BELLWETHER
COHORT 01 · DESIGN PARTNERS · Q3 '26 · forming

Quality needs a new interface.
Lead the release.

Release confidence for your CI. One ledger for humans and the agents you ship with.

Failure inbox
Every failure carries a verdict, confidence, and reasons[]. Never a black box.
Release confidence
One call, with the evidence behind it: should_block_release(). Override is one click.
Reads from your CI
GitHub Actions · Playwright · Cypress today. One YAML step. Built to absorb your full pipeline.
Agent-readable
Every screen has a JSON peek and an MCP tool. Humans, APIs, and agents read the same ledger.
On the name

Old word. A bellwether is the lead sheep that wears the bell, the one whose movement the flock reads to know which way the day is going. We took the name because that's the gap. Your CI emits a thousand signals an hour. Someone has to be the bell: the neutral, audible thing that says this is real, that isn't. Bellwether is that bell.

§ 02 · MANIFESTOTHE SHIFTBWX/01
02 · The shift→ a worldview, not a feature grid

Software changed. QA changed. The way teams decide what to ship hasn't.

  1. 01 · WAS

    QA was a handoff.

    A team owned testing. They wrote suites, gated releases, and the rest of engineering trusted the green check. The interface to quality was a person.

  2. 02 · IS

    QA is shared.

    Agents write tests. SDETs review them. Engineers triage failures in Slack. Testing is faster and noisier; ownership is everywhere and nowhere. The handoff dissolved.

  3. 03 · NEEDS

    QA needs a decision layer.

    Reruns and gut feel don't scale to a thousand runs a day. What's missing is a neutral ledger that classifies every signal, names the owner, and tells you when to ship.

    Bellwether is the bell.

§ 03 · DIAGNOSISTHE OLD WORKFLOWBWX/02

This is your quality decision layer today.

It's a Slack thread. Flaky test triage runs on tribal knowledge, "rerun it once," and the muscle memory of whoever happens to be online. CI failure triage is whatever the on-call engineer can remember at 4pm on a Friday. It works. Then your test count crosses a threshold and the noise becomes the signal.

§ 04 · MECHANISMTHE DECISION LAYERBWX/03
04 · The new decision layer

The same failure, twelve seconds later.

Bellwether reads your CI events as they land. Every failure ships with a verdict (flake, regression, infra, or new), a confidence score, weighted reasons, a recommended action, and the likely owner. The event model is neutral by design: build, deploy, and security signals join the same ledger as we widen coverage.

Same ledger · three surfaces

One classification, three calls: the human triage page, the REST API, and the MCP tool agents call. Same evidence, same confidence, same owner. No surface tells a different story.

HUMAN
bellwether.run/triage/fail_8820
The triage page. One verdict, weighted reasons, recommended action.
API
GET /v1/failures/fail_8820
Same JSON your dashboard reads. Hook it into Slack, Jira, your release bot.
AGENT · MCP
classify_failure(fail_8820)
Your release agent calls the same tool your engineers read. One ledger, both audiences.
§ 04b · TRUSTHOW WE WON'T LIE TO YOUBWX/03b
04b · Trust

No black-box verdicts.

Three things we wrote into the spec before we wrote a classifier. They are the difference between a confidence score you can trust and a number that gets ignored after the first wrong call.

COLD-START CAP
First 7 days, confidence is capped at 0.7.
On a new repo with no override history, every verdict ships with a low ceiling and a default of `unknown`. We earn calibration before we claim it.
INSUFFICIENT HISTORY
No history, no number. We return null.
If a test has run fewer than five times or the cluster lacks enough signal, the confidence field is `null`, not a fabricated low score. The product is honest about what it doesn't know.
FEEDBACK REQUIRED
Every verdict (UI, Slack, MCP) has a “was this right?” affordance.
Overrides are first-class events. The next classification on the same fingerprint reads your last override before the model fires. The learning loop is the moat; the affordance is non-optional.

Spec refs: CLASSIFY-05 · CLASSIFY-06 · LEARN-01 · LEARN-02. The learning loop is the moat. We built the affordance first.

§ 05 · POSTUREWORKS WITH YOUR STACKBWX/04
05 · Posture

The neutral layer. Not another silo.

Bellwether sits across your stack. We read CI events; we don't replace your test framework or your tracker. Coverage starts with test signals and absorbs build logs, deploy gates, and security scans as we widen.

GH
GitHub Actions
CI events · YAML step
PW
Playwright
Test runner · traces
CY
Cypress
Test runner · runs
SL
Slack
Failed-run summaries
JR
Jira
Auto-draft tickets
LN
Linear
Route to owners
What we read
test names · errors · stack frames · retries · durations · git SHAs · branches
What we never read
source code · env vars · secrets · user data inside fixtures · production traffic
● SOC 2 Type I · in flight● Self-hosted ledger · q4 ’26
§ 06 · PROOFDESIGN PARTNERSBWX/05
06 · Pilots

Three design partners. Twelve weeks. Numbers, not adjectives.

Targets · Cohort 01 · Q3 ’26 · not yet measured
−68%
TIME-TO-TRIAGE
Median minutes from CI failure to a verdict an engineer can act on. Cohort target: drop p50 from 21 → 7 minutes by week 12.
−41%
FALSE BLOCKERS
Releases held by flake-misclassified failures. Cohort target: prevent 41% of those holds without rolling back regressions.
3.2×
RELEASE DECISION SPEED
Median time from “RC built” to ship/block call. Cohort target: replace the meeting with a Slack click backed by a confidence score.

Cohort 01 starts Q3 ’26. We’ll publish weekly numbers (baselines and progress against these targets) once pilots are running. No retroactive revisions; the first reading is the first reading.

§ 06b · MEMBERSHIPTHE COHORTBWX/06b
06b · Cohort→ quarterly cadence, soft windows

Cohort 01 for Q3 '26.

A small first cohort. Real CI traffic. Direct line to the founders.

We onboard design partners one cohort per quarter. Cohort 01 is the first set of teams shipping on Bellwether. Their classifications, overrides, and owner-mappings shape V1.1. Their feedback is weighted higher in the learning loop than anyone who comes after, because they showed up first.

What you get

  • V1 access from day one
  • Co-design on V1.1
  • Founder Slack channel
  • Classification feedback weighted ×3 in the learning loop

What we ask

  • Real CI traffic, not a sandbox
  • One review call per cohort
  • Permission to learn from your labelled overrides
Bellwether

Lead the release.
Don't just run the tests.

Cohort 01 is forming for Q3 '26. Cohort 02 picks up Q4.

One repo, one CI, one Slack channel. Four minutes to set up.

No new dashboards to live in.

Talk to a founderRead the docs

no credit card · revocable token · works with what you already ship on