Somewhere in the last two years, every firm on earth became an “AI-native company” by buying the same subscriptions as everyone else, and we think that is a polite fiction. If AI is going to sit inside serious security work, then the firm doing that work has to practise AI the way it practises security: hands on the machinery, findings written down, mistakes owned in public. So we launched our own research lab. This page is what it works on, what it publishes, and what it is building next.
Everything in this library clears a twenty-five point standard before it ships, from statistical power to hand-checked citations, and every finding carries an evidence class so you can see what it rests on. We publish the standard itself, and we would frankly enjoy being held to it.
The Birchlogic Research Standard · publishing with the first entryDoes safety survive compression?
Every regulated deployment quantises its models to fit real hardware, and almost nobody has measured what that does to the safety behaviours the full-precision model was aligned with. We extracted a refusal direction once at full precision and applied it unchanged across int8 and NF4 arms, and the finding so far is uncomfortable in an interesting way: the geometry drifts under NF4 harder than int8, and the behaviour transfers anyway.
Status · Paper in preparation, independently reverified inside the lab before anything ships.
What does an on-premises security model actually cost?
The sovereign-deployment story is sold hard and measured badly, including, at first, by us. Our own first measurement pass failed our own review, so the work is being re-run under the provenance protocol, and we would rather publish late and right than early and wrong.
Status · Re-running. No numbers until it clears.
- 2 reserved
Two further programmes are being scoped now, and they will appear here when they are real.
Everything above feeds one machine.
The agents do the work that eats a security team alive: evidence collection across your stack, policy sets that stay current instead of going stale between audits, control monitoring that notices drift the week it happens rather than the week the auditor does, and first drafts of the two hundred questionnaire answers every enterprise deal now demands. That work runs continuously, which is what makes a serious programme affordable at a price where our competitors can only offer you a few hours of somebody’s month.
What the agents will never do is decide what matters. Scope, judgment, the conversation with the auditor when they push back, the call on which of forty red flags is the one that counts, and the signature on anything that leaves the building: that stays with a person whose name is on it, in every engagement, with no exceptions, because we have read enough machine-generated compliance work to know exactly what it is worth without a senior reader.
And so you have it from us rather than from a sales call: the workbench does not make junior work senior. It removes the volume work so that senior time goes where senior time matters. A firm telling you its AI replaces judgment is telling you something about its judgment.
Enterprises are pouring serious money into having their own AI, and at the end of it most of them are running the same model, trained on the same data, through the same harness as everyone else. The capability is real and it is everywhere, which is exactly the problem: when every firm’s outputs come from the same intelligence, no firm’s outputs sound like the firm. There is no individuality left in what comes out, no self, and for a professional organisation whose entire value is how its people think, that should be alarming.
Because what makes your firm’s work yours was never the model. It is the judgment of your senior people, the way your firm argues, the precedents it reaches for, what it refuses to say, the know-how that walks out of the building every time an expert retires. No amount of spend on your own AI captures any of that today.
So we are building a memory layer: a system that captures expert and organisational knowledge and sits in front of any model you choose, frontier or your own, so that the outputs carry your firm’s identity at your experts’ level. Not another model. A memory that makes whichever model you use produce work that could only have come from you.
We think that matters more than owning your own AI, in a world where intelligence itself is being commoditised. The intelligence is becoming a utility. The identity is becoming the asset.
It grew out of our work on judgment in regulated decision functions, and it is in build now.
The lab works out of Delhi and Singapore, and the second of those is not an administrative fact. Singapore is where the rules for regulated AI are being written fastest right now, with MAS drafting AI risk guidelines, the MindForge handbook already published, and IMDA circling agentic systems, and we would rather build governance-grade infrastructure next door to the people writing the rulebook than read about it from a distance. If your regulator is going to ask hard questions about AI in the next two years, there is a fair chance the question gets drafted within a kilometre of our Singapore office.
- Delhi5th Floor, Statesman House, Connaught Place, New Delhi 110001
- Singapore1 Scotts Road, #24-10, Shaw Centre, Singapore 228208