What ships, in what order.
Nine phases, each ending with something demonstrable. Phase 0 first, because every later phase writes to the ledger.
Gets better without teaching itself to lie.
“Learn from how engineers fixed errors” is the right idea, but three distinct mechanisms usually get conflated there, and only two are safe. We'd rather ship accretion and compilation first and say plainly that policy learning is not ready.
Every resolved incident becomes a structured record: symptom signature, discriminating checks, root cause, remediation, verified by, outcome, trust tier. No behaviour change, no risk. This ships first because it's the least interesting to build and unblocks everything else.
When an engineer writes “restart app-b and rerun the batch job,” that's a claim. 3AM compiles it into assertions plus an action sequence. Every learned artifact is falsifiable and re-validated by replay against historical resolutions, and compilations that stop reproducing lose their tier automatically.
An agent optimising a verification script learns to make the script pass, not to fix the cause. Every proxy has an exploit, and an optimising agent will find it: satisfying a trial-balance check by writing a balancing journal entry is literally a catalogued incident. We defer this until the reward signal is trustworthy.
Every learned artifact is falsifiable, and every compilation is validated by replaying it against historical resolutions. A learned artifact is re-falsified on every use: if it stops reproducing, it loses its tier automatically.
The sequence is the product.
Each phase ends with something demonstrable. Phase 0 comes first because every later phase writes to the ledger. Phases 1 and 2 can run in parallel.
monorepo, naming, ledger library and event schema, build pipeline (Nuitka, cosign, SBOM), licence tool
A signed, compiled image runs, refuses an invalid licence, and writes a verifiable hash-chained ledger.
Prometheus/Alertmanager, Grafana, Splunk, Datadog, Git hosts, Jira, ServiceNow, Kubernetes, SSH, SQL, HTTP executors
Each connector passes the SDK conformance suite against a sandbox or mock; Pinata connects only through them.
Slack (Socket Mode), Symphony, PagerDuty, xMatters, Twilio IVR, e-mail, escalation, SSO identity, quorum
One-tap approve/reject round-trips on every channel with no inbound ports; each decision is in the ledger with approver identity.
discovery, per-repo shards, incremental indexing, federated search
300 synthetic repos indexed incrementally on a CPU-only VM within a set time budget; alerts resolve to service, repos and owners.
applies_when, query adapters, technology packs (MySQL, Postgres, Oracle, MSSQL, JVM, nginx, Kafka, Kubernetes, Linux)
The same check class confirms the same failure on Prometheus and Splunk data; packs are only readable in memory with a valid licence.
concurrency, per-service autonomy, policy engine, daily digest
A week of shadow mode on Pinata produces digests; flipping a service to L1 goes through approvals.
setup wizard, estate map, incident timeline, approvals, audit explorer
A new user connects Pinata end to end through the wizard alone, without touching config files.
Helm, Compose, installer, preflight, GPU detection, CPU profile, air-gap bundle
A clean VM goes from install to shadow-mode proposals on Pinata in under 30 minutes (target).
a second legacy stack (for example .NET/MSSQL or Python/Postgres), smoke, fast and nightly suites
Both estates scored through the same product, held-out sets included.
- Held-out control set: scores must collapse when the corpus is removed
- Time-to-mitigate improving against baseline
- Collateral damage flat or improving (this is the guard)
- Abstention precision: abstains when it should, not more
- Human override rate decreasing over time
- Upgrading the model before the verification layer exists: that buys a more confident narrator saying the same unverified thing
- Letting the retrieval index read ground-truth files at query time, so every score becomes meaningless
- Training on our own root-cause score, which is gameable by construction
Point it at a monolith that scares you.
We’re onboarding a small number of teams running legacy core systems we can’t rewrite. The install runs on your hardware, behind your firewall, connected through a guided setup. So if you’ve got a service where the on-call rotation has learned to dread the pager, we want to see it.
Shadow mode by default: proposes and logs, touches nothing until you approve it.