOPO, Totalis, Trepa). The ledger rates Bayse Markets, a prediction exchange with real pedigree, at 8.0, so the house already knows infrastructure beats apps, which is its own stated T8 preference. Of 661 genuinely in-lane companies in L2, the memo found exactly one selling the engine rather than the app. Convallax (#8, fit 7.0) is the named company: options on Polymarket outcomes sold as an RFQ venue with a published audit, two people in New York, Alliance already on the cap table, and nothing settled with real money yet. Doefin, below the cut, is the same shape pointed at Bitcoin hashrate with a Nomura and Merrill derivatives career behind it.

The sourcing method #

Eight lanes were run in two waves against the Harmonic API, each seeded from the fund's own positions in that lane (Botanix, Xverse, Ord.io, Bound and Saturn for L1; Pear, Vest, Trendex and Wager for L2; Raze and LAGOON for L3; pump.fun, Aura.fun, worm.wtf, Exverse and Bubblemaps for L4; Steel.dev and Textile.dev for L5; Moe, Haloo, Pinned Social and Bo AI for L6; Sophia Space, Cosmic Shielding and Kind Designs for L7; Theta Neurotech and Hirechain for L8). Each lane ran similar_companies on its seeds plus between 44 and 121 keyword queries, pooled and deduplicated on root domain, applied the hard filters, scored the survivors on lane relevance plus Daxos shape, hand-read the top of the ranked list, and wrote fifteen names up. That produced 119 pooled unique shortlist names, of which 64 received a deep card built from Harmonic plus an independent verification pass (GitHub and npm and PyPI and crates.io APIs, SEC EDGAR full-text and Form D, USAspending and NSF award APIs, DefiLlama, App Store and Play lookups, ClinicalTrials.gov, Companies House, state registries, Wayback, Hacker News Algolia, CourtListener, and direct fetches of the companies' own rendered bundles). The fifty ranked here are those 64 minus the fourteen lowest on fit.

Pool sizes by lane: L1 2,462 unique from 3,327 raw rows; L2 1,866; L3 1,373; L4 5,368; L5 1,725 from 1,740 raw; L6 4,544 Harmonic companies plus a separate App Store sweep of 3,376 unique apps; L7 3,733; L8 3,234. Roughly 24,300 unique company records in total, with unknown cross-lane overlap, so that number is a ceiling on breadth rather than a count of distinct companies.

The exclusion universe, and the one place it failed #

Every pooled domain was checked against a 303-entry exclusion file holding the 198 external domains that appear across 328 Fireflies call transcripts from 2023-06-29 to 2026-10-01, the house benchmark and rated set from reference_daxos_benchmark_distribution.md, and all 28 resolved portfolio domains. Matching ran on root domain, each parent domain suffix, and bare company name over four characters, which is why botanixlabs.com caught botanixlabs.xyz. Exclusion hits by lane: L1 10, L2 1, L3 6, L4 6, L5 1, L6 1, L7 3, L8 0.

The file is not fully authoritative and a reader should know exactly how. In L7, two companies that are unambiguously already in the house set surfaced in the shortlist and had to be removed by hand: BESXAR, a canonical 7.5 benchmark anchor, and Aethero, which sits on the BORO-SM list where all partners voted in. The file does contain sophia.space, cosmicshielding.com, kinddesigns.com, satlyt.ai, katomed.ai, wootznano.com, teroai.co, kragon.space, delivr.ai and auron.solutions, so benchmark companies were partly loaded. The gap is partial rather than total, which makes it more dangerous than a wholly empty file would be, because it invites trust. Separately in L8, ABILITY Neurotech was excluded by hand as prior house work that the file did not catch. Treat every "zero exclusions" finding in this report as "zero matches against an imperfect file", not as proof of a virgin lane. That said, the L2, L5 and L8 findings survive the caveat comfortably, because one hit in 1,866, one hit in 1,725 that was the portfolio company itself, and zero hits in 3,234 are not the kind of numbers a few leaked names would change.

The Harmonic-only funding rule, and where Harmonic was wrong #

Every funding, valuation, stage, round-count, investor and headcount figure in this report comes from the Harmonic API, pulled 2026-10-04. No web funding number was used as a primary source anywhere. That rule exists because web figures are stale and wrong, and it held up: it is the rule that produced the correct $2.1M for Daimon through the Prospinity entity rather than the $0 on the Daimon record.

The rule also has a cost, and it should be stated plainly. Harmonic was independently contradicted in at least seven places in this set. Creed shows $500,000 against $4,700,000 actual, understating by roughly nine times. Second shows $0 against an unconfirmed $5.1M press figure. Lygos shows a $1.1M equity seed with no corresponding Form D, while the only third-party-attested capital figure is $190,000 from one investor. Parse shows $500,000 against YC-derived coverage of $125,000. Elastro was verified to $2,325,000 and not the $2,330,000 carried. Six companies are flagged by Harmonic itself as EXISTS_BUT_UNDISCLOSED, meaning Harmonic believes money was raised that it cannot see. And in L8, Harmonic returned an empty investors array and empty funding-round detail for every single company in the shortlist, including the funded ones, so nobody can tell who is already on a cap table in that lane. In L1, funding_rounds came back empty for every top-15 company. The safest reading: $0 in this report means "not recorded" and the +1.00 EV bonus that rides on it is provisional until a founder confirms it on a call.

Harmonic also mis-resolves entities, which is a data-quality item for whoever maintains the portfolio file. The thesis already documented six wrong or doubtful resolutions among 28 positions, including Hopper resolving to a Montreal travel app at $710.5M rather than Hopper Labs and LAGOON at $0, and Pinned Social resolving to an Essex marketing consultancy. The L6 sweep hit the same class of error live: calai.app resolved to MyFitnessPal founded 2010, and two of its four given seeds were known bad resolutions. Harmonic stage fields are not trustworthy on crypto companies, since Botanix carries OUT_OF_BUSINESS against a live mainnet. Harmonic headcount is LinkedIn-derived and systematically inflated on small companies: Studley shows 19 because it counts a user and UGC creators, Symmetry shows 33 on $0, Turbo AI shows 17 against a recorded peak of 445. And Harmonic's web_traffic was null on every record in the L3 pool and is a poor proxy for a mobile-first consumer app, which is why L6 had to run a parallel App Store methodology.

The $15M hard filter #

Applied before scoring in every lane, per house rule. It removed 85 companies in L2, 107 in L4, 78 in L3, 153 in L8, 62 in L5 and the bulk of L7's reduction from 3,730 to 1,072. Nothing in the ranked fifty is above it; the largest raise in the set is Creed at $4.70M. Three named companies were dropped on it that are worth knowing about anyway, because they are the comparables the fund will price its own positions against: Orbital at a $5M pre-seed with sixteen investors including a16z, which is the most direct Sophia Space competitor found; Tolan at $30M and 183,130 App Store ratings, which is what the companion sub-lane is being priced against; and Noise at $7.1M with Paradigm, which the L4 memo calls the most credible attention-market instance in the entire pool.

What could not be verified, across the set #

Revenue. No verified revenue figure exists for any company in the ranked fifty. Not one. Where revenue is mentioned it is founder-reported, press-reported, or inferred from a pricing page.

Retention. No cohort data, DAU-over-MAU, or repeat-usage rate exists publicly for any consumer company here. The L6 memo says this plainly: rating counts measure installs rather than retention, so every company in that lane could carry the Moe failure (LLM cost at 142 percent of revenue) and this method would not see it. Six of the fifteen in that lane are inference-heavy chat or companion products where that failure is live.

Customers. Roughly three-quarters of the fifty have no customer confirmed from the customer's side. Vaults.fyi's case studies all trace back to vaults.fyi's own documentation. Microsandbox's "teams building with microsandbox" is adoption language. Zep's customers page returns a 404. Pantheon names nobody at all.

Trading volume. DefiLlama's derivatives endpoint is paywalled, so every traction number drawn from DefiLlama in this report is locked value rather than traded volume. For a perp venue, volume is the number that matters and it is not available.

Adverse checks. CourtListener returned a 22-to-24-hour anonymous-tier throttle on at least six companies. California FTB suspension status was WAF-blocked on every path. SBIR.gov returned HTTP 403 for every query, so no SBIR or STTR sweep ran in L8 at all. Reddit returned 403. No litigation or insolvency search was attempted in Georgia, France, Argentina, Switzerland, Estonia or the BVI, which covers six companies in the fifty. In L7, CourtListener on thirteen names and Hacker News on twelve returned nothing about any of them, and the memo is right that this is absence of coverage rather than a clean bill of health.

Federal awards. USAspending indexes prime contracts and grants only, missing subcontracts, classified awards, sub-threshold OTAs and state money, and name collisions degrade the search. No federal award was found for ten of the fifteen L7 companies, and "not found" is not "does not exist". NIH RePORTER was verified against a control and returned zero awards for every L8 hardware candidate, which is weak evidence rather than strong, because an organisation registered under a different name from its trading name would be missed.

Where a reader should be sceptical #

Be sceptical of any score above 7.0 in this set. The house formula pays +1.00 for a sub-$1M raise and +1.00 for a US address, so two of the points in most high scorers are bonuses rather than judgment. Check the base. BoxLite's base is 6.00, Microsandbox's 6.25, Cua's 6.25 and Lightpanda's 6.67, which are defended. But eunomia-bpf's base is 5.08, Convallax's is 5.00 and Viridian's is 5.50, so each of those is a 7-plus built substantially on bonuses, and the thesis names this artifact explicitly and lists five prior instances of it, the worst being Neuramorphic at 4.00 fit against 1.50 general, annotated "vaporware".

Be sceptical of the recall, not just the precision. The most important methodological finding in the exercise is the L5 one: twenty-six nameable companies in agent infrastructure, including Mem0, Letta, Langfuse, Modal, Firecrawl, Exa and Composio, never appeared in a 1,740-company pool built from 44 keyword queries, and five of the final fifteen in that lane arrived only through a deliberate named-company pass. Harmonic's keyword index does not reliably retrieve a company from a phrase describing what it does, and its matcher is phrase-literal, so "defense autonomy software" and "SBIR phase II hardware" both returned zero. This report is a floor on what is out there, not a census. Anything not surfaced by these queries is simply absent, and the real lanes are larger than the pool sizes above.

Be sceptical of the $0 column. Twenty-six of the fifty show it and section 5.2 lists the dozen where it is probably wrong.

Be sceptical of the one IN. BoxLite is a single-founder company with 42 contributors and roughly six sustained committers at "headcount 2, nothing raised", which is unexplained, plus an investor named against $0 and 0 rounds, plus a contested HQ between Menlo Park and Singapore. The verdict is IN conditional on three answers, and the first of them, what funds six committers at zero raised, drops it to WATCH if it reveals undisclosed capital or an undisclosed parent.

Be sceptical of the ranking at the margins. Ranking on fit and breaking ties on company score is the right rule for a partner who wants calls to make, but it buries good companies. Anemoy at 7.75 company ranks 37th. Swerve at 7.50 ranks 26th. Superserve at 7.33 ranks 20th. Fairgate at 7.33 ranks 12th. If the question were "which of these is the best business" rather than "which should Daxos call", the order would be substantially different, and the COMPANY column is there so a reader can re-sort it.

Finally, be sceptical of the lane boundaries. Two companies in the ranked fifty sit in lanes they do not really belong to, because the lane that surfaced them owns them: Pantheon AI and TariffSense are filed under L8 "Neurotech, bio devices and vertical AI" and are in substance vertical-AI compliance software with nothing bio about them. Firehorse is the same. That is a labelling artifact of how the sourcing was split, not a judgment about the companies.