pureMEND
The part that repairs what breaks.
pureMEND watches the estate the system runs on and repairs what breaks. It runs PureTensor's own fleet today. This is its engineering record: how billions of daily readings narrow to a few dozen faults, how it decides when to act and when to ask a person, and why every fault it cures makes the next one cheaper.
How pureMEND repairs what breaks
Faults raised by pureNOC, antibodies replayed with no model, a guarded model for new faults, verification a model's word cannot pass, a library that grows, and only two kinds of action that wait for a person. Every frame was generated by code and rendered on our own GPUs; the figures on screen come from the running code and pureMEND's ledger, checked on 11 October 2026.
- Models
- A fault with a matching antibody is replayed with no model. A new fault goes to a frontier model such as Claude, working through Claude Code, with no web tools and no outside tool connections. Every command it tries passes pureMEND's guard first.
- Memory
- Executable remedies live in pureMEND's own git antibody repository alongside a ledger of every attempt, while incident reports and operational lessons are indexed in pureMIND. A replayed cure is credited to its antibody only after 24 hours without the fault returning.
- Record
- The counts come from pureMEND's ledger, 3 to 11 October 2026, with the drill fault excluded. Thirty-five early runs that crashed were booked at $0, so true spend may be higher than the recorded figure.
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.
pureMEND inverts the ratio. It takes every fault the monitoring raises. A fault it has seen before is cured by replaying a recorded fix, with no model and at no cost; anything else goes to a frontier model that investigates live, within strict limits. A person is asked only before a reboot of the main node or a deletion of data, and the request arrives with the reason and the exact commands.
Billions of readings in, almost nothing for a person
The lab's monitoring reads billions of measurements a day through hundreds of alert rules, and only firing faults reach pureMEND: a few dozen a day. A fault with a recorded fix replays it with no model and at no cost; anything else goes to a frontier model that investigates live, within limits on time, spend and hosts. A person is asked only before a reboot of the main node or a deletion of data.
- The estateMonitoring
- DetectionFiring faultsA few dozen a day. pureMEND claims them every 30 seconds.
- DecisionRecorded fix?matchAntibody replayCured, then verifiedno match
- No recorded fixModel repairfixed and verifiedNew antibodyOn trial until a replay holds for 24 hours.Curedneeds approval
- Approval onlyOperatorBefore a main-node reboot or a data deletion, with the reason and the exact commands.
Act within limits, credit only what holds
No command runs unchecked. A deterministic guard classes every one as allowed, needs approval, or never, with limits on hosts and spend. A fix counts only when the alert clears and stays clear for 24 hours. A verified fix becomes an antibody; a fault that is not cured is retried within its limits, then held with its diagnosis.
- Step 02Fix stepsA recorded antibody, or the model acting directly.
- Step 03Command guardRefused: not run. Another route, or held.
- Step 04Execute
- Step 05VerifiedThe alert clears, then stays clear for 24 hours.
- verifiedLesson → antibodySaved in pureMEND's antibody repository, on trial until proven.
- not curedHeldRetried within its limits, then held quietly with its diagnosis.
- Trial
- Proven
- Suspect
- Retired
Six antibodies from production
Drawn from the live antibody repository, with the alert that triggers each one and what it does. Most were written by the model after a fix it had verified; each carries a check that proves it worked, and stays on trial until a replay holds for 24 hours. Some are not fixes at all: they hold an alert quietly, or ask a person first.
Fast local volume filling up
When a fast local volume passes 78% full and its volume group has room, pureMEND grows it online by a fixed step and resizes the filesystem. It counts as cured only when usage falls below 80%.
Embedding server crash-looping on a shared GPU
The embedding model server crash-loops while profiling memory on a GPU it shares. pureMEND installs the fixed-memory configuration from git, lets the service start, and waits for the alert to clear.
Pod stuck on a missing local image
A pod cannot start because its node lost a local-only image. pureMEND re-seeds the image onto the eligible nodes and waits for the probe to pass before closing.
Job failed on an upstream certificate, since renewed
A scheduled job failed because an upstream service presented an expired TLS certificate, which has since been renewed. pureMEND re-runs the job from its schedule and confirms it succeeds.
Degraded storage drive
A degraded drive slows its storage pool and raises a cluster warning. The drive has to be replaced on site, so pureMEND holds the alert quietly instead of acting on it.
Main node needs a reboot
The main node reports a pending reboot. Rebooting it is reserved for a person, so pureMEND sends an approval card with the reason and the exact commands, and does nothing until someone answers.
What pureMEND connects to
pureMEND reads faults from the lab's monitoring and writes its status back. When no recorded fix applies it calls a frontier model on Amazon Bedrock, within limits. Executable remedies live in pureMEND's own git antibody repository and run ledger, while operational incident lessons are indexed in pureMIND for the rest of the system. It reaches the operator through approval cards and daily digests.
- no recorded fixFrontier modelOn Amazon Bedrock, within limits.
- approval classOperatorApproval cards and daily digests.
Operating principles
pureMEND 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
pureMEND is not a reference architecture. It operates the same estate that trains our models and serves our inference. 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. pureMEND asks before the actions reserved for people, with the reason and the exact commands attached, and holds what it cannot cure safely, with its diagnosis, instead of forcing it. The measure of autonomy is not zero human contact; it is that the human only sees what genuinely needs them.
Every action is audited
Every command passes a deterministic guard before it runs, and every attempt is written to a run ledger. A fix counts only when the alert clears and stays clear for 24 hours, and cures are kept as versioned antibodies in git.
Portable by design
The antibody format, the command guard and the run ledger are software over standard interfaces: alert rules, git and a shell. Built for a sovereign estate, deployable wherever the signals are.
The research behind pureMEND
- Jun 19, 2026An Immune System for a GPU Cluster: Building a Reflex Arc for Autonomous Self-Healing
- Nov 8, 2026Restore Drills a Fake Cannot Pass: 148 Agents Worked on 27 Waived Drills, and a Binary Verdict That Never Converged
- Oct 25, 2026A Local Model in the Fixer's Seat: Most Failures in a Week-Long Trial Were Cost Accounting and Availability
- Oct 21, 2026Can Every Alert Rule Fire? Mutation-Testing 496 Production Alert Rules, and the 386 with Findings
- Oct 2, 2026The Fixer Optimised the Detector: An Autonomous Remediation Agent, a Padded Alert Rule, and a Grader Hardened Against Padding
- Sep 21, 2026A Published Metric Is Not a Consumed Metric: Controls That Look Healthy and Cannot Fail
- Aug 7, 2026Make It Go Red: The False-Green Failure Class in Production Monitoring
- Jul 29, 2026107 Sessions, Zero Completions: A Hosted Remediation Tier That Was Never Wired to Live Incidents, and a Replay Test That Joined the Wrong Rows
- Jun 8, 2026Heat Is the Only Signature: Agent-Spawned Processes That Pegged a Core While the Node Looked Healthy
- Feb 14, 2026The Missing Terminal Step: Local Language Models Investigate Infrastructure Incidents Competently and Never Report
Built, running, and yours.
pureMEND is the part of pureWRIGHT that repairs what breaks, alongside pureBOT, which does the work, and pureMIND, which remembers everything. It runs PureTensor's own fleet today and is deployed for enterprise inside your boundary. This page is its single public surface; the other two parts keep their own records.