Mavera DocsWhat changed: the reasoning upgrade
The new platform doesn't just look different — it thinks differently. Here's what that means for your work.
Inspect the reasoning upgrade
Review a plan, sources, approval boundary, and final evidence trail.
Open the sample planned run in Mave.
Good afternoon
Choose a surface from the sidebar to begin.
- 1Open a planOpen the sample planned run in Mave.
- 2Inspect the planOpen the proposed steps before approving them.
- 3Inspect sourcesCheck which claims come from current sources and which are inferred.
- 4Approve the bounded runApprove the local plan and watch each step execute.
- 5Read the auditOpen the completed reasoning and evidence trail.
If you used the previous Mavera, the biggest change isn't the interface. It's the brain.
Old vs new, honestly
The previous platform ran on an earlier generation of models. The new platform runs on Fable 5 for orchestration, conversation, and judgment, with Sonnet 5 doing high-volume respondent work. What that generation gap buys you:
| You'll notice | Because |
|---|---|
| She plans before answering | On non-trivial questions you see a concise plan, assumptions, method, and cost before work runs. |
| Plans instead of reactions | Ask something big and you get a decomposed plan — audience, method, cost — that you approve before a credit is spent. The old platform picked one tool and hoped. |
| Honest refusals | Audiences now decline questions their profile can't support instead of hallucinating an opinion. This was the single biggest quality complaint on the old platform. |
| Follow-ups actually follow | The full conversation, results, and business context ride along on every turn. "Why?" after a study result gets a real answer about your result. |
| Panels disagree like humans | Respondent diversity is measured (lexical overlap, opener diversity, length variation) — synthetic groupthink gets caught and re-run. |
| Hours-long autonomy | Programs plan → execute → read findings → re-plan, dozens of iterations, with cited web sources. The old platform capped out at single tool calls. |
What this means in practice
Inspect the "why." On the old platform, explanations were plausible narration. Now a concise reasoning summary shows the plan, assumptions, and evidence used. It is an inspectable work trail, not hidden chain-of-thought.
Give harder tasks. The instinct from the old platform was to break work into tiny pieces. Stop. Hand over the whole job — "figure out our pricing page strategy and validate it with the audience" — and let the planner decompose it.
Push back. Fable 5 updates on argument. If a result looks wrong, say why — "that segment split doesn't match what we see in sales calls" — and it will genuinely re-examine rather than apologize and repeat.
The numbers are computed the same way — better inputs
One thing did NOT change: numbers still come from code counting recorded synthetic responses, never from the model writing statistics. What improved is the quality of the responses being counted — sharper persona adherence, less generic filler, measured diversity.
Where the old work went
Your old-platform personas, files, and usage history were migrated — audiences carry a MIGRATED provenance tag, files sit in the Files page library, and old usage shows in your billing view. Nothing was lost; everything got a better engine.