Part 1 named six failures. Part 2 named the engine that corrects them. Part 3 told you how to live inside it. This part is the one that earned its own page.
It is about the claim the rest of the architecture exists to make: Carabase doesn’t just remember what you said — it knows what you mean. That sentence sounds promotional. It is not. It is the literal output of the layer that sits above retrieval — the layer that consolidates a thousand scattered signals into a named view, attaches the receipts, and keeps it current. The recall half is the part everyone will eventually ship. The synthesis half is the part nobody has shipped, and the part that compounds.
This is the long-form home of that claim. Five sections. The Synthesis Tax, in depth. How a view is formed from scatter, with worked examples. Taste and positions as one mechanism. Knowing when to shut up. And the line the privacy story has to hold, because if any part of this reads as “Carabase profiles you,” the whole edifice falls.
The Synthesis Tax, in depth
Part 1 introduced the Synthesis Failure as the sixth indictment. Part 2 named the Synthesis Tax as the third cost the status quo charges. This is the longer version: where the tax actually shows up, why nobody has named it before, and why it is the load-bearing part of the new claim.
You pay the Synthesis Tax every time somebody asks you what you think.
What do you think of the new vendor — should we sign? You start assembling: the demo three weeks ago, the reference call that went sideways, the price compared to what we paid last time, the gut-feel you texted a friend about, the thing the CTO noticed and you nodded at without writing down, the deal’s analogue to that other one in 2024 that didn’t work. Twenty minutes of head-archaeology before you can return a sentence. Twenty minutes that don’t go on a timesheet.
What do you think of the engineer you’ve been mentoring? You assemble: six months of one-on-ones, two strong PRs you remember and probably three you’ve forgotten, a moment in the standup last week where something landed differently than usual, the read from their last skip-level, the comparison to where you were at their tenure. Half an hour of work to return a paragraph.
What should you order off this wine list? You assemble: where you’ve drifted this year (off crisp whites, into oxidative things, away from Burgundy that’s become too expensive), what you said to Dan after the dinner in March, the Overnoy you keep recommending and have started feeling slightly cliché about, what you actually felt like when you opened the door of this restaurant. The assembly happens in seconds for a low-stakes case and minutes for a high-stakes one. Either way, you pay.
The recurring shape: the view exists, but it is not anywhere. No single document contains it. Nobody — not your second brain, not your AI assistant, not the most attentive friend — can point at a file and say “that’s what she thinks.” The view lives smeared across the evidence, and the only synthesizer ever invited to assemble it was you.
Why hasn’t anyone named this tax before? Three reasons.
One: synthesis is invisible work. Nobody bills for it because nobody bills for thinking. The work happens before the sentence comes out of your mouth, in your head, at 11 pm on a kitchen floor, and the only artefact is the sentence.
Two: the technology to actually do anything about it has only just arrived. Without local-first capture across enough surfaces, you don’t have the scattered signals. Without entity resolution across sources, you can’t cluster them. Without an LLM cheap enough to run synthesis overnight on a year’s worth of fragments, you can’t name the view. Without retrieval fast enough to surface the view in the moment, you can’t use it. Each of those was prohibitively expensive in 2023. None of them is now.
Three: the demos are bad. “The engine forms a view of your taste in wine” does not slap in a tweet the way “the engine remembers last Tuesday” does. Recall has a punchline. Synthesis has a long answer. So the industry shipped the punchline and skipped the long answer, and the long answer is where the actual product lives.
That is the tax. It compounds with every year you accumulate context. It is the part of your day that you don’t even feel as work because you have been paying it forever. We are going to zero it.
How a view is formed from scatter
Take wine. It’s the case the demos under-sell because everyone thinks taste is unserious. Taste is the universal case — everybody has it, almost nobody writes it down, and the mechanism for forming a taste-view is identical to the mechanism for forming a deal-view. Get the wine case right and the rest follows.
Here’s what the engine actually does, on a hypothetical year of yours.
It has watched 84 receipts land — merchants, restaurants, wine bars. It has watched 47 photos cluster into wine-shaped scenes: your hand around a glass, a bottle on a table, a printed list with the “Jura” section underlined. It has watched 22 messages to friends about specific bottles — “the Overnoy was unreal,” “don’t order Sancerre at this place,” “a friend brought a bottle that made me reconsider whites again.” It has watched 4 Readwise highlights from natural-wine writers — Alice Feiring on Jura, a Jancis Robinson column on oxidative whites, a Punch piece on low-intervention. It has watched 11 photos of restaurant wine lists where the trail of evidence ends with a receipt for the low-intervention pick.
It has watched, also, the negatives. It has noticed that you ordered a Sancerre once in April and didn’t order one again. It has noticed that you texted a friend saying “I think I’m off crisp whites this year” in May, and then the receipt pattern actually changed in June. It has noticed that the Burgundy line has gone quiet — the receipts stopped, the messages stopped, the photos stopped. Silence is data too.
The dream cycle runs at 3 am. The worker that does this is not a chatbot. It is a clustering pass that builds candidate clusters from the evidence, a naming pass that proposes a one-line label for each cluster, a verifier that checks the label against the evidence and rejects the ones that don’t hold, and a consolidation pass that merges related clusters into a single named view. The output is a small set of named views, each with: a one-line claim, a confidence score, the strongest 14–18 pieces of evidence, the date of the most recent reinforcement, and a hold-or-cool status.
By morning, you have a view sitting in your engine. “Drift: crisp whites → oxidative / low-intervention, since Q1. Hold: high.” You can open it. You can read the receipts. You can edit the one-line claim. You can fork it. You can kill it. You can ask “help me pick off this wine list” and the engine answers from the view, not from the most recent receipt — which is the whole point.
A view is not a fact. A fact has one true value at a time. A view is a consolidated position with evidence behind it — the kind of thing a colleague who’s known you for years would say if you asked. The view is editable because you hold the position, not the engine. The view’s evidence is visible because you can be wrong about your own taste; the receipts are how you check. The view goes cold over time if no new evidence reinforces it — that’s the engine’s way of saying “you used to think this; we haven’t seen signal for it in four months; verify before you cite it.”
This is the substrate the rest of Carabase’s “knows you” claim sits on. There is no model in there generating opinions about you from a personality test. There is only a clustering pass, a naming pass, a verifier, and a consolidation pass, run over evidence you produced and the engine harvested. The view came from you. The engine just made it legible.
Taste and positions are the same mechanism
Now run the same mechanism at higher stakes.
You’re in a deal. Over six weeks, the engine has watched a dozen meeting transcripts, three redlines on a term sheet, a long Slack thread with your lawyer, the gut-check you texted your co-founder, four prior deals where you held the same line in the same way. It clusters those into a view: “your position on the indemnity cap.” It attaches the evidence. It knows the live value of that cap — the one you actually signed — is 1× ARR in the v3 term sheet, two weeks ago, not the “we’re flexible on caps” message you sent eighteen months ago in a different conversation. When someone asks what your position is, the view returns 1×, with the v3 doc as the live evidence, and your prior consistency as the supporting pattern. You hold the position, the engine renders it back to you faster than your memory could.
One mechanism. Two calibrations. The wine case and the deal case run on the same view-forming machinery. What changes is calibration: how strict the engine is about which evidence outranks which, how loud it gets when sources disagree, and how willing it is to return a view at all.
In the wine case, calibration is loose. The cost of being wrong is small — you order the wrong glass, you laugh about it, the view updates. The engine will happily synthesize a view from receipts and messages because the worst-case downside is a mediocre Tuesday dinner.
In the deal case, calibration is tight. The cost of being wrong is large — you cite the wrong number in a closing email and now the lawyers have a problem. The engine treats signed documents as the source-of-record, contemporaneous notes as supporting, and confident-sounding messages as paraphrases that don’t get to override the document. Disagreements between sources of comparable authority get flagged, not silently resolved. Views in this register hold to a different bar.
The temptation, when reading this, is to assume there must be two products — a fun one for wine and a serious one for deals. There isn’t. There is one engine that you have already met. It exposes the calibration as a property of the domain, not as a separate codebase. Same retrieval, same provenance, same moss-and-ember discipline. Same view-card with the layered edge that says “sits on evidence.” What changes is one calibration parameter and a couple of guardrails. That’s it. That’s the whole architecture commitment behind “one engine, two registers.”
If you find yourself wanting to assign your taste to a different category from your positions — one playful, one serious — that’s a category mistake the rest of the industry will probably make. Taste is not unserious. Your taste is information about you that has been thrown away every year of your professional life because no tool was good enough to consolidate it. The engine treats your taste with the same care it treats your positions, because under the hood it is the same thing: a view of who you are, derived from evidence you produced, made citable.
It knows when to shut up
A chatbot answers everything, confidently, including the things it’s wrong about. This is the deepest tax the LLM era has imposed and the one professionals quietly absorb every day: a discount factor on every model output, varying by domain, that you mentally apply because the model will not apply it for you. It is exhausting. It is also corrosive, because the parts of the model output that are right and the parts that are wrong arrive with identical confidence.
Carabase abstains. The abstention contract is the property that, more than any other, makes the engine credible at professional stakes.
It hedges when the evidence is thin. If you ask about a fact and the only supporting source is a single low-authority artefact — a Slack message, an offhand voice memo, a forwarded email — the answer comes back framed as a belief, not a fact. “You mentioned X once in a Slack message to David. No other source confirms. Verify before relying.” You get the belief, you get the weakness, you make the call.
It flags when the most authoritative source is stale. A signed v3 term sheet says 1× ARR. A v4 redline sitting in your inbox unsigned disagrees. The engine’s answer reads: “Last authoritative value: 1× ARR, as of the v3 term sheet (signed). The v4 redline (unsigned, in your inbox) may supersede this — verify before you cite it.” Out-of-date authoritative facts are not silently down-weighted. They are surfaced with the disagreement attached, because the disagreement is the load-bearing signal.
It refuses and flags when sources of comparable authority disagree on a value. A signed counterparty document and a contemporaneous note that say different things about the cap is a question for a human, not a model. The engine returns both, surfaces the conflict, asks for your call. There is a Review queue item generated. Nothing is silently picked.
The most consequential of these is the one nobody else will ship: when your own confident message contradicts a signed source-of-record, Carabase believes the document.
The principle behind that is authority is not authorship. A source you didn’t write can be the most authoritative source on a fact about you — a contract you signed, a document a lawyer prepared, a record an institution keeps. A source you wrote can be the least authoritative source on the same fact — a confident message at 2 am, a paraphrase from memory, a generous summary you sent to a friend. Mixing these up is the failure mode that has destroyed careful work since the typewriter. Carabase, by construction, will not mix them up.
The consequence is a property no chatbot can credibly claim: an engine that can be right against you. Most of the time this is invisible. Occasionally it saves you from an embarrassment — the cap you confidently remembered was 2× was actually 1× in the document you signed, and the verifier caught it before you sent the email. Rarely, it saves you from something worse. The reason to trust it against anyone else is that you have watched it be right against you. That is the trust property. It is not optional.
The matching design property: a hedged or flagged answer should look different from a confident one. We’ve added a quiet confidence chip on every answer and a distinct flag treatment for the refuse-and-flag case. The visual is calibrated to be honest, not promotional. The engine telling you it’s 0.62 confident on something is more useful than a trust-me badge that’s always green. The UI’s job is to make “the engine is hedging” legible at a glance.
The line that has to hold
Everything above lives or dies on one sentence: it knows what you think because you told it — across a thousand fragments — and it shows you the receipts.
That sentence is the entire privacy contract. It must hold everywhere, in every surface, in every demo, in every conversation. If any part of the experience reads as “Carabase profiles you,” “Carabase infers traits about you you didn’t express,” or “Carabase knows you better than you know yourself” — the whole edifice falls. Not because the privacy story would be weaker. Because the product would have become a different product, and a worse one.
The architectural commitments that make the line hold are the same ones Part 2 specified, applied to views. Restating them, because they matter most here.
Provenance is mandatory. Every view carries the evidence that produced it. You can open a view and see the 14 receipts, 6 photos, 22 messages it was clustered from. There is no view without receipts. A view without provenance is a hallucination by another name; the engine does not ship those.
Moss, ember, and the derived edge are an epistemic contract. What you authored is moss. What the engine harvested from your sources is ember. What the engine derived from both — a view — sits on a visible layered edge, with the evidence collapsible beneath. You can always tell the difference between something you said, something a connector pulled, and something the engine consolidated. This is not a design flourish. It is a promise that the engine will not pretend its synthesis is your authorship.
It is yours, materially. The substrate runs on your hardware. Bodies stay on your drive. Zero public ports on the Host. Row-level security. The Host code is open source. If we disappear tomorrow, your engine, your views, and the receipts behind them keep working. This was Part 2’s structural commitment; the view layer inherits it without exception.
Per-class opt-in. Forming taste views from your photos and messages is one capability. Forming professional views from a deal corpus is another. They are separate switches, both off by default. You opt in once per class, and for high-stakes professional knowledge, per scope — a per-deal grant, a per-client grant, not a single global flag. Deal knowledge from one engagement does not prime answers in another; client A’s substrate does not leak into client B’s. The scope is enforced as a capability grant at the read path, not as a promise.
And the framing rule that has to hold in every external sentence the team writes: the claim is consolidation of what you expressed, with receipts. Never divination. The engine does not psychometrically infer hidden traits. The engine does not generate opinions about you from a personality model. The engine does not know you better than you know yourself. The engine reads what you wrote and what the connectors gathered, clusters the scatter, names the view, attaches the evidence, and shows you. The synthesis is computation over evidence you produced. If that distinction blurs in a sentence, the sentence is wrong and needs to be rewritten before it ships.
What this comes to
The status quo gave you a chatbot that forgot you every Tuesday. The second-brain era gave you a job as an unpaid librarian and called it a feature. The retrieval era will give you a memory feature in 2027 that’s a worse version of the same recall promise. Each of those treated the missing piece as a technical problem at the wrong layer.
The missing piece is synthesis. The thing you have been doing yourself, at 11 pm, on a kitchen floor, for your entire professional life. The view-forming the colleague who’s known you for years does without thinking about it. The thing that compounds because it actually models you, rather than searching the things you typed.
Carabase forms views from your scatter, on your hardware, with the receipts visible, abstaining when the evidence is thin, flagging when the source has gone stale, and trusting the document over the message — including when the message is yours. It does not profile you. It consolidates what you have already expressed, makes it legible, and hands it back.
It doesn’t just remember what you said. It knows what you mean. That sentence is not marketing. It is what the engine does, in the order it does it, with the property guarantees attached.
If you have read this far, you are why we are building this. The private alpha application is below. So is Part 3, if you want the practical guide to actually using it. And the rest of the manifesto, if you want to start from the indictment.