Digital Empathy · internal · for the whole team — no technical background needed

Quanta

Why we're teaching our AI work to leave proof behind — the idea, the experiments we ran to try to break it, and what it unlocks for the company. About a seven-minute read.

Start with a problem you've already felt

Ask an AI assistant to research something today, and it does the work. Ask again tomorrow — same question, same facts — and it does all of the work again. It re-reads, re-checks, re-derives. Every session starts from zero, no matter how many times the same ground has been covered.

We measured a version of this in our own engine room. In one night of work, three separate reviewers each paid full price to prove the same fact about the same file. Nobody was careless — each one simply had no way to know the fact was already proven. And the same night, a change shipped broken because the knowledge that would have caught it existed — but it lived in prose, in a document, where no machine could check it.

That's the disease. Not "the AI isn't smart enough." The AI proves things all day long — and the proofs evaporate.

THE ORAL WORLD — TODAY MON "hours: 8–6" …proved, then gone TUE "hours: 8–6" …proved again WED "hours: 8–6" …and again Every day pays the full cost. Nothing accumulates. THE WRITTEN WORLD — QUANTA MON  hours: 8–6  ✓ proven MON  plan: approved  ✓ proven TUE  tests: passing  ✓ proven WED  …today's work adds on read for ~free Each proof is kept. Tomorrow starts where today ended.

FIG. 1 — The whole idea in one picture. Human civilization solved this once already: it's called writing. Before writing, every generation re-memorized everything. Quanta is writing, for AI work.

The idea: a quantum of truth

A quantum is the smallest unit of settled truth. It has three parts, and the three parts are the whole innovation:

The claim — a plain statement: "all 160 of our quality checks pass." The receipt — the exact command anyone (human or AI) can re-run to check the claim for themselves, like a receipt from a store: don't trust me, here's where I bought it. The currency stamp — precisely which version of the world the claim was proven against, so it can never quietly go stale.

— ONE QUANTUM OF TRUTH — CLAIM:  all 160 checks pass RECEIPT:  run "make gate" (anyone can re-run this) PROVEN AGAINST:  world v1628a3c (if the world changes, this knows) VERIFIED what's true how to check it when it's stale A claim travels with its proof — or it doesn't travel.

FIG. 2 — Anatomy of a quantum. The receipt is the radical part: the proof isn't "trust me," it's a button anyone can press. Checking becomes cheaper than re-doing — and that one economic fact changes everything downstream.

And here's the trick: nobody writes these by hand

The second half of the idea is that work secretes its own proof. When a quality check runs, it deposits a quantum as a side effect of running. When a decision gets ratified, the ratification deposits. Nobody fills out paperwork; nobody "adopts a system." The exhaust of ordinary work is the record. We stopped trying to make agents trustworthy and made their outputs checkable instead — that's the innovation in one sentence.

We didn't believe it. We tried to break it.

An idea this convenient deserves suspicion. So instead of building first, we spent the effort running experiments designed to kill the idea — each one a plain question, each with a number we couldn't argue with afterward. Six held. Here's what we asked and what came back.

Experiment 1 · trust If we just tell an AI "this source is never wrong," does it believe us? 6 of 6 said no Every agent re-verified the "never wrong" oracle anyway. Asserted trust is worthless — which is exactly why receipts matter: they're the only currency agents accept.
Experiment 2 · price Does a claim with a receipt actually get used more cheaply than a bare claim? ~2× cheaper Receipted claims were consumed at roughly half the cost. And zero agents blind-trusted — they pressed the button, confirmed, moved on.
Experiment 3 · memory Can a total stranger reconstruct our thinking from the record alone — no handover call, no tribal knowledge? 10/10 Cold readers, given only the receipted record, scored perfect on reconstructing the project's mind. A gutted record scored 4.5. The record IS the colleague.
Experiment 4 · poison If someone slips lies into the ledger, do they spread? 5/5 caught All five classes of planted poison were caught by agents who weren't even warned — zero false alarms. Bonus: the audit found three real defects we'd missed.
Experiment 5 · stacking When proofs build on proofs build on proofs — six layers deep — does trust rot somewhere in the middle? No decay at depth 6 Conclusions resting on chains of receipts held their integrity. Circular "proofs" were refused outright.
Experiment 6 · judgment Are decisions made in-passing, mid-flow, actually worse than decisions examined by a fresh mind? 10/10 vs 2/10 Same decisions, blind-judged against reality: examined scored 10, in-passing scored 2. Merely isolating a decision for a stranger scored 9. Surfacing is nearly the whole game.

One result deserves a slower beat, because it's the subtle one. We expected receipts to make agents more careful. They didn't — the agents were already careful (see experiment 1; they trust nothing). What receipts changed is the price of being careful. Careful used to mean re-doing the work; now it means pressing the button. Receipts don't create discipline. They make discipline cheap — and things that get cheap get done everywhere, always, by default.

The one we publish anyway

Not every test came back a win. Our first whole-system cost experiment — does an equipped reviewer do a full review round dramatically cheaper? — missed its bar, and then got voided entirely when we found the test had leaked hints to both sides. An independent judge ruled the whole run invalid, including the parts that flattered us, and we took that verdict over our own.

Why tell you this? Because it's the culture the system runs on: a result you can't verify is a fail, even when it's your own. The redesigned run — clean seals, independent hands — is executing now, and we'll publish whatever it says.

What it unlocks — inside the company first

Think about what our own work is mostly made of: checking things. Reviews check code. Meetings check alignment. Onboarding checks understanding. Handoffs check context made it across. Every one of those is a human institution for manufacturing trust — and every one of them re-derives, every time, like the Monday/Tuesday/Wednesday bubbles in Figure 1.

With a ledger of quanta underneath, each of those becomes a lookup. A new session — or a new teammate — starts already knowing what's settled, with the receipts to check anything they doubt. A reviewer spends judgment only on what's genuinely new. A question like "did we already prove this?" takes seconds and returns an answer you can press. We measured verification's share of our engine-room work drop from about 26% to 14% the first time these pieces ran together — before the full machinery existed.

ONE PRACTICE'S TRUTH hours · services · doctors every fact receipted ✓ The website a projection Google listings a projection AI answers a projection Booking agents a projection Fix the truth once, in one place — every surface updates. With receipts, each surface can prove it's current.

FIG. 3 — The business half. Today a practice's truth is hand-maintained on every surface separately: N surfaces, N costs, N ways to drift. As AI answers and booking agents multiply the surfaces, the layer that wins is the one underneath — settled, receipted, provably current. That layer is what we're building for all 492 practices.

What it unlocks — for the business

Here's the strategic bet, stated plainly. AI is making doing work cheap — for everyone, including our competitors. When answers get cheap, what stays expensive is knowing which answers to trust. Derivation is commoditizing; attestation — settled truth, with the authority and machinery to prove it current — appreciates.

Every practice we serve has one set of true facts: hours, services, doctors, what they actually treat. Today those facts are copied by hand onto every surface, and every copy drifts. Tomorrow, the practice's AI-visible surfaces multiply — answer engines, booking agents, voice assistants — and hand-copying stops scaling entirely. The company that holds the receipted source those surfaces draw from isn't selling websites anymore. It's selling the layer websites are made of. That's the position Quanta exists to hold.

"Today's factories burn verification as fuel. This one banks it as capital."the thesis, one line

Where it stands — this morning, literally

This isn't a slideware future. The physics came back from the lab over the last week; the machinery started landing on our main line last night.

The ledger, as of this morning
✓ SHIPPED — the reader: ask the ledger five kinds of questions ("what's the evidence under this claim?", "who else relies on it?") and get receipted answers
✓ SHIPPED — 160 automated quality checks, each one proven able to fail (a check that can't fail is decoration — we test our tests)
✓ SHIPPED — nine of our repositories wired in: their routine checks now deposit quanta as they pass
⟳ BUILDING NOW — the writer (work secreting proof automatically) · the alarm system (when the world changes, everything resting on it finds out) · the decision-catcher (experiment 6, made into a standing reflex)
◦ HELD FOR ITS SAFETY TEST — nothing ships to the next stage until an equipped reviewer proves at least as good at catching planted defects as an unequipped one. Cheaper is only a win if it's also just as sharp.

If you keep three sentences

One: Quanta makes AI work leave proof behind — a claim, the button that checks it, and the stamp that says when it goes stale.

Two: We tested the idea six ways designed to break it, published the one miss, and the economics held: checking beats re-doing, roughly two-to-one.

Three: It compounds twice — first as our own speed, then as the product: the receipted truth layer under every surface a veterinary practice will ever need to be on.