Case 23 · Graduation Day
NDAI
A genuinely felt problem — Arrow's information paradox told through a deal that dies the moment it goes public — carried by one clean mechanism sentence, then surrendered by an over-time apology, a half-remembered author list, and nothing for the room to do next. Scored as if pitched to investors; it is a research briefing, so a 1 on the ask is normal.
01 Scorecard · 12 dimensions
02 Klaff 5-lens audit · 2/5
THE ONE REORDER — Open on the a16z story cold, name Arrow's paradox as the villain, and end on one line the room can act on — where to read NDAI — instead of the over-time apology.
03 Delivery
78 fillers ("um" ×23, "like" ×23, "uh" ×20) — nearly one word in eleven. The mechanism section survives it; the open and close don't. Ran over its slot, acknowledged on stage.
04 Beat-by-beat
| Time | Beat | What works | The fix |
|---|---|---|---|
| 0:00 | Intro recovery | Honest — "I'm not gonna lie, it's not exactly what I did" | Cut entirely. Reacting to the emcee's framing spends the pattern-interrupt window on meta-commentary. |
| 0:20 | Brandon callback | "Information markets we exist in are very inefficient" ties the talk to the day's arc | Keep to one clause; it currently delays the story. |
| 0:50 | MetaDAO / a16z story | The best 60 seconds of the talk: a deal dies because evaluation had to happen in public | Don't introduce it with "Not really important" — that undercuts the story right before it does all the work. |
| 1:50 | Generalization + soda factory | "You kind of give up the most valuable part of an interaction" — the thesis in one line | The soda example is redundant after MetaDAO; one concrete beats two. |
| 2:30 | Arrow's paradox + NDAs | Naming the economics and the incumbent (NDAs: expensive, slow, imperfect) gives the room a handle | Rehearse the garbled line — "we don't promise to share information… we promise not to share information about it." |
| 3:00 | NDAI mechanism | "There's only one bit of information that leaves it, like yes, deal or zero, no deal" — the sentence the room will remember | Add the missing proof beat: what does the paper actually show? [confirm: a result, construction, or theorem] |
| 3:45 | Programmable privacy | Challenges the privacy-vs-experience trade-off from yesterday's discussion — a genuine idea | This is the insight; it deserves the close, not the slot after the mechanism. |
| 4:20 | Close | — | "I'm way over time… I guess it's good enough" is the anti-close. Replace with one pointer + one sentence of why it matters. |
| 4:35 | Conditional Recall (prompted) | "If you can define conditions and commit to conditions under which something is remembered or not, what does that enable?" | It only surfaced when prompted ("another paper you forgot to mention"). Integrate it in one planned sentence or drop it. |
05 The gaps investors block on
- The speaker's relationship to the work is never stated — presenter, co-author, or fan? The room cannot weight the claims.
- Zero results from either paper — the TEE-negotiation mechanism is described but nothing is shown to work.
- The one-bit claim invites the obvious objection — real deals need price and terms, not just yes/no — and never pre-empts it.
- The best story is self-undercut: "Not really important" said about MetaDAO, the example that carries the talk.
- No pointer — no link, venue, or next step for either paper. [confirm: where NDAI and Conditional Recall are published]
Pick the deal-room wedge. The talk floats between two papers and three audiences — DAO investors, M&A counterparties, privacy theorists. If NDAI is ever pitched as more than a briefing, the fork is programmable-privacy-as-paradigm vs. the deal-room wedge: NDAI as the NDA replacement for deal evaluation under Arrow's paradox. Take the wedge — it is the only place the talk already has a felt, specific situation and a named, hated incumbent.
“NDAI is the NDA replaced by hardware: both sides send agents with their secrets into a sealed enclave, and the only thing that ever leaves is one bit — deal, or no deal.”
06 The rewritten pitch
Andreessen Horowitz wants to buy ten million dollars of a token at a twenty percent discount. But the deal has to be evaluated by a DAO — in public. So the moment the offer is visible, everyone pumps the token, and the deal is dead. That's not a crypto quirk. It's a 50-year-old problem with a name: Arrow's information paradox — your counterparty has to evaluate your information before the deal, but revealing it gives away the most valuable part of the interaction. A factory with a secret soda recipe can't be acquired without handing the recipe to the acquirer — at which point, why acquire?
Today we patch this with NDAs: expensive, slow, and imperfect. NDAI — a paper by [confirm full author list] — replaces the NDA with hardware. Each side sends an agent, loaded with its private information, into a trusted execution environment. The agents negotiate inside. The enclave is built so that exactly one bit ever leaves: one, deal; zero, no deal. No leakage, no lawyers. And before you object that real deals need more than one bit — price, terms — [confirm: the paper's answer to richer outputs].
The bigger idea is what the authors call programmable privacy. Yesterday's discussion treated privacy as a trade-off — keep your data or trade it for a better experience. These systems break that trade-off: you commit, in code, to exactly how your information can be used, and markets that were impossible become efficient. A companion paper, Conditional Recall, by Shin and Christoph Schlegel from Flashbots, asks the mirrored question: what becomes possible if you can commit to the conditions under which something is remembered?
Both papers are at [confirm link/venue]. If you hold information you can't reveal but need evaluated — a deal, a recipe, a position — read NDAI. It's the NDA that's math instead of lawyers.
Brackets are the founder's homework — the copy is done, the proof is theirs to supply. Every claim must match what's actually built.
07 Do these before the next take
- State your relationship to the work in the first sentence — the briefing's weight depends on it.
- Move the a16z story to 0:00 and cut the emcee meta-commentary; never introduce the load-bearing example with "Not really important."
- Get the author list right — "probably someone else who I forgot to mention" is the cheapest fix in the talk. [confirm NDAI authors]
- Add one proof beat per paper — what NDAI shows, what Conditional Recall shows, one sentence each.
- Pre-empt the one-bit objection (price/terms) before the room raises it.
- End with a pointer, not an apology — and time the talk to the slot; it ran 5:07 with an on-stage "I'm way over time."
- Filler drill — 8.8/100 ("um" ×23, "like" ×23, "uh" ×20) is the highest-friction number in this review.