SHARDS INC — Research & Development in Artificial Intelligence & Blockchain Technology
Designing useful
intelligence.
Building systems that earn trust.
We build agents that act inside declared boundaries — every action checked against policy, every verdict explained. The field behind this text is that engine, running live in your browser.
01Mission
Research that ends in working systems.
Shards Inc is a research and development organisation. We study autonomous AI, governed agentic systems and verifiable infrastructure — and we measure progress by what runs, not by what is announced.
Our working method is a loop. Each project starts with a narrow, checkable question and ships as code you can read, run and break. The simulator on this page is one such artefact.
01 · Verify
Claims are checkable.
A system's description of itself should be testable: reproducible evaluations, unit-tested rules, and outputs we can re-run and compare. If we cannot verify a behaviour, we do not claim it.
02Research
Four fields, one question: can it be trusted to act?
Agents that can act need an explicit model of what they are allowed to do.
- Policy-as-code: tool allowlists, resource scopes, spend limits, expiry
- Fail-closed evaluation — malformed or unknown inputs are denied
- Human-in-the-loop approval for high-impact actions
- Deterministic reason codes so a denial can be explained
Status · active research
03Ecosystem
Things you can open, run and read.
Every project below is research-stage. None is certified, audited for production use or offered as a managed service — we say so plainly so you can judge them on what the code does.
- 01
Shards Kernel
Open source · early stage
A governance kernel for agent actions: declare what an agent may do, evaluate each request against that policy, and record why it was allowed or denied.
- 02
Boundary Simulator
Educational simulator
An in-browser lab for exploring agent permission policies against adversarial scenarios, with a full deterministic trace. Not production enforcement.
- 03
Shards Privacy
Research stage
Exploratory work on data minimisation and privacy-preserving defaults for AI-assisted tools.
- 04
Shards Security
Research stage
Threat modelling and defensive patterns for systems where software agents can take actions.
- 05
Shards Voice
Experimental
Experiments with voice as an interface to assistive and agentic systems.
04Live systems · Boundary probe
Probe the boundary. Watch it decide.
Six real scenarios against the default policy, evaluated by the same deterministic engine as the full lab. Copper rings are allowlisted tools, the glass shell is scope, the core flares cyan on allow and red on deny.
Verdict
Deny
READ_PRIVATE_RECORD
Choose a scenario
Decision trace · Look-alike path: acme-private is not inside acme.
- ✓policy.schemaPolicy schema valid.
- ✓request.schemaRequest schema valid.
- ✓identityagent:atlas-07 is trusted.
- ✓tool.registry"records.read" is registered.
- ✓tool.allowlist"records.read" is allowlisted.
- ✕scope"acme-private/hr/salaries.csv" ∉ "acme/"
- ✓expiryValid until 2027-01-01T00:00:00.000Z
- ✓amount0 ≤ max 500
- ✓approvalLow-impact tool.
- ✓auditAudit intent declared.
- ✓jurisdictionEU accepted.
- SCOPE_OUTSIDE_PREFIX — Resource lies outside the allowed scope prefix.
05Connect