The text behind each main slide, kept off the slides.
Each main slide now carries one short body. What it used to say is here, keyed by slide, unchanged in substance.
02 What if: the other four reasons researchers gave
They said: It makes things up, on exactly the technical details that matter. What if every answer were grounded in your own verified record, and cited it, claim by claim?
They said: It reasons impressively about everything, and shallowly about the field I actually work in. What if it kept its frontier-wide reach, and knew your lab’s work down to which finding superseded which?
They said: It only ever answers. It never brings anything of its own. What if it kept thinking overnight, and came back with ideas of its own?
They said: Real problems do not fit in a context window, and managing it is unpaid work. What if context weren’t limited, and managing it weren’t your job?
03 The Category: the two funded bets, and where memory tools sit
AI has two funded learning bets: retrain the weights (personalization, one model gets smarter) and generate new data (robot labs). Both are real. Hypotheses are not cures; the world still has to answer. And neither bet can carry what a lab knows across models, sessions, and people.
Memory tools sell storage and recall for a single model. Dozens of funded teams compete there, inside one layer of the component. That crowd is our supply chain.
Our real competition is the in-weights bet. We win where the mechanism predicts we should, on currency, scale, and accumulation. We ran that comparison as one pre-registered experiment and published the kill condition.
Sapience covers both axes at once: breadth and nowness, where a frozen model gives you one at the cost of the other. Across fields, it bridges work no single context can hold and knows your lab down to which finding superseded which. Across time, it holds what you settled this morning and what a field learned a decade ago, and reasons over both.
04 Proof: the other two results, and the unit of knowledge
It knows things nobody told it. On aggregation questions whose answers exist in no single input, the write-time aggregation gate adds 22.8 points (95% CI +8.1 to +38.2) over a compute-matched counter emitting the identical sentence; exactly neutral on single-episode controls (paper v4.3). A storage system returns what you put in. This returns answers that were never stored.
Current knowledge: when facts change, 98 percent current-answer accuracy on 108 update probes, 70 points over similarity retrieval (3 seeds). Give the strongest retrieval baseline write order and it draws level, which tells us the gain comes from tracking order, not from a weak comparison.
Knowledge travels in two containers today: papers, fifteen pages, years late, with provenance no machine can read; and weights, where everything is absorbed and nothing travels out. The knowledge object is the container we built instead: one claim carrying its evidence, its reasoning, its typed links, and its author. The reasoning is worked out once at capture and reused after that, where a session log makes every later query re-derive it. That is why the store stays small and fast to reason over even at field scale, and the typed edges are what the discovery engine reasons over. The structure itself does measurable work all the way into the weights: replaying structured objects beats replaying the same content as raw prose by 10 to 12 points at a matched token budget (3 seeds, pre-registered).
05 In the field: what the product does overnight
Overnight consolidation is live: it comes back in the morning with grounded connections between your own open problems, each one citing the objects it came from.
The researcher quotes in the appendix are unedited. They describe one brain compounding across providers, surfacing new ideas and connections, and cross-domain links a researcher would not have found alone.
06 Where the time goes: the mechanism behind the 72 percent
The oldest loop in science, ideas crossing fields, still runs on decades: Boolean algebra waited 83 years for circuit design. In research first, across academia and industry, where the price of frozen knowledge is highest, and then in knowledge work generally.
Structured, cross-domain matching covers that gap. On imports that actually happened, the true source sits in a top-40 slate a median of a decade before anyone reaches for it (a floor, not a cherry-pick).
Reasoning outside the box is where current AI is weakest: it completes patterns close to what it has already seen. Structured knowledge lets the engine trade off deliberately between nearby associations and far analogies, matching by mechanism across fields. The imports in the figure are that tradeoff at work.
Science is where this matters most, because nowhere else does knowledge move this fast and travel this badly. And the discovery architecture compounds: every brain that joins adds structured, typed knowledge the matching engine reasons over, so it gets better the more it connects. One case from the figure: given a 2025 zinc-battery problem blind, the engine surfaces a 1986 solvation-chemistry source in its slate, 39 years after it was published.
08 Timing: data supply, cost structure, and what is open
The industry’s data supply curve is running out: frontier labs now bid millions for defunct companies’ internal chat logs. The scarcest training signal is expert reasoning, and the store captures it at write time: structured, current, owned by the people who produced it.
Capital-efficient by design: we do not pre-train. The cost advantage is structural. An 8B open model on the store beats a model 30x its size reading the truncated raw context. A stronger reader still helps within its own window, and either way there is no per-token rent, so no vendor owns us.
The standard is open. The network is open to join. The intelligence that runs it is ours. Revenue starts as subscription, and a larger share may later come from royalties on transacted value. Pricing is live; the table is in the appendix, since it is not what this round is priced on.
Seed round, weighted to senior research and engineering hires.
09 The Horizon: the substrate argument, drawn out
The same knowledge object that gives one person’s AI memory - a claim with its evidence, its reasoning, and a name attached - is exactly the object that lets two people’s AIs connect. Matching, consent, and attribution are properties of the substrate, not features bolted beside it.
Publishing in the age of AI: knowledge broken into small credited claims, with a feed. Contributions surface by what they solve, and anyone can plug their own knowledge into a live discovery process.
Each field joins the shared haystack by consent, with credit. A blind match keeps sharing safe: both sides learn a connection exists; the insight stays hidden until both opt in. Sharing between two brains is topic-scoped, read-only, revocable, across vendors. A verified connection between two labs is not yet on record; it is a named milestone of this raise.
Consider two materials labs that choose to collaborate through Sapience. They share one continuously learning system: it learns from their combined work as it happens, holds both labs’ full context where no window could, and surfaces connections between them, credited to whoever established each claim. And a measured path carries what it learns into the weights: in controlled, pre-registered runs, adapter-trained retention lands 47 to 75 points above the no-adapter control while learning the new domain as well or better, across two model families, three seeds each; a lab result, not what ships today (the mechanism, drawn out: Appendix A3). No frozen model offers this, at any size.
The only superintelligence that matters is personalized superintelligence: yours, grounded in a living model of your work, compounding with every mind that chooses to connect.
10 The Ask: wedge, sizing, and cost structure
Science is the wedge, not the ceiling. The substrate runs on any working knowledge, and the named second beachhead is builders: a repository’s decision history is working knowledge at its most concrete, and the coding result earlier in this deck is that market’s evidence. The wedge is the world’s researchers inside roughly $3T of global R&D a year; the reach is a billion knowledge workers, under $115B of enterprise-AI spend in 2026 heading toward $270B. Science is the hardest version of knowledge work, which is why the architecture carries to the rest without being rebuilt.
How the number is sized. The core plan is people: roughly 24 senior research and engineering hires on 24 months of runway, serving open weights rather than pretraining. Beyond the team, the round buys the two milestones that cannot wait: the weight path productized, and the network reaching self-sustaining density while the system-level learning space is still open. We hire from the same pool as the in-weights labs, at the same level, without the pretraining bill. The result is a frontier-lab raise with the compute line removed.
Capital-efficient by design: we do not pre-train a frontier model (~6N FLOPs per token, trillions of tokens, many runs). We serve open weights (~2N per token) and build the learning system on top, so compute is ~12% of the raise rather than ~70%, and the largest line is people: growing the full-time team toward ~24 across engineering, research, and go-to-market, on roughly 24 months of runway.
