Plans
A written intention before an action, so the reasoning can be reviewed rather than reconstructed afterwards.
THE PLATFORM
One runtime. Every operation.
CortexVigil is a vision AI platform. The engineering agents work inside a git repository. On the surface they share nothing; underneath they are built the same way and held to the same rules.
The shared part: the four-step loop every solution closes, the governance layer that decides what a system may do, and where it runs. Written for the people who review it before a pilot — security, platform, and whoever owns the audit.
01 — THE STACK
Four layers. The middle two are ours.
The top and bottom belong to you — your sensors, your systems of record — and we do not ask you to move either. What we bring is the part in between: something that can reason about what it sees, and something that decides what it is allowed to do about it.
LAYER 01 — YOURS
Sources
LAYER 02 — OURS
Agent runtime
LAYER 03 — OURS
Governance
LAYER 04 — YOURS
Systems of record
Tools
A declared, finite set of things it can do. Nothing it can reach is undocumented.
Memory
What happened on this site, in this repository, last week — so it stops asking the same question twice.
Evaluation
Scored against cases kept from real deployments, so a model change is a measurement rather than a hope.
02 — THE LOOP, IN TWO WORLDS
The same four steps, in a control room and in a repository.
Read down a column for one product’s loop. Read across a row to see why they are one system underneath.
| Loop step | CortexVigil | Engineering agents |
|---|---|---|
| 01Observe | Every camera on every site, continuously — not the four a person happens to have on screen. | The repository, the ticket, the CI history and the conventions the team already follows. |
| 02Decide | Is this a violation on this site, in this zone, at this hour — by your rules, not a generic safety model’s. | What change would satisfy this ticket, and is it a change this team would accept in review. |
| 03Act | Raise the incident, page the supervisor, log the vehicle, clip the evidence — inside your systems. | Write the change, run your CI, open the pull request on your branch. Never merge it. |
| 04Account | What was seen, which rule fired, who acknowledged it, what was overridden and by whom. | Every tool call, file read, refusal and retry against the run, reviewable at the next audit. |
03 — GOVERNANCE
The interesting part is what it refuses to do.
A system allowed to act on your site or in your codebase is a security boundary before it is a productivity feature. The limits are not written in a prompt where a clever input can argue with them. They sit outside the model, they are enforced whether the model agrees or not, and hitting one is a logged event with a reason attached.
“The agent decided to” should never be the end of an incident review. That is the whole design goal.
Scope limits Zones, directories, services and record types it may never touch — checked before the action, not after.
Named human gates Specific people on specific decisions. A gate the system can tick itself is not a gate.
Secret and network isolation No production credentials, no unlisted outbound calls. The sandbox is the default, not the enterprise tier.
Data stays where it starts Video can stay on the site. Source can stay in your VPC. Neither becomes anyone’s training data.
A REFUSAL, AS IT IS LOGGED
$ svakrit audit show run/418 --refusals REFUSAL ENG-DEP-00 run 418 agent 11 · engineering at 14:24:31 wanted add dependency [ratelimiter@2.1] rule no new third-party dependencies source policy/engineering.yaml:22 outcome wrote against existing client noted tradeoff recorded in PR body [1] refusal · [0] overrides · [0] silent skips
WHY IT IS A RECORD AND NOT A LOG LINE
A refusal names the rule, the file the rule lives in, what the agent wanted, and what it did instead. That is the difference between a system you can certify and a system you have to trust.
Silent skips are counted too. An agent that quietly does nothing is a failure mode, not a safe default, and it shows up in the same summary.
04 — DEPLOYMENT
The intelligence travels to the data.
Not the other way round, and not through anyone else’s cloud unless you choose it.
USED BY CORTEXVIGIL
On the edge
An edge node at the site does the inference. Video never has to leave the premises — only the events, and the clips you ask for.
USED BY THE ENGINEERING AGENTS
In your VPC
Self-hosted inside your own network, with the model provider you have already approved. Your source never crosses a boundary you did not draw.
AVAILABLE FOR BOTH
Hosted by us
When you would rather not run anything: the same runtime, the same governance layer, the same audit trail, in [region].
05 — SECURITY
What a security review asks, answered in one place.
Nothing below is new. These are the controls described in 03 and 04, collected here so a reviewer does not have to go looking for them.
Enforced outside the model
LIVEScope limits, named human gates, and secret and network isolation are not written in a prompt where a clever input can argue with them. They are enforced whether the model agrees or not, and hitting one is a logged event with a reason attached.
Per-tenant isolation
LIVEEvery site, repository and customer is its own tenant. A rule authored in one tenant, and the evidence it produces, are not reachable from another.
Data stays where it starts
LIVEVideo can stay on the site behind an edge node. Source can stay inside your VPC. Neither becomes anyone’s training data, on any deployment option above.
An audit trail, not a log
LIVEEvery decision, action, refusal and silent skip is written down with the rule that produced it and the file that rule lives in — reviewable months later, by someone who was not there.
Formal certification status, penetration-test reports and a completed security questionnaire are handled in the technical session below rather than published here.
Want the version with the architecture diagrams in it?
We will walk your security and platform people through the runtime, the policy model and the deployment options before anyone talks about a pilot.