Research begins with a question we cannot yet answer well.
Six programs, each built around open questions. We publish what we find, including what did not work, and we say how sure we are.
PrinciplesThese are the rules the lab holds itself to. Each one came from a mistake we made or watched someone make.
How we work
- A claim names its evidence.We say whether something was seen, derived or guessed, and we never let a guess wear the clothes of a measurement.
- A failure stays in the count.When a check cannot be run, it is reported as not run. A number that improves because the hard cases were dropped is not an improvement.
- Authority follows reach.What an action touches in the world decides who must approve it, not what the action is called.
- The simplest reliable method first.Plain code before a script, a script before a model, a local model before a remote one. We move to the next step only with a reason we can write down.
- Nothing is lost by accident.Knowledge leaves a system only through a door with a name on it: removed on purpose, replaced with a record of what came before, or archived where it can be found.
- People decide. Software keeps the promise.Our tools propose and carry out. The decisions that matter stay with the people accountable for them.
The programs
Program 1Running in our own operations since early 2026.
Governed agency
AI systems are moving from answering to acting. We study how to grant them real work without granting them the decisions that belong to people.
- Where exactly does a suggestion become an action, and can software find that line on its own?
- How should approval work when an AI performs hundreds of small actions an hour?
- When a person approves once, what does that approval cover, and when does it expire?
Program 2Instruments built; calibration ongoing.
Evidence and honesty
Language models are fluent whether or not they are right. We build ways to see the difference from the outside.
- Can a system detect that it is filling a gap with something it imagined, before it states it?
- What should a report do with the cases it could not check?
- How do you know a detector works if it has never fired?
Program 3Our lab runs much of its work on its own hardware.
Local-first intelligence
A business that sends everything to a remote model is renting its judgment. We test what can run on hardware it owns.
- Which tasks need a frontier model, and which only seem to?
- How should a shared machine divide itself between a person working and AI working in the background?
- What is lost, in quality and in cost, when the work stays home?
Program 4In daily use; the open questions are about scale.
Memory and continuity
Work spans days and people. An assistant that forgets each session cannot carry it. One that remembers everything cannot be trusted with it.
- What should an assistant remember, and who decides when it forgets?
- How does a belief that was true last month get rechecked before it is relied on again?
- How do separate parts of an organization share memory without sharing what each must keep private?
Program 5Applied through Method ML's client work.
AI and the work of firms
Professional firms run on numbers that have to agree: hours and invoices, the ledger and the bank, the plan and what was done. That makes the back office the honest test of whether AI helps.
- Which back-office work can AI take on today without anyone checking it twice?
- How do you move an old system forward while the firm keeps running on it?
- What does a firm need to see before it trusts a number software produced?
Program 6Measurement under way; no conclusions published yet.
The economics of machine labor
The price of a token says little about the cost of a finished, verified piece of work. We measure the second and report it beside its denominator.
- What does verified work actually cost once retries, review and human attention are counted?
- Which ratios can be made to look better simply by doing more of the wrong thing?
- When is a person still the cheaper and better choice?
Working with the lab
To discuss the research or a business licence for IntentOps, write to hello@rndbi.com.