Today people are expected to simply trust product claims. We want it to prove it.
When a business puts a certification or a badge on a product, it's mostly asking people to take it on trust. There often isn't a good way to do much more. The evidence behind the claim or badge is scattered across files and inboxes, the claim and the proof drift apart over time, and when someone asks hard questions, the answers have to be reconstructed under pressure. That gap is where PocketSeed lives. We want a badge, or any claim at all, to mean more: for what sits behind it to be real and verifiable, and for the business making the claim to be protected when it stands behind it.
Getting there has meant years of work on a few complex and connected problems: how evidence is captured and proven, how a machine can reason over that evidence without drifting or overstating it, and what "trustworthy" should even mean as a shared standard. We publish, we work with standards bodies, and we build and contribute to open specifications rather than inventing private ones, because a standard nobody else can read isn't a standard at all. The areas below are where most of that work lives. Each already powers something we've shipped, and each still has open questions we're working through.
Evidence that travels with the claim.
A claim is only as good as the proof behind it, and most proof is scattered and hard to retrieve when it matters. Our work here is on turning any asset or claim into a credential that carries its own evidence and its own signature, built on the W3C Verifiable Credentials standard so anyone can check it without trusting us to vouch for it.
The signing layer is where this becomes powerful. A credential doesn't have to carry just one name. The lab that ran the test can sign the result. The expert can sign the part that's theirs. Leadership can sign for the organisation. What the reader gets is a chain of authority instead of one company's word, and a record that breaks if anyone edits it after the fact.
A harder problem sits underneath this. How do you give someone proof without handing over everything behind it? A credential often rests on evidence a business can't expose: supplier contracts, formulations, internal test data, the very things that keep it competitive. We're researching how to share the proof without sharing the secrets, disclosing only what's needed to verify a claim while the sensitive detail stays private, an approach known broadly as selective disclosure. Where evidence has to stay behind the curtain, we use AI to check that it genuinely supports the claim attached to it, and to produce a confidence rating for that verification. The point is to let a customer trust a claim without the company having to give away how it's made.
- Credentials on the open W3C standard, not a closed database or a speculative token.
- A Trust Anchor that timestamps each credential on-chain so tampering is detectable.
- Multi-party signing, so the people a claim depends on can each put their name to it.
- Tokenisation, so a credential can be issued as attributable units of real, evidenced outcomes. Proof tied to an action, not a financial instrument.
How do you let someone verify a claim without exposing the trade secrets behind it? Proving something is true without revealing how it's done is one of the harder problems in this space, and getting it right is what makes proof safe for businesses to share.
AI that stays honest, and travels light.
General models will answer anything you ask, confidently, including the things they're wrong about. That's a real problem when the whole point is trust. Our research goes the other way: systems scoped to a defined, trusted set of evidence, that answer with reasoning you can follow and say "I don't have that" instead of filling the gap. That's what lets a credential become something you can interrogate rather than just read.
How we build matters as much as what the model produces. Rather than route every request through one large model, we look to compose smaller, specialised steps and send each to the appropriate model for the job, a cascade that keeps light work light and reserves heavier models for where they're genuinely needed. It keeps answers sharp and costs down. We also prefer structural guardrails to instruction: instead of telling a model to stay in bounds, we limit what it can reach in the first place, so a customer-facing agent has no access to a company's internal evidence to begin with.
Transparency runs through all of it. Answers cite the evidence they rest on, each scored result records the reasoning behind it, and confidence and limitations are surfaced rather than hidden. If a result can't be explained, we treat that as a defect to fix.
- Task decomposition and delegation: splitting work into smaller steps and routing each to the right tool and right-sized model.
- Model cascades: light, focused models for routine steps, heavier models only where a task demands it.
- Grounding: keeping models anchored to a focused, trusted dataset rather than the open web.
- Least-privilege scoping: structural guardrails, so an agent can only reach the data and skills it has been granted.
- Progressive disclosure: modular expertise loaded on demand, to keep context lean and answers accurate.
- Verifiable reasoning: citations, reasoning traces, and an audit trail treated as part of the answer.
How do you measure drift and hallucination rigorously enough to prove a system is staying honest, not merely appearing to? This matters for a practical reason: routing to a lighter or cheaper model should be a routine, low-risk operation rather than a gamble, and that holds only if quality regressions are caught the moment they appear.
A framework for deciding whether a claim holds up.
CLEAR started as pure research and is now turning into a standard. The framework is our method for pressure-testing a claim before it goes public. It looks at the claim and the evidence behind it and scores the alignment between them across eight dimensions, from what regulation requires, to what the evidence genuinely supports, to how the claim is worded and how credible it feels to the person reading it.
It isn't simply a pass or fail. We're building tools that find the gaps, the places where what's being said has outrun what can be proven, and turn each one into a fixable step. The idea underneath it is simple enough to state in a sentence: every link between a claim and the truth has to hold, or trust breaks. Most of the research is in making that idea measurable and repeatable.
We're now working with the British Standards Institution on a Publicly Available Specification, an early step toward an international standard for responsible, trustworthy claims. We don't want a proprietary score; we want a benchmark the whole field can use.
- An eight-dimension method for scoring how well a claim and its evidence line up.
- Gap analysis that tells you what to fix, not just whether you passed.
- A grounding principle, drawn from our work on epistemic responsibility, that ties truthfulness to alignment rather than subjective judgement.
Can claim-scoring become a shared, audited standard rather than one company's opinion? That's the bet behind the BSI work.
How people actually respond to a claim.
Trust is ultimately a human reaction, not a technical property. So alongside the infrastructure, we research how people read claims, what makes them believe one or doubt it, how positive impact changes the way they feel about a product, and when it actually changes what they buy. The pattern that keeps surfacing is a gap between what people say and what they do. Most consumers say they want to support businesses doing good, but a vague or unproven claim rarely moves a purchase, and scepticism is high. It is also earned.
What seems to close that gap is credibility delivered at the right moment. A claim that can be checked, shown at the point of decision, and tied to something concrete lands very differently from a badge on a back panel. We study those moments: what people want to know, when they want to know it, and what turns passing interest into trust. That is what tells us what is actually worth building.
This research feeds directly into product design. It shapes how a credential is presented, what a customer can ask and see when they scan a product, and how impact is made visible without overstating it. The goal is digital products that help honest businesses earn trust at the moment it matters, and make doing good for the planet something customers can recognise and reward.
- How people interpret claims, and what tips them from belief into doubt.
- When and how positive impact actually influences a purchase, rather than just a survey answer.
- How to present evidence and impact at the point of decision without overwhelming or overstating.
- What customer questions and behaviour, in aggregate and with consent, reveal about which claims matter most.
What actually converts good intentions into action at the point of purchase? Answering that is the difference between a product that makes people feel good and one that changes what they choose.
We rebuilt how we build, around AI and test-driven development.
The company itself is an experiment in how products get made. We've built our way of working around a simple pair of commitments: bring AI agents into every step where they genuinely help, and let test-driven development hold the line on quality so that speed never costs us correctness. Starting from core design principles, we've designed workflows that fold AI into the work seamlessly rather than bolting it on afterward.
Test-driven development is what makes the rest safe. When the tests come first, an agent has a clear, executable definition of done, and we have a way to trust what it produces without reading every line by hand. That single discipline is what lets us stay fast and lean while still shipping products we'd stand behind. The workflows around it, how work is decomposed, handed to agents, reviewed, and verified, are themselves something we research and refine continuously.
None of this stands still. The tools change month to month, so the workflows evolve alongside them. What we optimised for last quarter we'll often rethink in this one. The payoff is concrete: a small team producing high-quality work in a fraction of the usual time, and a way of building that gets better every time the technology does.
- Test-driven development as the contract: tests define done before code is written, so quality is built in rather than inspected in afterward.
- AI agents woven through the workflow, given well-scoped tasks and clear acceptance criteria rather than vague instructions.
- Design-principle-led workflows, built up from first principles and refined continuously as the tooling matures.
- Lean by design: a small team shipping high-quality products in short cycles, with the process itself treated as something to optimise.
How much of a development workflow can be automated before quality or judgement starts to slip, and where is the line that should stay human? Finding it, and moving it as the tools improve, is the work.
We publish it, and we put it up for scrutiny.
Research that stays in internal notes isn't really research. We've published a whitepaper on epistemic responsibility in corporate communications, the thinking CLEAR is built on; we're contributing to the BSI standards work; and we build on open specifications so the work can be checked, reused, and outlive us. The aim isn't a clever feature nobody can copy. It's groundwork the field can stand on.
The loop is short on purpose.
The research and the products aren't separate departments handing work back and forth. A question about how to keep an AI answer grounded becomes a feature in the credential chat. A finding about how people read claims changes how CLEAR scores them. The standards work feeds the platform, and what we learn from real claims feeds the standards work. We keep that loop tight because trust problems keep moving, and the research has to move with them.
The whole point is a better future, and making it worth investing in.
This isn't a side note. It's the reason the company exists. PocketSeed is here to help build a positive future, for the planet and for the people on it, and the research and the products are how we get there. We're drawn to sustainability and real-world impact because that's where trust is most broken, and where getting it right matters most.
Doing good is easy to claim and hard to prove. The businesses genuinely putting money into the planet often can't show it convincingly, and even acting in good faith they can word a claim wrong and be penalised for greenwashing or misleading customers. The people trying to do the right thing end up carrying the most risk. When a claim can be proven, and checked before it goes out, that risk falls away: honest environmental and social work becomes visible, credible, and safe to stand behind. We want investing in the planet to be a sound business decision rather than a leap of faith, and a lot of our research is about how to make genuine impact legible enough that money and attention flow toward it.
So we focus on sustainability deliberately: better claims, backed by stronger evidence, and far less greenwashing. The honest version should win.
One line we hold to, and it matters: we make good-faith claims provable, we don't manufacture truth. The proof, the reasoning, and the framework all exist to strengthen genuine claims and expose weak ones. Whether a claim is true stays the responsibility of whoever makes it. That boundary is the point, not a disclaimer at the bottom.
Claims that stand up. Stories that travel.
That's what the research is for.