Research
An agent is only as reliable as the context it acts on. This is a program of falsifiable measurements on one property: that the information an autonomous agent communicates, stores, resolves, and acts on stays reliable across every stage.
Yao Tsakpo · Independent research
A long-lived agent passes every fact through four stages. Each is a place the fact’s integrity can be lost, and each is a place it can be measured. We use agent context integrity as the organizing term for the whole property, and treat each stage as a separate, testable component rather than a claim to have founded a field.
| stage | component measurement | status |
|---|---|---|
| Communicate | Pointer vs payload: reference-based context acquisition | measured |
| Store / resolve | Provenance-tiered conflict resolution in agent memory | this paper |
| Verify | Verification integrity (faithfulness of acted-on context) | forthcoming |
| Act | Execution integrity: a boundary between claim and action | measured |
Store / resolve · 2026
Mainstream agent memory keeps recency and relevance but discards the one signal that decides which conflicting fact to trust: where it came from. Attaching a discrete, write-time, unforgeable provenance tier and resolving conflicts by tier measurably improves accuracy on conflicting facts. Two closed-form models, an exact simulation, and a five-model empirical panel.
Read the paper→Store / resolve · reproducible walkthrough
Same question, same conflicting facts, two memory stores. As conflict density rises a flat store's accuracy decays toward chance while a provenance store holds at 100%. The hands-on companion to the provenance paper: the derived models, the measured confirmation, and every step needed to reproduce it.
Run the walkthrough→Act · 2026
An agent that acts turns a claim into a consequence, and the seam between the two is owned by nobody: the judgement is the authority. Measured on a physical sorting task, where a model's uncertain verdict moves a part into a customer's bin as confidently as a certain one. Adding a boundary converts those errors into deferrals, and the whole tradeoff curve is reported rather than one flattering point.
Read the measurement→Communicate · 2026
Sending an agent a resolvable reference rather than a raw payload lets the receiver reconstruct richer, verified context from far fewer transmitted tokens. The communication-stage measurement of the same program.
Read the measurement→Verification is the natural remaining component. Framing them under one property is what lets each measurement reinforce the others rather than stand as an isolated engineering note: the act-stage measurement was taken on a physical task with no email in it, which is what distinguishes a property from a feature of the place it was first noticed.
The measurements are implementation-independent: each is reproducible over any store, in well under a hundred lines, with any inference endpoint. Where a running system is needed to verify a claim end to end, the work uses Inbin as the reference implementation, an event store whose write surfaces separate an authenticated owner channel from an agent channel. Inbin is one instantiation of the requirements, not a dependency of any result.