How we raise reliability
No agent is trusted. Every output is checked.
This site never says an agent can be trusted. It says how reliability is raised, in three ways: what an agent does is traced, what it is told and runs on is tested, and every output is checked. It is not a promise that outputs are correct. It is a promise about how a wrong output gets found.
A test set is a fixed list of test cases that an agent's instructions or a model must pass before they may be used for a role.
Three legs
-
Traced
Everything an agent does is recorded, and can be read back.
Every model call is recorded: project, role, model, tokens and cost. Measured You can read who said what to whom in the team. Measured
Not yet: reading them in plain words in a workspace of your own. Planned, not built Traces hold full words by default, which is a risk to your data too (issues #51, #84).
-
Tested
No instruction or model change is used before it passes its test set.
The gate exists, and it is a command a person runs. Measured; gate manual
Planned: every standard agent ships its own test set, and the system enforces the gate. Open models are measured against a top commercial model on the same test set. Planned, not built
-
Checked
AI writes the checks, the checks run without AI, and a person approves.
A person's approval, and a separate reviewer who cannot change the work, are in place. Measured, with a recorded deviation
Planned: where a rule can be checked by code it is checked by code, and every fault found becomes a test. Planned, not built
What you would see
- Every task on one list, with who or what is doing it, what it touches and what it waits on.
- A trace you can open: when the builder started, when the change was opened, who answered, who is waited for.
- The cost of each model call, and the clashes that were refused.
The drawing shows where this is going. Part of it exists today in a first form.
The honest limit
- A person runs the test-set gate today. The system does not yet enforce it (issue #115).
- Nobody has shown that a reviewer catches a real fault (issue #28). The reviewer is another agent, so it is checked too.
- All reviewers use one model family (issue #82).
- The prompt-injection suite measures. It does not certify.
The technical detail
The claims behind this page, with their states as the evidence page records them.
| No. | Claim | State | What it serves | Check or issue |
|---|---|---|---|---|
| 8 | Every model call is recorded: project, role, model, tokens, cost. | Measured | Every call traced (F6) | trace_checks |
| 9 | You can read who said what to whom in the team. | Measured | Full team tracing (F18) | team_trace_checks (6 of 6); #84, #159 |
| 10 | A separate reviewer checks work before you see it. Nobody has shown that a reviewer catches a real fault. | Measured (pair runs) | Independent review (F13) | reviewed_checks; #28 |
| 13 | Nothing is published or merged without a person. No self-approval. The recorded deviation: whoever holds a project's key (Temporal token) to the engine's workflow service can send a decision under any name. | Measured, with a recorded deviation | Human approval (F9, S9) | inbox_checks; #32 |
| 20 | A test set must pass before a model or instruction change is used. The gate is a person's command. | Measured; gate manual | Test-set gate (F5) | golden_checks; #115 |
| 60 | A role moves to an open model only after a test-set run against a top commercial model. The nearest built part is the test-set gate (claim 20). | Not built | Test sets (F5), open against commercial | Decided 2026-10-13. Not built. #314 |
| 69 | Each standard agent ships a test set. The system, not a person, enforces the gate. | Not built | Test sets (F5) | #115; #316 |
| 70 | Code checks what code can. AI-written checks run without AI. Each fault becomes a test. | Not built | Checks by code (F29) | #296 |
| 71 | You can read each agent's test results and traces in plain words. | Not built | The organisation's product owner's workspace (F17) | #113, #154 |
Reliability on each question
-
Question 1
Build what actually matters
No agent is trusted, so every change is checked before you see it: a separate reviewer looks at it, and only you accept it (claims 10, 11 and 13).
Read the answer -
Question 2
Choose your models, keep your freedom
This is the "tested" leg: a model or an instruction change is used only after it passes its test set, and open models are measured against a top commercial one (claims 20 and 60).
Read the answer -
Question 3
People and AI, building together
This page is the "checked" leg: AI writes the checks, the checks run without AI, and a person approves (claims 13 and 70). The reviewer is another agent, so it is checked too: nobody has shown yet that it catches a real fault (claim 10, issue #28). Each standard agent is also meant to ship its own test set (claim 69).
Read the answer -
Question 4
Use AI and stay compliant
The "traced" leg is the base of the records: every model call is recorded and you can read who said what to whom (claims 8 and 9). Traces hold full words by default, so they are a risk to your data as well as a record (claim 25).
Read the answer -
Question 5
Fit your people and your budget
Spending is checked before it happens: a call that would pass your limit is refused (claim 5), and the cost of each run is recorded (claim 8).
Read the answer
Tell us you are interested
B Team is not offered to anyone yet. NGOs come first. If you want to hear when that changes, or you would like to be one of our first pilots, write to us.
Write to: [contact address to be decided]
This draft has no form and collects no data. Once there is an address, what happens next is meant to be this: we tell you whether we can work with your organisation yet. We promise no date and no price.