Skip to main content

Product

An investigation you can read back

Production Master turns approved production signals into a cited report: a conclusion, the exhibits under it, the hypotheses it rejected, and a proposed action that waits for a human.

Report surfaceconclusionexhibitsapproval gate

Interactive walkthrough

Follow one incident from signal to gate

The evidence workspace below uses synthetic data to show the reading order: what happened, what was tested, and where a human takes over.

checkout-api · 14:23–14:31 UTC

Synthetic signalE-01

Checkout latency rises immediately after deploy a83f9c

The synthetic checkout-api p99 series increases from 410 ms to 2.9 s after the deployment window.

Source boundaryMetric
1 / 6

Three commitments the report has to keep

They are the same three on every investigation, and each one is visible in the artifact the product produces.

01

Gathers, then cites

Every log line, metric window, deploy diff and config change it reads becomes a numbered exhibit. Nothing enters the report uncited.

claim → E-0n → raw source

02

Argues against itself

Competing hypotheses are tested and the losers stay in the report, with the evidence that killed them. A rejected theory is a result, not a deletion.

considered · rejected · why

03

Stops at the gate

It proposes the change and waits. Remediation is a decision a human makes, holding a report they can check line by line.

proposed → awaiting approval

What happens between the alert and the answer

Seven stages. The investigation can stop at any of them, and it always stops before production changes.

  1. 01

    Detect

    An alert, a deploy, or a request opens an investigation with an explicit evidence window.

  2. 02

    Gather

    Read-only connectors pull the deploy diffs, metric series, logs, cluster events and config changes inside that window.

  3. 03

    Exhibit

    Each source becomes a numbered exhibit with its origin, timestamp and volume recorded.

  4. 04

    Hypothesise

    Candidate explanations are written as falsifiable statements, not as prose.

  5. 05

    Challenge

    Each candidate is tested against the exhibits.
    Rejections are kept, with the exhibit that rejected them.

  6. 06

    Report

    The surviving explanation, its citations, its confidence and the ledger of rejections are assembled.

  7. 07

    Gate

    A remediation is proposed and the investigation stops.
    A human decides whether it is applied.

A rejection is a result

Every hypothesis the investigation opened stays in the record with its verdict and the exhibit that produced it.
You can see what was ruled out before you are asked to trust what was not.

PM-SAMPLE-001Hypothesis ledger4 considered · 1 held
rejected
Cache degradation
Cache hit rate stays within baseline across the whole incident window, and no cache-layer change was deployed. [E-04]
rejected
Upstream provider latency
Upstream latency does not change during the window, and checkout-api is slow on requests that never reach the provider. [E-04]
rejected
Traffic spike
The latency series steps rather than ramps, and it steps at the deployment boundary rather than with load. [E-01]
held
Connection-pool exhaustion after deploy a83f9c
The pool ceiling drops 80 to 20, latency steps immediately after, and the logs show zero idle connections at the new limit. [E-02, E-01, E-03]

The product stops before production

Production Master prepares a remediation and the evidence for it, then holds.
Applying the change is an action a person takes, on a report they can check line by line.
There is no autonomous remediation path and no configuration that creates one.

  1. 14:28 UTC

    Alternative challenged

    The alternative does not correlate with the latency increase, so it is not supported by this sample evidence.

  2. 14:30 UTC

    Cause supported

    Confidence remains below certainty because the sample does not reproduce the issue or include a post-change recovery observation.

  3. 14:31 UTC

    Human approval

    Production Master prepares the action and evidence; it does not execute a production change without human approval.

How it runs

Choose a deployment model without changing the evidence contract.

Read-only by default

Connectors are scoped to read.
Production Master proposes changes; applying one is a separate, human-initiated action.

Runs where your evidence lives

Self-host the runtime via Helm inside your own infrastructure, or use the managed service.
The investigation engine is identical.

No host lock-in

The runtime depends on no specific cloud or execution platform.
Hosting choices sit behind explicit interfaces.

Bring your own model keys

Point the runtime at your own LLM provider credentials, or use ours on the managed plans.

Read one investigation before you believe any of this

The evidence workspace is the whole product story — citations, rejected hypotheses, confidence and the approval gate.

Self-host via Helm — available now to closed-beta teams.