Evidence and trust

One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

Watch the plan while it runs

An investigation takes a few minutes. A box appears in the chat labeled Thinking… with a live timer. Stage cards stack up inside it. Each card shows a title, a spinner while it works, and the model's own reasoning as plain prose.

These are the stages that matter most to you:

• Breaking down the question — your question is split into 3 to 5 sub-questions.
• Picking the relevant topics — the search area. Open the card to see which topics it entered and how close each one was.
• Forming candidate explanations — the rival theories, written before it looks at any evidence.
• Convening the node jury — the long one. A progress bar counts records as they are judged.
• Pressure-testing the verdict — a PM lens and an engineer lens argue over the result.
• Writing the cited answer — the case file is assembled.

A card with a chevron is clickable. The jury card is the one worth opening, and it renames itself to Jury verdict in once it lands. It lists the records that survived, each with the quote the juror had to copy out of it.

When the run finishes, the box collapses to Thought for 3m 12s. Your timeline is not gone. Click it to open the whole thing again, case file included.

Read the answer, then the case file

The answer is written prose with citation markers in it. Under the answer sits the case file: a confidence badge and a stack of sections. A section only appears when it has something in it, so no two case files look the same.

Follow a citation to the record

1. 1
   Find the evidence chain

   It sits at the top of the case file, already open. Each row is one claim. The mark beside it says whether a verifier could tie that claim back to its source, or could not, or did not check it.

2. 2
   Click the citation chip

   The chip at the end of the row is a button. A chip with no record behind it is greyed out and will not click — that claim is framing, not evidence.

3. 3
   Read the record in place

   It opens right under the claim, not on a new screen. You get the record's own text, or its real field values when it is telemetry with no prose.

4. 4
   Keep walking

   Facts lists what Vernais pulled out of that record. Connected to shows the records it links to, so you can follow the trail further.

The same click works everywhere a record appears. The chips in Evidence graph, the rows in Matching records, the Sample failing events under measured impact, and the open links in the graph exploration trace.

Read the competing explanations

Before it looks at a single record, Vernais writes 2 to 4 candidate explanations. The rules on that step are strict. Each one must name exactly one specific cause. They must be mutually exclusive, so evidence for one is evidence against another. Each must be testable on its own. And at least one has to be a genuine alternative — an outside cause, or a problem that was already there before the change you suspect.

That last rule is the important one. It forces the system to carry a theory that would embarrass its favourite. Without it, an investigation grades its own first guess.

Hypotheses considered shows each theory with a badge: likely, possible, or ruled out. Ruled out then lists the losers, each with a reason line. Read those reason lines closely, because two very different outcomes share that heading:

Check the confidence badge

A colour-coded badge sits above the case file: green for high, amber for medium, red for low. The model states a label first. Then post-checks run, and if a check objects, the label drops. The prose is rewritten to match the badge, so the words and the badge can never disagree.

hypothesis only is a different badge, and it lives on the timeline rather than the case file. When the final gate refuses to certify a cause, the candidate-explanations card is retitled Candidate explanations — causal verdict withheld and wears that badge. Open it and it names the reasons. Confidence is forced to low.

Several checks can pull a high down to medium, and the caveat line names whichever fired. Two come from the causal check. One: a symptom in the timeline predates the first dated change you blamed. Two: the error it named covers only part of the cohort that is actually failing. Both are worth acting on — a medium here usually means the pattern is real and the mechanism is incomplete.

A third is on you. Name a tool in your question, and it may turn out to supply under about a third of the grounded evidence. The answer then calls that angle tangential and knocks the high down. Confidence covers what each level should change about your next move.

Read the measured impact

Inside are the real counts from your event tools. A by-version table with a failure percentage per version. Top errors with how often each fired. The most-affected users. A bar per event name. And sample failing events you can click open like any other record.

These counts are scoped to this incident, not to your whole workspace. The scan starts from the records that survived the jury and counts the telemetry the graph ties to them.

Take the tracking metric

After the answer lands, Vernais derives a metric for the problem it diagnosed. The section reads Tracking metric (North-Star) when it settled on one, and Tracking metrics when it derived a set — a north star plus the supporting metrics around it.

Each card shows the metric name, the value right now, and the formula in monospace. Under that sit the numerator and denominator with the exact events counted on each side, why it picked that metric, and a suitability badge.

Check that the events in the numerator are the failure you care about. The formula is shown so you can disagree with it.

Read what it couldn't check

That section carries two things: the data it went looking for and did not find, and the evidence that pointed the other way. A grey line at the very bottom of the case file adds a machine check: NLI ground-check: 7/10 claims entailed by their sources. Any caveat sits on the same line.

Develop your intuition

The caveat line is where a run admits it is uneasy. It shares that grey line at the bottom with the ground-check, and it is the fastest read in the whole case file. Four things show up there, and each one changes what you should do next.

Related resources

One constraint shapes most of what follows: when Vernais answers from YOUR CONNECTED DATA, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

The four levels

hypothesis only is a refusal, not a failure. Vernais found records that describe possible friction, but could not establish that they caused your outcome. So it withholds the verdict, keeps the facts it does have on screen, and sets the level to low. The caveat names what was missing. That is one of three things: outcome telemetry that clears the evidence contract, a claim chain the entailment check can verify, or a causal path its own cited evidence supports.

What moves the number

• The model's own call — the answer states a level. That is the starting point, not the verdict.
• The mechanism coverage check — if the error named in the root cause covers less than 80% of the worst-hit cohort, a causal high is capped at medium. The caveat prints the real percentage.
• The timeline check — if a symptom is dated earlier than the first dated candidate change, a causal high is capped at medium. The change cannot have caused something older than itself.
• The root-cause quote check — an entailment model re-reads the root-cause sentence against its cited sources. If none of them entail it, high drops to medium.
• The named-tool check — name a tool, and if under a third of the surviving evidence comes from it, high drops to medium and the caveat calls that angle tangential.
• The withheld verdict — the causal evidence gate refused the verdict, so the level is forced to low and the answer is labelled hypothesis only.

Why a causal answer can reach high

A causal answer can reach high. There is no blanket cap on why-questions. The cap is medium only when one of two named objections fires, and the caveat prints the reason.

• Coverage — the error named in the root cause covers less than 80% of the highest-failure cohort. The answer says so, with the real percentage.
• Order — a relevant symptom predates the first dated candidate change.

With neither objection, the cap stays high. There is a third guard worth knowing. The root-cause quote check would normally knock high down to medium, but it yields when the structure underneath is strong. Strong means all of: one version failing at 85% or more against a clean comparison build, at least 30 failures on it, and three or more tools reporting failures. That exception exists because the entailment model scores correct free prose near zero. Without it, the strongest cases would be marked down for being well written.

What stays the same when you re-run

• Stable — the level, the surviving records, and the leading hypothesis. The question breakdown is cached against the question itself, and the candidate explanations against a fingerprint of your graph, so the hosted-model steps before the jury cannot drift between runs.
• Varies a little — the wording. The models that write the prose are hosted and not byte-identical, so two runs can phrase the same finding differently.
• Changes on purpose — add data and the fingerprint changes. The cache misses, the candidate explanations are made fresh, and the number can legitimately move. If the answer changed, the data usually moved.

So a level that flips run to run is a signal, not a glitch. Check what synced between the two runs.

Initiative confidence is a different number

The confidence on an initiative's Validate tab is not the investigation badge in another costume. It answers a different question and it is built a different way. Do not read one as the other.

The six factors are scored out of 100 between them, with 60 as the decision-ready line. Each factor gets its own bar showing its score out of its maximum, so you can see which one is holding the number down.

Each hypothesis card is anchored to that measured number, then scaled by its own evidence balance. Receipts for it nudge the score up, capped; receipts against it pull it down, harder and also capped. The cards sort highest first.

Related resources

One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

What the message means

You asked a why-question. The investigation ran, then ended with this:

I couldn't find grounded evidence for this in your connected tools. The closest area was <topic>, but no records there actually speak to your question — so I won't guess. Try naming a specific tool, metric, or time frame, or rephrasing the question.

Here is the machinery behind it. Every candidate record is judged on its own by a small AI juror. The juror votes one of three ways: this supports the idea, this contradicts it, or this is neutral. To take a side, the juror has to copy a real quote out of that record. If the quote is not really in the record, the vote drops to neutral. Records that survive with a real quote become your evidence.

This message means zero records survived. Not few — zero. That is why there is no confidence badge on it. There is nothing to be confident about, so nothing is claimed.

Work through the causes in order

1. 1
   The brain has not been run here

   Open Data. If it reads No knowledge-graph data yet, the map does not exist yet. Syncing a tool brings records in raw. Corvex is the separate step that turns them into the graph Vernais reasons over. Go to Integrations and run Corvex once. It runs 15 stages and takes a few minutes. Running it needs an elevated permission, so ask your workspace owner if you cannot see the button.

2. 2
   The tool holding the evidence is not connected

   Open Data and read the Source filter. Each row carries the real number of records that tool put into this workspace's graph. If the tool you expect is missing there, Vernais has never seen its data. Connect it and sync it. Each workspace has its own graph, so a tool connected in another workspace does not count here.

3. 3
   The question named a tool that does not hold it

   Naming a tool folds all of its records into the search and seeds the search there. A named tool with no records in your workspace is dropped instead. Three tool names are ordinary English words — teams, linear, notion — so they count as tools only with context. Write in Linear or linear ticket, never the linear drop in signups. If you are not sure the evidence lives in that tool, drop the name and ask again.

4. 4
   The window has no data

   A live sync is capped, not a full history backfill. Each stream stops at 20 pages or 2,000 records. Analytics exports such as Amplitude and Mixpanel cover about the last 30 days. Ask inside the period you actually synced. To check what you have, open Data, filter to that tool, and sort by Most recent.

5. 5
   The question was too broad

   Vernais matches your wording against the topics it discovered in your own data. A question with no symptom and no subject matches nothing worth entering. "why is our product bad?" gives the search nowhere to push off from. Rewrite it with a symptom, a subject, and an anchor — the next section shows how.

Rephrase so the question can land

What this message is not

Questions people ask

Develop your intuition

These are starting points, not rules. When you hit this message, ask yourself:

• Was the question too broad — did I give it a symptom and a subject?
• Did I name a tool that does not hold the evidence?
• Did I ask for a cause when I only wanted a count?
• Does my wording use the words my data uses, or my own shorthand?
• Has Corvex ever been run in this workspace?
• Am I asking about a period my synced data actually covers?

Related resources

One constraint shapes most of what follows: when Vernais answers from YOUR CONNECTED DATA, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

A tool that cannot say "I don't know" cannot be trusted when it says it knows.

It will not invent a cause

Every candidate record is judged on its own. A small AI juror reads it and votes: does this support the explanation, contradict it, or neither? To take either side, the juror has to copy a real quote from that record. The system then checks that the quote really appears in that record.

No quote means the vote is thrown out. It drops to neutral. So a record cannot become evidence because it sounded right. Only because a sentence in it actually exists.

Why — an invented cause needs invented evidence. Invented evidence has no quote behind it, so it never survives the vote. This is the single check that makes the rest of the answer worth reading.

It will not store a cause as a fact

Vernais writes down what your data literally says. Two records touch the same version. One ticket names another by number. A charge failed at 4:02. Each of those has a receipt.

"The v2.2.1 deploy caused the promo failures" is in no record. Nobody wrote it down. So Vernais never stores it. The cause is worked out fresh, the moment you ask, from evidence gathered right then. Even the timeline you read is derived at question time.

Why — a stored cause ages badly. It gets copied into a doc, quoted in a review, and believed months after the data it came from moved. A derived cause carries its evidence with it, every single time.

It will not answer a data question it cannot ground

For any question about your connected records — a count, a list, a top 5 — the model has to call a data tool before it answers. It cannot answer from memory. It cannot guess a number. It cannot even tell you it lacks access without trying first, because the tools are its access. When a tool returns nothing, it says so in those words.

An investigation runs a harder version of the same rule. If not one record survives the jury, the run stops and tells you so. It reports that it found no grounded evidence, names the closest area it looked in, and says it will not guess. No confidence badge appears, because there is nothing to be confident about.

It will not take your word over your account

Two kinds of source exist here, and Vernais keeps them apart.

Those are all you, speaking through different channels. Your own site repeating your claim adds nothing. It is the same assertion wearing a citation. So Vernais will not web-search to settle who you are — a search there hands your words back to you and makes them look proven.

It will not argue, either. It says once that the account lists one thing while you are telling it another, points you to Settings → Profile, then answers your real question. Edit the account and the answer changes.

Why — your title shapes what the AI assumes about your work. A title it adopted because you repeated it would be a fact with no source, and it would outlive the conversation that created it. Asking a fourth time is not evidence.

It will not edit your code

Ask about the source and it plans, searches by meaning, searches for exact strings, reads files, and follows the call graph. It answers with exact path:line citations, and a checker on the server verifies each one against the files it actually opened.

It cannot change a line. This is not a rule it was asked to follow nicely. No write tool is registered, so there is nothing for it to call. Its shell access is an allowlist of read-only inspection commands.

Two more edges sit around it. Only a repo URL you typed in that message can be cloned — not one from earlier in the thread, and not one it found itself. And the whole capability is permission-gated: without codebase.analyze, one of the AI permissions an owner assigns, the code tools are left out of its belt entirely.

What it answers without hesitation

The refusals above are narrow on purpose. They cover your data and your identity. They are not a rule about knowledge in general, and reading them that way would make the tool useless.

Ask what a transformer is and it explains. Ask for an opinion and you get one. Ask about the news, a price, or a public company and it searches the web, reads the pages, and cites them with numbered chips. Tell it not to search and it answers from what it knows. It will not refuse the question.

The answer always tells you which of those worlds it came from. That is the whole point of the constraint, not a side effect of it.

Related resources

One constraint shapes most of what follows: when Vernais answers from your connected data, every claim has to trace back to a record it can point at. It will not invent a cause. A question it cannot ground in your data returns nothing rather than a guess. Questions about the outside world, or about general knowledge, it answers like any good assistant — and tells you which world the answer came from.

Who you are is a record too. Your account is that record. Vernais reads it on every message, and it does not take your identity from anywhere else.

Two kinds of source

Vernais sorts information about you into two buckets, and it never mixes them.

The second bucket is the surprising one. Your own website is not proof. It is you, speaking through a different channel. If you tell Vernais you are the founder, and then send it a link to your site that says you are the founder, it has one claim, not two. Nothing was confirmed. The link is the same claim wearing a citation.

The answer to "so who am I?" does not change because you asked a fourth time.

Why it holds the line

This is an anti-impersonation guard. Your name and title travel: they sit on your profile card, and they are attached to the things you create. A title accepted in a chat box would spread from there.

Picture the rule without the guard. Anyone in a chat box could type a title and be believed. Anyone could paste a link to a page they made an hour ago and be believed twice. Repetition would work, because a model that gives ground once tends to give ground again. So the guard is not about doubting you. It is about the fact that a chat message and a self-published page carry the same weight — and that weight is not enough to change a record.

What it does when you disagree

It says so once, plainly, and moves on. You will get something close to: your account lists this title, you are telling it that one, and it can only verify the account. Then it answers your actual question.

It will not tell you it is taking the correction. It will not promise to treat you as the new title going forward. It will not call your account stale. It cannot know that, and it cannot edit your account from a chat message. Saying otherwise would be a promise it has no way to keep.

That second line matters. A web search about you can only return what you published about yourself. It would hand back your own claim with a source number next to it. That produces confidence, not evidence — so the search is refused for you and allowed for everyone else.

Change the record, not the chat

Open Settings → Profile. Three fields shape how you are described:

• Display name — up to 80 characters. Shown on activity, on the member list, and on anything you create.
• Role / title — up to 120 characters. This is the title Vernais reports when you ask what your role is.
• About you — up to 300 characters. A short line about what you do, shown on your profile card.

Save, and the change is authoritative. It is not a request Vernais weighs against your chat history. It is the record. Changing the email you sign in with works differently — you confirm from the new address before it takes effect.

What it remembers between chats

Vernais keeps durable preferences across conversations. How you like answers shaped, what you care about, a standing goal, a constraint you work under. Those carry from one chat to the next, so you do not repeat yourself.

Identity claims never make that trip. When a turn ends, a background pass pulls out anything durable you revealed. A guard then reads each candidate before it is written. Any line that asserts your name, role, job title, or employer is dropped on the spot. The same guard runs again when memory is read back — so if a claim ever slipped in, it stops surfacing and heals itself.

So the split is clean. Your profile answers who you are. Memory answers how you like to work. Neither one can quietly become the other.

Related resources