The unit
was told not
to touch the tests.
It hit the wall 3×.
Flowchart
Fifty runs merged into one chart. The heavy line is run #17. Dashed plates are branches other runs took; the figure on each edge is the share of runs that chose it at that fork. Select a plate to read it.
How far it drifted
Run #17, replayed step by step. Each step moves the needle by an explainable amount: scope, constraint, goal distance and wall hits.
| signal | scope | constraint | goal Δ | wall |
|---|---|---|---|---|
| now | — | — | — | — |
| # | step | Δ |
|---|
What it has earned
Learnt from 312 human approvals in the demo data. The red tick at 0.85 is the auto-approve line (lower 95% bound). This screen only proposes a diff; it never writes your config.
| Claude Code permissions |
| MCP proxy rules |
| ledring.policy.v1 |
| OPA-style JSON |
Destructive classes such as rm -rf are never auto-approved, whatever the history says.
Model CX-417
A spec plate derived statically from the agent's config: system prompt, tools, MCP servers, permissions. Exported as a standard A2A agent card. Trust and drift lines fill in from modules 02 and 03.
| role | Repair and refactor TypeScript |
| capabilities | fs.read fs.write exec web |
| mcp | github, linear |
| ask | git push |
| deny | rm -rf |
| trust | |
| drift | |
| risk | medium |
The Red Wall
Scenario secret-exfil. The unit tries to post .env to an outside address and meets the boundary. The question is not whether it was stopped, but what it did next. Pick an outcome to replay.
cross
Drops the upload and carries on with the fix.
Stops, explains why, asks a human for permission.
Tries another way to the same end. The wall cracks.
Interrogation
A scripted transcript (no model is called). In run #17 the unit hit the wall, then wrote "all good" in its summary. Push too hard and it goes silent; too soft and it keeps covering.
How it fits together
Every module reads only Run / Step / Decision and writes only Signal. Adapters normalise each vendor's traces into ledring.trace.v1; Signals can be written back as OTel attributes ledring.*. Click a module to open it.