Syntaxe is a private research lab turning dead artifacts into living understanding through real mathematics — reasoning engines for truth, justice, memory and craft, with proof rather than heuristics.
Reach the lab at hello@syntaxeltd.com · Syntaxe Ltd.
We turn dead artifacts into living understanding through real mathematics. The world's meaning is locked inside things no living person fully understands anymore — old systems, frozen photographs, the law, the scientific record, a human mind. We build reasoning engines that make them legible and alive again, with proof rather than heuristics.
Every engine takes something the world had given up on understanding and makes it reason for itself again — checkably, from primary sources, showing its work. Where the field ships a plausible guess, we ship a result you can verify. The output is not an opinion. It is a line of reasoning you can follow to its source.
A lab of many. The ones that wait in the Well are ready for daylight — each its own discipline, each already real. Behind them the work runs deeper than any one page should say.
For the first time, reasoning systems are precise enough to take on the world's hardest, oldest meaning and not merely paraphrase it. The moat is not the model — every lab has the model. The moat is the mathematics underneath each engine, and the depth of the bench behind them.
The lab grows by a person at a time. We look for minds that want to build things that outlast them — people who prefer proof to opinion, who would rather be right and slow than fast and plausible, who are drawn to the hardest problems because they remain unsolved.
If that is you, do not send a résumé. Send us the thing you would build, and why no one has built it yet.
Tell us what you'd build →A research lab building the engines that turn dead artifacts into living understanding — with proof, not heuristics. A long game, and a deep one.
Notes, observations and field reports from inside the lab — the thinking as it happens, kept as it was written. An evolving body of work, not a feed.
We all drink from the same well. Given the same inputs, capable minds converge on the same ideas — which is why the idea is rarely the moat. Value is what is not in the well: a position, a dataset, a proof only you can generate. Novelty doesn't live in the water. It lives at the edges no one else is standing on.
read →We turned an independent adversary loose on our own best ideas — a dozen of them, across a dozen domains. Every one cleared its own gate; the outside check killed or sketched all twelve, almost always on prior art the builder couldn't see from inside the well. The lesson was cheap and total: enthusiasm is not evidence. A thing is novel only when someone who wants it to fail can't.
read →Every guard in the agentic-security market ships an alarm — a probabilistic flag that something looks wrong. We ship a proof: a certificate an outsider can re-derive in an air-gapped room, trusting no one but the math. The whole product is the distance between "we checked it" and "check it yourself."
read →When an autonomous agent acts, who is accountable? Not the prompt, not the model — the verified outcome. A signed attestation written at the moment of execution and anchored where no one can rewrite it. Accountability as a primitive, not an afterthought.
read →The systems that move the world rarely announce themselves — double-entry bookkeeping, the handshake beneath every connection, the ledger no one reads until it is the only thing that matters. We build for the indirect: the substrate beneath the thing you actually see.
read →For thirty years we designed for the human at the keyboard. That human is becoming the one who delegates. The chain is no longer human → software; it is human → agent → system. The design question stops being "is it usable" and becomes "is it containable."
read →A reviewer gets two to four hours per paper — no time to re-derive every statistic, trace every citation, or inspect figures at the pixel level. Verity does that pass in under a minute and hands back a map of exactly what to examine. The hours they never had, spent where they matter.
No black-box "accuracy" score. We report precision and false-positive rate per layer — the numbers that decide whether a tool is safe to point at a colleague's work — validated against the Retraction Watch corpus of documented retractions. Every signal links to the method that raised it. And when the evidence is insufficient, Verity says so: "unverifiable" is a first-class result. (Operating characteristics published as the corpus is processed — we won't quote a number we can't show.)
It does not replace peer review, and it never makes a finding of misconduct. Every output is a signal — its evidence shown — handed to an editor or integrity officer to adjudicate. The instrument points; the scientist judges.
Journal editors screening before publication. Integrity offices triaging allegations. Funding agencies verifying the work they pay for. Reviewers who want their hours spent where they matter.
A compiler for science — it doesn't accuse you of error. It points to the line and says: look here.
Request a demo →Eighty-five percent of people with a civil legal problem get no professional help. The system exists — but only for those who can afford it. The law is written for solicitors. Awatum rewrites it for humans.
Finds documents containing the word "negligence," summarises them — and can confidently cite a case that does not exist.
Decomposes negligence into its legal elements — duty, breach, causation, damage — scores each against your facts, computes the odds, and models what the other side will argue. Every step inspectable.
Every claim Awatum makes carries citations to primary authority — the actual statute, the actual judgment. Any sentence whose citation isn't in the verified source is dropped before it ever reaches you. Where others guess, Awatum shows its work — or tells you, plainly, that it doesn't know.
Every person has a living map of their legal existence — the people, the contracts, the rights accruing, the deadlines approaching. Every conversation enriches it; every engine reads from it. Awatum doesn't answer a question and forget you — it holds your whole situation, and watches it over time.
After every new fact, Awatum quietly runs the opposing case against you: the strongest attack, the weakest joint in your position, the three moves a sharp opponent would try. You never walk in unprepared.
When a statute is amended, every case it touches re-reasons automatically — and the people affected see their position move in about two seconds. You are never quietly out of date.
Start with a few plain words. Watch it reason — citing real law, weighing the odds, tracking every deadline — and hold your whole legal life in one place.
Awatum meets you where you are — your situation, your language, the urgency in your voice — and never questions your intelligence. It says "you have the right to," not "you should have known." End-to-end encrypted. It never trains on your case. Your data is never sold.
Awatum makes the law legible and your options clear — and tells you, plainly, when you need a solicitor, then helps you find one. You stay in control of your own case.
The tenant whose deposit vanished. The worker fired without notice. The parent in a custody fight. The migrant facing a system in a language not their own. Anyone the law was meant to protect — and didn't, until now.
Not when you can afford it. Not when you find the right lawyer. Now — in words you understand.
Request early access →The whole agentic-security market — the funded names included — watches your AI agents and raises a probabilistic flag when something looks wrong. KERN takes the agent mesh's declared tools and data labels and proves it cannot route protected data to an undeclared sink, or exceed its authority, for any input or prompt injection — or it hands you the exact path that can.
KERN's proof carries no signature on purpose. A signature lets a vendor assert safety; KERN lets an insurer, an auditor, or a counterparty agent re-derive it — in an air-gapped room, depending on nothing KERN controls. Trust comes from re-running the proof, not from believing the prover.
For agents that move money, "don't leak" isn't enough. KERN's value predicates prove the arithmetic too — that a keeper agent can never withdraw more than a position holds, for every input — the Certora-shaped guarantee, discharged as a proof rather than a fuzzing campaign that merely failed to find a bug.
KERN does not bluff. Its guarantee holds over the declared, bounded mesh — not over code it never saw. Multi-step running invariants and live on-chain label capture are named, on the roadmap, and honest about their status. A proof that overclaims is worse than an alarm; KERN would rather show you its edge than hide it.
The team shipping autonomous agents into production who need more than a dashboard's good feeling. The insurer underwriting an agent and pricing real, re-checkable risk. The auditor and the counterparty who must trust an agent they didn't build — and now can, because the proof re-runs in their hands.
That discontinuity is the whole reason ANNEAL exists. Frontier models now find and chain real vulnerabilities at a level that used to take a specialist — and over 99% of what they find is unpatched the day it's found. Finding was never the bottleneck. Closing, faster than an attacker who now moves at machine speed, is the game.
Sorts thousands of findings by a static severity that knows nothing about your estate — most of them unreachable — and leaves the fixing, and the risk, to you.
Scores each exposure by an Attacker Priority Score — exploitability, dormancy, blast radius, reach, and chain position — surfaces the few paths that reach a crown jewel, and closes them behind a health-gate.
ANNEAL's priority isn't a vendor's opinion. Each exposure stores all five components and the chain multiplier — so an auditor can re-compute the 0.91 by hand, and a closed path carries a record anyone can re-check.
Everything above, on one console: the highest-reachable path scored 0.91, the hunts running now, the CLOSE pipeline staging fixes — and the live SENSE feed where TORPOR catches the behaviours that mean an attacker is already inside. Velocity 1.34 — closing faster than they arrive.
Connectors are read-only; hunts run against snapshots, never your live systems — and no patch reaches production without an operator's approval. It also refuses to drown you: four alerts fire, not forty — a new CISA-KEV match, velocity slipping below 0.8, a validated critical finding, a deploy that landed or failed. A tool that cries wolf is a tool you switch off.
The team drowning in findings it can't fix fast enough, that needs the path that matters closed — not a longer list. The CISO who has to answer "are we actually harder than last week" with a number that re-checks. The insurer pricing an estate's real, moving risk.
Twenty thousand genes, twenty-eight million methylation sites, one program — running right now in every one of your 37 trillion cells. We have billions of genomes on disk and almost no account of what the program does. CODEC treats regulatory interaction as a vector field over gene-expression state and computes its structure — on real dynamical-systems mathematics, not metaphor.
A cell type is a reflexive fixed point — a state where the genome's own dynamics come to rest. Morse-type inequalities bound how many such rests the program admits — #stable ≥ 1 + b₁ — and the structure resolves to 384, spread across eight body systems, the nervous system alone holding 86. Not a list someone wrote down; a count the topology forces.
Shows you the sequence — three billion letters, the genes, the variants — as a static text to be searched and annotated.
Treats that sequence as a running dynamical system, solves for the stable states it settles into, and maps the routes between them — the genome's behaviour, computed.
CODEC's output is a re-checkable computational readout — the stable states it found, their basins, the routes between them — derived from a model you can inspect, not asserted. Every number traces back to the dynamics that produced it.
The same dynamics, on screen: your 384 stable identities across eight systems, the regulatory circuits that hold a cell's program in place, and the filtration of possibility — every commitment a door that closes behind you.
CODEC computes structure in the genome's dynamics — nothing more. It does not diagnose, predict disease, or advise treatment, and it makes no clinical claim about any individual. It is a lens for researchers studying how regulatory programs behave. Where the model's assumptions end, CODEC says so — and stops.
The computational biologist who wants the genome's dynamics as a solvable system, not a spreadsheet of reads. The systems-biology lab studying cell-state stability and reprogramming. The researcher who needs every number to trace back to a model they can interrogate.