Project summary

The conversation-analysis market measures sentiment, topics, and talk-time, but ships no measure of depth — whether an exchange actually moved something or just performed. Over seven years I recorded and transcribed my own conversations across eleven contexts (work, legal, medical, family, facilitated introspection): 50,000+ segments where the producer of the record and the judge of its depth are the same person. On that corpus I validated a five-pattern, five-level depth taxonomy, and the signal turns out to be physical — the deepest segments run slower and more silent (167 vs 200 words per minute; 1.22s vs 0.79s pauses; two pre-registered results, p≈4.5×10⁻⁶ and p≈0.003). This project turns the instrument into a public benchmark: can frontier models tell genuine reasoning from fluent performance?

What are this project's goals? How will you achieve them?

Three deliverables. First, a human-adjudicated, held-out gold-standard label set built under a pre-registered protocol — the current pipeline labels are AI-applied and quality-gated, and the gold set is what makes every claim testable. Second, published per-model results for frontier models (Claude, GPT, Gemini, open-weights) on the depth-recognition task, including a null result if none beat chance. Third, the taxonomy specification and evaluation harness released openly so others can replicate on their own data. The relevance to safety work is direct: depth discrimination is the same muscle as deception and sycophancy evaluation — telling whether communication is performative.

How will this funding be used?

$5,000 (minimum) funds adjudication tooling, the first gold-set slice, and the harness release. $25,000 funds the full six-month open release: my time writing the specification, a substantially larger adjudicated set, and multi-model evaluation runs beyond the roughly $4,800 in committed compute credits I already hold.

Who is on your team? What's your track record on similar projects?

Solo. Before this: initial team at R3 building the Corda blockchain platform (I worked on its first commercial deployment, into the $2.4T international gold market), Head of Financial Services at TradeLens, the Maersk–IBM venture, where the Citi Asia-Pacific implementation enabled $1.2B in paperless trade finance, and ADP Ventures (consumer-permissioned data platform scaled to a $25M projected run rate). Since 2023, alone: 70+ self-running edge services, a public medical-research portal (revasserkernel.com, 10,000+ papers indexed for rare-disease literature synthesis), live commercial products (revasser.nyc, revaddress.com, developer tooling on npm). Seven years collecting the corpus, two years building the instrument.

What are the most likely causes and outcomes if this project fails?

Three honest risks. The corpus is single-subject — that is precisely what eliminates inter-rater disagreement, but it means transfer beyond one identity is an open question; the benchmark is designed so others can test the taxonomy on their own data, which is the only real answer. The current labels are AI-applied and not yet human-validated — the gold set this funds is the fix, and if adjudication shows the taxonomy discriminates worse than claimed, that result gets published too. And I am one person running real businesses alongside this; the deliverables are scoped so each lands independently.

How much money have you raised in the last 12 months, and from where?

$0 external to date — everything so far is self-funded from my own products. Pending applications this month, disclosed in full: Emergent Ventures ($25,000, applied 2026-09-06) and the EA Transformative AI Fund ($75,000 over 12 months, applied 2026-09-06 — that application funds my time for the broader research program; this Manifund ask funds the open-release slice, and if both land I will take the union of deliverables and reduce or return the overlap). Also pending: Anthropic AI for Science (API credits) and Google for Startups (~$2,000 infrastructure credits).