Sentinel
Infrastructure that diagnoses, heals, and remembers.
This is not a product page. It is the engineering record of the operating layer that runs PureTensor's fleet: how millions of daily signals are triaged for the cost of almost nothing, how the system decides when to act and when to wake a human, and why every failure it survives makes the next one cheaper.
Monitoring that only alerts is half a system
Classical monitoring ends at the pager. The dashboards watch, the rules fire, and then a human does everything else: reads the graphs, remembers the last time this happened, finds the runbook, types the fix. The operator is the remediation engine, and the operator does not scale.
Worse, alert volume corrodes the signal. A rule that pages for everything teaches the team to read nothing. The failure mode of monitoring is not silence; it is noise that trains people to ignore it.
Sentinel inverts the ratio. Every signal on the estate passes through a triage tier cheap enough to run continuously; failures the system has seen before are handled by replaying a proven fix; genuinely novel ones get deep reasoning. What reaches a human is the residue neither tier could resolve, and it arrives with the diagnosis already written.
Millions in, a fraction of a percent out
The economics only work as a funnel. A triage tier cheap enough to read every signal runs continuously; known failure shapes replay a proven fix without further reasoning; only the genuinely novel earns expensive attention, and only the unresolved earns a human.
Act where certain, escalate where not
No fix executes on a hunch. Every remediation is a signed plan that must pass a policy and blast-radius gate before it touches the estate, and every execution is graded afterwards. Success becomes an antibody. Anything less becomes the operator's briefing.
Six antibodies from production
Drawn from the live corpus, fingerprint and remediation included. On familiar patterns the triage tier recalls an antibody with a 91.3% hit rate, at 99.94% triage availability. Every entry below has fired against the real estate; the invocation counts are the audit log talking.
Stale RBD lock blocking pod startup
A pod fails to mount a distributed block volume because the previous mounter died without releasing its lock. Sentinel verifies no live mounter exists, releases the lock, and retries the mount.
Node proxy wedge on storage-tier nodes
Service connectivity probes fail from a specific node subset. Sentinel restarts the node agent, validates that routing rules propagated, and re-runs the probe before closing.
Unit file corruption after package upgrade
In-place upgrades occasionally rewrite a unit's start command with literal escape sequences. Sentinel diffs against a healthy peer, restores the clean unit file, and reloads the daemon.
Monitoring targets pointing at retired hosts
Scrape failures with non-existent-domain errors reveal targets for hosts that no longer exist. Sentinel cross-references the active inventory, prunes the stale targets, and regenerates the manifest.
Storage driver stale node registration
After a node rotation the storage driver retains a stale node identity and volume attachments hang. Sentinel forces the driver to re-register and confirms attachments drain.
Image architecture mismatch on mixed nodes
A single-architecture image scheduled onto an incompatible node crashloops with an exec format error. Sentinel patches a scheduling exclusion and flags the missing multi-arch manifest for a human.
Perception, orchestration, memory
Sentinel is the layer that watches and acts. Beneath it, pNexus routes work across inference tiers and tools; beneath both, pureMind remembers. Three names in engineering, one system in operation, and each layer keeps its own written record.
Operating principles
Sentinel was not designed on a whiteboard and installed. It grew inside a working fleet, one survived failure at a time, and its principles are the ones that survived with it.
It runs our own fleet, daily
Sentinel is not a reference architecture. It operates the same estate that trains our models, serves our inference, and runs this website. Every claim on this page is exercised in production before it is written down.
Escalation is a feature
A system that never asks for help is either perfect or lying. Sentinel treats a clean, well-briefed escalation as a correct outcome: the measure of autonomy is not zero human contact, it is that the human only sees what genuinely needs them.
Every action is audited
Remediations execute from signed plans, pass a policy and blast-radius gate, and are graded against what actually happened. The record of what was done, by what, and why survives in versioned memory.
Portable by design
The triage tiers, the antibody corpus, and the memory substrate are software over standard interfaces. Built for a sovereign estate, deployable wherever the signals are.
Documentation, not sales.
Sentinel is operational, documented here so the shape of the operating layer is on the record. This page is its single public surface; the orchestration and memory layers keep their own records. Operators running comparable estates who want a deeper briefing can reach the team through the contact form.