The Workflow
Five phases. One spine.
Identify. Validate. Decide. Act. Measure.
One arc, one thread
Three tools, three copies, none of them connected.
The tracker lives in a spreadsheet. The story lives in a slide deck. The judgment lives in a consultant’s memo. And the moment anything changes, all three go stale.
Vernais runs the whole arc on one spine. A cause becomes a plan. A plan becomes work. Work becomes a locked metric you watch move. Every hop is grounded, every ID is human-readable, and nothing is re-typed between phases.
01 · Identify
Identify. What’s actually happening?
An investigation just told you why. Identify turns that finding into the anchor everything else validates against.
- Reads the AI-written problem statement in plain language.
- Reviews the captured signals the investigation surfaced.
- Drills into clusters to separate the real pattern from the noise.
- Pulls the bounded signal set the investigation actually traversed, not your whole firehose.
- Clusters those signals with HDBSCAN into on-topic vs noise, so a spurious mention doesn’t distort the picture.
- Breaks the evidence down by source mix, integration plus outcome, so you see where the problem lives.
- Writes a problem statement you can read in one breath.
Signal clusters, related vs noise
Each cluster is a group of Signal records the graph tied together. Related clusters describe the problem. The noise cluster is set aside, visible but not counted: the same honesty that governs the graph.
Source mix by integration + outcome
A breakdown that answers “where is this coming from, and how is it landing.” Failed Stripe charges, PostHog onboarding drop-offs, Sentry errors, each with a grounded count, each carrying its filter recipe.
02 · Validate
Validate. Is the diagnosis sound?
Identify says what’s happening. Validate proves whether the explanation holds, and starts building the number you’ll measure.
- Reviews confidence-scored hypothesis cards and pins the leading one.
- Marks each attached piece of evidence relevant or not.
- Reviews the math behind the candidate North-Star metric before it’s locked.
- Builds hypothesis cards straight from the investigation’s evidence ledger, each with a confidence score and an ID.
- Correlates every signal that arrives (a page, chat message, meeting transcript, or file) against the problem statement.
- Surfaces a candidate tracking metric with its full formula and receipts, ready for you to read the math.
Confidence-scored hypothesis cards IN-42-H1
Each hypothesis comes from the investigation’s own reasoning, carries a stable confidence, and never silently flips run-to-run. Confidence is honest by design: a causal claim is capped at “medium” unless the source text verbatim entails it.
The correlation ledger (cosine)
Every new signal gets a cosine score against the problem statement, so you see related vs unrelated at a glance. A related page can bump or spawn a hypothesis. You keep the veto: mark-as-evidence overrides the machine either way.
The candidate metric’s math card
Before anything is locked, you see exactly how the North-Star would be computed: the formula, the numerator, the denominator, the events it counts. A closed grounding lock means the metric can only be built from real events the field index actually discovered. The model cannot invent a number.
03 · Decide
Decide. What’s the fix?
Validate confirmed the cause. Decide picks the response from Vernais’s drafts and from your team’s own ideas.
- Reviews AI-drafted solutions against the root cause.
- Promotes cluster-born solutions worth pursuing.
- Attaches evidence as supporting or contradicting.
- Picks the current solution: the one Act will build.
- Generates default solutions directly from the root cause.
- Carves signal-born solutions out of the clusters and out of anything your team proposed in a doc, chat, or meeting.
- Deduplicates incoming signals so the same idea doesn’t clutter the board twice.
- Maps each signal to solutions as supporting or contradicting evidence.
Default and signal-born solutions IN-42-S1
A default solution reasons from the diagnosed cause. A signal-born solution is a real proposal your team made: a fix suggested in a meeting or a 
Slack thread, carved out, given an ID, and linked back to the source signal. Both sit on the same board for you to judge.
Evidence attached as supports / contradicts
When a new signal proposes or challenges a fix, Vernais attaches it to the relevant solution with its stance. You can see, at a glance, what’s backing each option and what’s pushing against it.
Pick the current solution
One solution becomes current. That’s the blueprint Act decomposes. You can change it; changing it re-plans downstream.
04 · Act
Act. Who builds it, and when?
Decide chose the fix. Act turns it into scheduled, assigned, trackable work, and finds the right person for each piece.
- Reviews the auto-decomposed tasks.
- Assigns owners, sets priorities and estimates.
- Moves work across the board and clears blockers.
- Decomposes the chosen solution into tasks, each typed tech, design, or non-tech, each with its own state machine.
- Schedules the work with a Gantt engine that respects capacity, dependencies, and committed dates.
- Auto-advances technical tasks from real git signals: a merged branch moves the task without a manual drag.
Auto-decomposed tasks IN-42-T1
The current solution breaks into concrete tasks, each with an ID and a type. A tech task follows an engineering ladder; a design task follows a design ladder; a non-tech task follows its own. The right workflow for the right work.
A Gantt engine that schedules for you
The plan isn’t a static bar chart. The scheduling engine reads each task’s estimate, its dependencies, and the assignee’s real capacity, then lays out an earliest-viable timeline. Move one task and the dependent ones re-flow.
The best person, ranked, with a plain “why”
For each task, Vernais ranks your teammates and tells you why in words, not a black box.
- Role-fit for the task type: an engineer for a tech task, a designer for a design task.
- Current load: how much is already on their plate this cycle.
- Earliest start: when they can realistically pick the work up.
Auto-advance from git
Technical tasks watch the real 
GitHub signals wired through the Act board, branch created, commits landing, PR merged. A merged task can move itself forward, so the board reflects reality without babysitting.
05 · Measure
Measure. Did it work?
The work shipped. Measure holds you to the number you set out to move, and proves the outcome.
- Locks the North-Star metric as an immutable baseline.
- Takes readings to append growth snapshots.
- Sets status: tracking, shipped, rolled back, or paused.
- Freezes the metric definition and its investigation-time value as an immutable baseline, no recompute, no re-parse, no drift.
- Replays the exact recipe over a fresh window each time you take a reading, appending a grounded snapshot.
- Rolls the locked metric up into Launches as a tracked outcome.
The metric math card
The formula, the numerator, the denominator, each with the receipt for how it was counted. The same math you reviewed as a candidate in Validate, now frozen.
Lock as baseline (immutable)
Locking captures the metric’s definition and its value at investigation time. That baseline never recomputes. It’s the fixed line you measure growth against.
Take a reading: snapshots that replay the recipe
Each snapshot re-runs the exact stored recipe over a fresh window. No re-derivation, no NLP re-parse: the same query, a new point on the line. That’s how a snapshot is trustworthy.
Status lifecycle + outcome sparkline
Set the initiative’s status: tracking, shipped, rolled back, or paused. Then watch the outcome sparkline move from baseline to latest reading.
The through-line
One thread runs the length of the spine.
An investigation opens an Initiative (IN-42).
Docs, chats, and meetings feed it evidence and solutions, each correlated, classified, and routed to the right phase automatically.
Solutions (IN-42-S1) decompose into tasks (IN-42-T1).
Tasks close, and a Tracking metric locks against its Baseline.
The metric flows to Launches as a measured outcome.
Every hop is logged in Activity. Anything important can surface in Inbox. Nothing is re-typed, nothing is invented, and every number carries a receipt.
You started by asking why. You end with proof of what shipped.