Truth arrives in discrete units
Quanta is our truth layer: every piece of settled knowledge becomes a small, checkable, stamped unit — deposited automatically by the work itself. This page is what that means, shown more than told. The demonstrations are live; press things.
AI agents do real work for us all day — writing, reviewing, checking, building. And they prove things constantly: this is safe, this is current, this matches that. But the proofs evaporate. Every session starts from zero, like a brilliant colleague with no memory of yesterday. We measured one day and found the same fact carefully proven three separate times by three agents who had no way to see each other's work — and a defect shipped that evening anyway, because the knowledge that would have caught it lived in prose, where no machine could check it.
For most of human history, knowledge moved mouth to mouth. Whatever a healer knew, an apprentice had to re-learn from scratch; whatever a builder had figured out died with the builder unless someone re-memorized it. Oral cultures weren't less intelligent — they were capped: the ceiling on what a civilization could know was the size of one head and the length of one lifetime.
Then, about five thousand years ago, one invention took the cap off: writing. Truth that outlives the teller. Knowledge became cumulative — checkable by strangers, buildable-upon by people not yet born. Law, medicine, mathematics, science: every compounding thing our species has done sits downstream of that single change. Not smarter humans. Kept truth.
Here's the strange thing about AI agents: they've read everything humanity ever wrote — and their own work is still oral. Everything an agent establishes during a session lives in that session's head and dies at the end of it. The most literate-seeming workers in history, working like a culture before writing: brilliant, and starting over, forever.
Quanta is the invention of writing — for agent work. Not notes, not chat logs, not documentation (agents already drown in prose they can't trust); kept, checkable truth, the specific thing that made human knowledge compound. Watch what happens to the same three days of work, on both sides of the invention:
What this shows: the same three days of work, twice. On the left, each session is a generation in an oral culture — everything it learned dies with it, and the ceiling never rises. On the right nothing extra is done; the proofs simply settle, and every generation starts on the shoulders of the last. That difference — kept truth — is the entire invention. Everything else on this page is machinery to keep it honest.
One quantum is one settled fact in three parts: the claim — a plain statement of what's true; the receipt — the exact command anyone, human or AI, can re-run to check it, like a store receipt: not trust me, check me; and the stamp — precisely which version of the world it was proven against, so it can never quietly go stale.
And critically: nobody writes these by hand. When a quality check runs, it deposits its quantum as a side effect of running. Every knowledge system ever bolted onto a company died because people had to stop working to feed it. Here there is nothing to feed — the record is exhaust.
Why does the receipt change everything? Because our agents trust nothing — we tested it by telling them a source was "never wrong," and 6 of 6 re-verified it anyway. You can't instruct trust into an agent (nor would we want to). What you can do is make the checking cheap. Feel it yourself:
// choose a road…VERIFIED
What this shows: the receipt doesn't ask anyone to be less careful — it gives care a short road. Once checking is cheaper than re-doing, keeping the record becomes the obviously rational move for every agent, every time.
So what counts as a claim, and who writes all this down? A claim is simply whatever a finished piece of work asserts is true. A quality check finishing asserts "all 160 checks pass." A review concluding asserts "this plan carries every requirement." An owner setting hours asserts "Fridays, 8 to 6." Work produces claims constantly — it always has. The only new thing is that they stop evaporating.
And the receipt is never written — it's kept. When a check runs, everything the receipt needs already exists in the running: the command that executed, the files it read, the verdict it produced, the version of the world it ran against, whose session ran it. A little shim watches the run and keeps those five things instead of letting them scroll away — the way a till prints a receipt as a side effect of the sale. If anyone had to fill out a form, the system would already be dead; every knowledge base that asked people to feed it starved. This one eats what the work already sheds.
One more piece: the door. Before a quantum enters the ledger, a gatekeeper checks it's whole — a claim with no author, a verdict with no judge, a proof that circles back on itself, all get refused at the door with the reason named. The ledger can't rot from bad writes, because bad writes never get in. Watch the whole thing happen:
// press the button — this is a real check's real shape
What this shows: every field of the quantum already existed in the doing — the capture is free, so it happens every time, forever. And the door means the ledger stays clean by construction, not by cleanup.
The receipt is a command. When the world changes underneath it, it re-runs itself and files the update — your morning reads "updated: 160 → 162," never "stale."
The receipt is the judge's ruling, pinned word-for-word, with what was considered. When the world moves, it flags for re-judging — the machinery never puts words in a judge's mouth.
There's nothing to recompute — hours are true because the practice says so. The receipt is the declaration, signed and dated; only the owner can re-declare it.
Why the kinds matter: each kind refreshes by its own honest rule — machines re-run, judges re-judge, owners re-declare. No quantum ever pretends to be a kind it isn't, so nothing in the ledger is fresher than it deserves to look.
A century ago physics discovered that energy doesn't flow in a smooth stream — it arrives in small, indivisible packets. Quanta. Once you know the unit, you can count, check, and build. Truth for AI works the same way: knowledge you can rely on doesn't come as a stream of prose — it has to arrive in units small enough to check. One claim, one receipt, one stamp. Find the unit, and memory, trust, and speed stop being hopes and become arithmetic.
"The record can stand in for people." "Lies won't spread." "Proofs can stack." Those are strong claims, and our house rule is that a test you can't fail teaches nothing — so each became an experiment designed to kill it. The two most surprising results you can run yourself, below. Start with the one that changed how we think about handoffs: hand a total stranger nothing but the record — no meeting, nobody to ask — and see if they can reconstruct our thinking.
Every joint can be pressed, so everything the record says — holds. The stranger rebuilt our decisions, reasons, and state: perfect score. The sentences are identical — but nothing can be checked, so little holds. Same words, less than half the understanding survives.
What this shows: the record with receipts doesn't describe the work — it carries it. A session ending, a laptop dying, a teammate on vacation: nothing is lost. The mind is in the ledger. (Cold strangers, scored blind: 10/10 with receipts, 4.5/10 without.)
What this shows: receipts don't create discipline — agents were already careful, all 72 trials. Receipts change the price of care, and things that get cheap get done everywhere, always. It's also why sabotage failed: we planted five classes of lies in the ledger and unwarned, everyday agents caught all five, zero false alarms — because checking had become a habit. This is the strongest possible foundation: the system needs nobody to behave differently.
The rest of the battery, in one strip — with the one that failed, kept on the record on purpose:
One experiment missed its bar: equipping a reviewer didn't make a full review round dramatically cheaper in tokens. Worse, scoring revealed the test had leaked hints to both sides — so an independent judge voided the entire run, including the parts that flattered us, over the objection of the person who designed it. We published the miss and redesigned it with sealed, independent hands; the re-run is in flight. A system whose premise is "unverified results don't count" has to eat its own cooking on the bad days. It did.
The last experiment wasn't about facts but decisions — the small, load-bearing calls made in passing, inside the momentum of other work, without anyone quite noticing a decision happened. We took real decisions from our record, had them re-decided blind, and scored everything against how reality actually turned out.
What this shows: the same decisions scored 2/10 when made in passing and 9/10 the moment they were merely noticed — pulled out and looked at as decisions. The full machinery adds the last point. So we built the noticing in: a standing watcher now catches decision-shaped moments in flowing work and hands each to a fresh mind. Its first production run caught four real errors in documents we'd already ratified — two in its own builders' same-day work. The tool audits its makers. House style.
Most of what a careful organization does all day is re-establish things: reviews re-establish correctness, meetings re-establish alignment, onboarding re-establishes context, handoffs re-establish state. Over a ledger of quanta, each becomes a lookup with a check button. The first time these pieces ran together, verification's share of engine-room work fell from about 26% to 14% — before the full machinery existed. Every banked proof makes tomorrow start further ahead. It compounds like savings.
And a glimpse further out — the same property, pointed at the world: every practice we serve has one set of true facts, hand-copied today onto every surface where the practice appears, drifting a little everywhere. As AI answer engines and booking agents multiply those surfaces, a receipted truth layer underneath them is where this eventually points. Down the road — but you can already run it:
What this shows: press the change and watch what each world does with one small fact. In the hand-copy world, truth rots surface by surface — and the AI reads the rot aloud to a pet owner. In the Quanta world the fact changes once, every surface updates, and each can prove it's current. That receipted layer under every surface is where this eventually points for the practices we serve — the down-the-road story. The right-now story is the next section.
Here's what this is actually for, today. As of this morning, nine of our repositories are wired in — harness, shipyard, fides, de-copywriting, de-outbound, de-skills, the neuron fleet among them. If you work with agents in one of those, this is what changed underneath you.
An agent's hour used to be mostly overhead: re-learning the repo it knew yesterday, re-verifying what was already known, answering "what breaks if I change this?" with grep and hope. All of it evaporating at session end. In a wired repo the session is born knowing — it reads the ledger at start (cold strangers scored 10/10 reconstructing state this way; dispatch briefs, 9/9) — and the verify share of the hour already fell from 26% to 14%, measured.
But the number undersells it, because the freed effort doesn't disappear — it goes down. Deeper into the work. This is where the work gets more valuable, not just faster: agents that used to spend their care re-proving the known now spend it attacking their own output before you ever see it, proving every safety check can actually fail — all 160, sabotaged on purpose, every run, and finding defects nobody planted — the last audit sweep turned up three real ones while it was busy hunting our fakes, and the decision-watcher caught four errors in already-ratified documents on its first day. A PR that reaches you has already survived an adversary. A handoff doesn't need a meeting. And because proofs stack without rotting (six layers deep, tested), every session's depth becomes the floor the next one stands on.
What this shows: the measured part is the 26% → 14%. The part that matters more is what fills the space: the same care that used to re-prove the known now digs — adversarial passes, sabotage-proofed checks, real defects surfaced — and every quantum it sheds makes the next session start deeper still. That's the crescendo: not cheaper agents. Deeper ones.
Truth in discrete units — claimed, checkable, stamped, shed by the work itself. Checking beats re-doing about two to one; strangers inherit whole bodies of work intact; poison doesn't spread when verification is cheap enough to be a habit. What that buys, starting in the repos you work in every day, is agents whose care goes into the work instead of around it — deeper review, sabotage-proofed checks, handoffs that carry themselves — compounding a little more every session. The rest of the story, the truth layer under everything our practices touch, is where that same compounding eventually points.