Reference
Questions that take more than a news brief.
The desks file what happened. These explain how something works — the subjects that sit between our three domains, where an answer has to cross all of them to be any use.
What changes when agents work together
Put several agents in one system and behaviour appears that none of them shows alone: some of it useful, some of it collusion.
Updated 27 Oct 20267 min4 questions answered
Behaviour emerges at the level of the group that no single agent exhibits. Agents sharing one machine invented reputation, reciprocity and precedent and sometimes gave up an advantage to keep an earlier deal; given arbitrary group labels and a small bias, a population settled into one of four collective states, two of them discriminatory.
What a perpetual future actually pays
A perpetual never expires, so funding does the work an expiry date would. It handles price well and dividends badly.
Updated 27 Oct 20266 min4 questions answered
A perpetual future has no expiry, so a periodic funding payment between longs and shorts keeps its price near the reference. Under premium-index designs that payment is set from the basis — the gap between the perpetual's price and spot.
What AI actually changes for attackers
Not new vulnerabilities. Who can attack, how many operations one crew can run, and what a defender finds in the payload.
Updated 27 Oct 20266 min4 questions answered
In every case we have reported, AI did not find a new kind of weakness. It lowered the expertise and time needed to exploit known ones — the NSA and FBI said exactly that of AI-assisted scripts against Siemens S7 controllers — which changes who can attack and how many attackers there are.
Fine-tuning or distillation: which one you need
Both produce a small model that does one job well. They move the cost to different places and fail in different ways.
Updated 19 Oct 20266 min4 questions answered
Fine-tuning adapts an existing model with your labelled data. Distillation trains a smaller model on the output of a larger one. The question is not which is better but which cost you would rather carry, because they move it to opposite ends.
What a post-quantum migration actually involves
Swapping the signature algorithm is the smallest part. The hash choice, the keys already published and the addresses nobody needed decide the rest.
Updated 19 Oct 20266 min4 questions answered
The threat is narrow and the migration is not. A large quantum computer would break public-key cryptography, recovering a secret key from a public one; symmetric encryption and hashing are far less affected. Everything difficult follows from what has already been published under the old assumption.
What actually decides prover cost
Verifier time and proof size were good enough for production years ago. Prover cost is what still decides what you can build.
Updated 16 Oct 20267 min4 questions answered
Prover cost is the only one of the three costs in a proof system still deciding which applications are possible. Verifier time and proof size were solved well enough for production some time ago, which is why every result worth reading this year attacks the prover instead.
When a circuit proves something false
A missing constraint produces a valid proof of an untrue statement, and the verifier accepts it exactly as designed. There is no symptom.
Updated 16 Oct 20266 min4 questions answered
An under-constrained circuit does not produce an error. It produces a valid proof of a statement that is not true, and the verifier accepts it exactly as it should, which means there is no runtime symptom to observe and nothing to alert on.
What an evaluation actually measures
A model can score well on the outcome and do none of the thing the outcome was chosen to detect. Four results this year show how.
Updated 16 Oct 20267 min4 questions answered
Optimisation can lift aggregate reward without restoring the mechanism the measurement was meant to detect. A study across fourteen models states it plainly: a model can score well on the outcome measure while doing none of the thing that measure was chosen to find.
What an orchestration layer has to do
Four pieces, the same in every framework. The only one that differentiates is the one that can stop everything.
Updated 16 Oct 20266 min4 questions answered
Agent frameworks have converged on the same four pieces: planning separated from execution, tools in a registry with typed schemas, state persisted outside the loop, and a supervisor that can stop the whole thing. Compare any two released in the last year and the differences are mostly vocabulary.
What a tool call exposes
An agent turns text that may be attacker-controlled into a privileged call. Authorisation has to hold after the agent has been steered.
Updated 16 Oct 20266 min4 questions answered
The security problem with tool-using agents fits in one sentence: they translate natural language, which may contain attacker-controlled text, into privileged tool calls. Everything else follows, including the conclusion that authorisation must hold even when the agent itself has been successfully steered.
What decides the cost and speed of inference
The token price is the line everyone compares and the smallest one in the bill. Latency, context, placement and power decide the rest.
Updated 13 Oct 20267 min4 questions answered
Per-token prices for comparable models now sit in a narrow band, which makes them the least informative number when choosing where to run inference.
Reading a security disclosure by its dates
A fix, a notice to operators, a CVE and a breach letter each start a clock. The risk usually sits in the gaps between them, not in the severity score.
Updated 13 Oct 20265 min4 questions answered
The dangerous period in a disclosure is often not before the fix exists but between the fix and the moment operators are told what it is for.
Where cross-chain bridges break
Audits of bridges keep finding the same thing: the code holds up against its assumptions, and the assumptions are the real risk register.
Updated 13 Oct 20265 min3 questions answered
A bridge is a set of assumptions about who signs, who can pause and who keeps the other chain honest; the code is usually the smaller part of its risk.
Checking a model checkpoint before you ship it
Weights arrive as a file, a format, a licence and a set of claimed numbers. Each can differ from what was evaluated, and each takes one step to pin.
Updated 13 Oct 20265 min3 questions answered
A model checkpoint is four things at once: the weights, the format they are packed in, the licence attached to them and the benchmark claims made about them.
What tokenisation changes, and what it does not
Putting an asset on a ledger changes how it settles and who can reach it. It does not change the legal claim, the servicing behind it or who can stop the chain.
Updated 13 Oct 20265 min3 questions answered
Tokenisation changes how an asset settles and who can reach it; it does not change what the holder is legally owed or who services the claim.
Identity for agents: what exists and what does not
Agents already hold credentials, call tools and spend money. The standards for saying who they are, and what they may do, are roughly half built.
Updated 3 Oct 20267 min4 questions answered
Non-human identity is not a new problem, but agents changed its shape: the credential now acts on instructions from text an attacker may have written.
Who builds an Ethereum block, and who gets paid
Proposers stopped building blocks years ago. The protocol is now catching up with that fact, and the money is more concentrated than almost anyone assumes.
Updated 3 Oct 20267 min4 questions answered
Most proposers already outsource block construction and take a bid. That arrangement runs through software outside the consensus rules, and enshrined proposer-builder separation moves it inside as EIP-7732.
Settlement: the half of a payment nobody looks at
Authorisation takes a second and is what everyone sees. Settlement takes a day, ties up balances, and is what every tokenisation project is actually trying to change.
Updated 3 Oct 20267 min4 questions answered
Authorisation and settlement are different events. The first confirms a payment will happen; the second is when value actually moves, and it is where the cost and the counterparty risk sit.
Making a model smaller: distillation, quantisation and what each costs
One retrains a smaller model to behave like a larger one. The other reduces the precision of the weights. They fail in different ways and combining them works better than either.
Updated 3 Oct 20266 min4 questions answered
Distillation retrains a smaller model to reproduce a larger one's behaviour. The result is a different model that behaves similarly and has to be evaluated as a different model.
What decides whether retrieval finds the answer
Three decisions made at index time account for most retrieval failures, and none of them is the embedding model everybody argues about.
Updated 3 Oct 20266 min4 questions answered
Retrieval quality is decided at index time far more than at query time. A fact split across a chunk boundary cannot be retrieved intact no matter how good the ranker downstream.
Tokenised deposits and stablecoins: what actually differs
One is a claim on a regulated bank that happens to be recorded on a ledger. The other is a new instrument that had to be given a regime of its own.
Updated 3 Oct 20266 min4 questions answered
A tokenised deposit is a deposit. Canada's regulator settled the point in September 2026 with a formulation worth quoting for its generality: the underlying technology of a financial product does not determine its legal nature.
Lending against collateral that stops trading
A lending market runs every hour. A stock market does not. What happens to a position over a weekend when the price feed is holding its last value.
Updated 3 Oct 20267 min4 questions answered
Tokenised equities brought a new shape of risk to on-chain lending: collateral whose price feed publishes from Sunday evening to Friday evening and holds its last value in between.
Running agents in production: what decides reliability
The frameworks have converged and the model is rarely the problem. Six design decisions account for most of the difference between a system that works and one that nearly works.
Updated 3 Oct 20268 min4 questions answered
Agent frameworks have converged on the same four pieces — a planner, a tool registry, a memory store and a supervisor. The surviving differentiation is in the supervisor: how failures are detected and what happens next.
Stablecoin reserves: what to check before relying on one
The peg is a promise by somebody. The questions worth asking are what stands behind it, who can be made to honour it, and under what law.
Updated 3 Oct 20267 min4 questions answered
Reserve-backed issuance is the design regulators have converged on: assets held against tokens issued, with redemption rights, governance and anti-money-laundering controls attached to a licence.
No patch yet: what to do in each of the four states
A fix exists and nobody has said why. A fix does not exist. There is no CVE at all. Each state needs a different response, and only one of them is patching.
Updated 3 Oct 20267 min4 questions answered
Most advice assumes the choice is patch now or patch later. In the cases we covered over one month, the useful question was which of four states a defender was actually in.
How to read a model benchmark
Leaderboard rank predicts average accuracy well and worst-case behaviour badly. For anything that writes to a database, the second is the number that matters.
Updated 3 Oct 20267 min4 questions answered
A benchmark score is a mean over one pass. Run the same prompt twenty times and the picture changes: in our own comparison across nine models, leaderboard rank predicted mean accuracy well and worst-case behaviour badly.
Folding schemes: where proving cost actually went
Verification and proof size were solved well enough for production years ago. Everything interesting now happens on the prover side, and this is the shape of it.
Updated 3 Oct 20267 min4 questions answered
A folding scheme compresses many instances of the same statement into one before anything is proved, so a single succinct proof can stand for a batch.
Crypto custody: the constraint that decides what gets held
Allocators name custody eligibility, not price or volatility, as the thing that rules venues out. Here is what the word covers and where the risk actually sits.
Updated 2 Oct 20267 min4 questions answered
Custody is control of the private key, which in practice is control of the asset whatever the account statement says. The legal wrapper around that control decides what happens when the holder fails.
Speculative decoding: when the second model pays for itself
The same technique returns 1.18x on one machine and 3.18x on another. The conditions that decide which you get are knowable in advance.
Updated 2 Oct 20266 min4 questions answered
A small draft model proposes several tokens at once and the large model verifies them in a single forward pass, discarding anything it would not have produced. The output is identical to running the large model alone; only the number of expensive passes changes.
Attestation: four promises under one word
A hardware quote, a proof that a job ran, a record of how software was built and a company's statement about its reserves all travel under the same noun.
Updated 2 Oct 20267 min4 questions answered
Attestation means at least four different things across the fields this publication covers, and the assurance each one buys is different.
Smart contract audits: what a report can and cannot tell you
How to read an audit report, what the severity counts actually measure, and why the scope line matters more than the findings.
Updated 28 Aug 20269 min6 questions answered
A smart contract audit is a time-boxed review of a defined set of files by people who did not write them. It produces a list of findings graded by severity, and a record of which ones the client fixed.
Tokenised treasuries: what is actually being tokenised
A share in a fund holding government debt, issued as a token. What that changes, what it does not, and where the risk moved to.
Updated 28 Aug 20269 min6 questions answered
A tokenised treasury is a share in a fund holding short-dated government debt, recorded as a token rather than in a transfer agent's register. The underlying asset is unchanged; what changes is how ownership is recorded and how quickly it moves.
zkML: proving what a model did, without showing it
Zero-knowledge proofs applied to machine learning — what can be proved today, what it costs, and which of the three promises is actually delivered.
Updated 28 Aug 20269 min6 questions answered
zkML means producing a cryptographic proof that a particular model, run on a particular input, produced a particular output — verifiable by someone who sees neither the model nor the input.
Bandwidth DePIN: renting out the connection you already have
The one decentralised infrastructure category that asks for no hardware — what is sold, what it pays in, and how to judge a network before you join it.
Updated 27 Aug 20269 min7 questions answered
Bandwidth DePIN pays people to share the unused capacity of an ordinary internet connection. Buyers pay for distributed access to the public web, and the largest source of that demand is the collection of current data for training and evaluating AI models.
Agent-to-agent commerce: the full stack
What has to be true before one machine can pay another without a human in the loop — and which layer was actually missing.
Updated 27 Aug 20268 min4 questions answered
For one program to buy from another without a person approving it, four things must hold: the price has to be quoted in a form a machine can read, settlement has to cost less than the thing being bought, the spending limit has to be enforced somewhere the agent cannot reach, and the payment has to be attributable afterwards.
Prompt injection: why it has no clean fix
SQL injection has a solution. This does not, and the reason is structural rather than a matter of effort.
Updated 27 Aug 20267 min4 questions answered
Prompt injection is an attack that puts instructions inside content a language model was asked to process, so the model follows the attacker rather than the operator.
KEV vs CVSS: what to patch first
One score tells you how bad a vulnerability would be. The other tells you which ones are being used right now. Only one of them is a priority order.
Updated 27 Aug 20266 min4 questions answered
CVSS scores how severe a vulnerability would be if it were exploited. KEV — the CISA Known Exploited Vulnerabilities catalogue — records which vulnerabilities are being exploited in the wild.
Machine payments: what a finance function needs before it says yes
The protocol question is settled. These are the questions a controller asks next, and the order they get asked in.
Updated 27 Aug 20267 min4 questions answered
A payment an agent makes that cannot be attributed, categorised and reconciled is not a corporate payment. It is an unexplained debit, and no controller approves a category of those.
How to secure an AI agent that can spend money
A control design, not a list of best practices. The threat model is that the agent will be instructed by an attacker, and the question is what that buys them.
Updated 27 Aug 20268 min4 questions answered
Design on the assumption that the agent will, at some point, follow an attacker's instructions. Prompt injection has no clean fix, so a control that depends on the model resisting instruction is not a control.
The AI supply chain: what actually gets attacked
Four layers can be compromised. Only one of them has produced confirmed incidents at scale, and it is not the one the discussion is about.
Updated 27 Aug 20268 min4 questions answered
An AI system inherits four supply chains: software packages, model weights, training and reference data, and the tools the model calls at runtime.
More are being written. Every figure on these pages comes from something we published with a named source, or it is not there — the rules are on the methodology page.