One effect per request
One request, one effect.
The Geometer vision
More room to act.
A reason to trust.
Our mission is to make consequential automation accountable while preserving freedom within explicit constraints.
A declared envelope. Several valid paths.Conceptual analogy
Declared operating envelope
Route A · actor chooses the path
Try the boundaries
Route A · permitted in this illustration
The actor chooses a route inside the same declared envelope. The boundary does not prescribe every step.
Inspect the illustrative record
- Constraint
- Keep proposed commands within the declared boundary.
- Proposal
- Follow permissible route A
- Decision
- Route A · permitted in this illustration
- Observed outcome
- Illustrated route only
- Assumptions
- Declared boundary and complete observation
- Limits
- Illustration, no real-world safety conclusion.
An explanatory analogy, not an aircraft safety model. It does not claim every real failure can be blocked. Without animation: either route can be permitted, advice can be ignored, changed assumptions require review, and missing evidence means unknown.
01 /Four markets today
The stakes go beyond
the task.
Games and live worlds
Restore lost items, reward quests, answer world events, hand out promo codes.
One restoration per ticket
Commerce and support
Refund orders, compensate tickets, change bookings.
Refund at most once
Money movement
Pay creators, disburse loans, pay claims, buy within a limit.
A claim is paid at most once
Operations
Grant roles, run lab protocol steps.
One grant per request
Research horizon, not a product
- Health operations
- Industrial equipment
- Energy and infrastructure
Longer-term directions beyond the four markets. Health operations would need their own domain owners and validation first. The physical-world idea
02 /Why accountability enables ambition
Not the brake.
The instrument.
The marine chronometer
A clock that widened the map.
It explored nothing itself. It made far more ambitious voyages possible to attempt, and to return from.
03 /Connected responsibilities
From intention to evidence.
- 1DefineWhat must hold?
- 2ExploreWhere could it fail?
- 3GuardMay this proceed?
- 4ProveWhat happened?
- 5AdaptWhat changed?
- 6DecideWhat next? Advice only.
04 /Broader assurance
Understand the decision.
Test the whole system.
Human oversight
Make review meaningful.
Test oversight under realistic workload.
How action approval works →Future research · Explainable AI
Make explanations useful.
Tested for faithfulness before we rely on it.
Read NIST’s explainability principles →Background: NIST AI Risk Management Framework.
05 /Progress you can inspect
Show the work.
Public, reproducible work comes next.
| Work | What you can inspect today | What comes next |
|---|---|---|
| Refund experiment | A deterministic reproduction shows two effects after a blind retry and one after reconciliation in the synthetic case. | Apply the method to authorized open-source workflows and publish the problems. |
| Evidence integrity | A local verifier accepts the intact synthetic evidence pack and rejects modified evidence against the supplied test inputs. | Complete assessment-certificate issuance and lifecycle verification. |
| Public audits | The illustrative certificate shows how scope, problems and exclusions will be presented. | Publish real case studies with their own version-specific certificates once supported. |
| Product release | Local Explore, Guard (in development and staging) and Prove in a free alpha. | Validate the improvement in real workflows. |
Earned authority
Give your agents more to do.
One rule at a time.
Start with one rule. People decide when to grant more.
Refund up to $50
At level 4. Ready for a person to grant level 5.
Future direction
Helping machines act
in the real world.
Our product starts with software agents. Over time, we want to explore how the same ideas could help machines understand their surroundings, plan useful work and act within clear limits.
Robotics makes it tangible: systems that recognize when they need help and remain accountable to the people who rely on them.
The longer view Long-term research directionRoom to actConcept 1
Build a picture of the world.
What is known? What is uncertain? A useful agent needs to recognize both before it chooses a next step.