Real progress and timing
The progress screen now reports what the extraction is actually doing: when Pass 1 finishes, and each of the three Pass 2 groups as it completes. The Results page shows how long each pass and each group took and how much each group wrote. If the model fails to return a field, that field is flagged Not returned for review instead of looking silent.
Three groups at once
With a Claude model, both passes now run on the model you pick. Pass 1 stores the PDF in the cache, and Pass 2 then extracts the 79 fields as three groups running at the same time, each reading the cached PDF.
Claude Fable 5.2
Fable 5.2 is the default model. Fable 5.1 remains available as the previous default.
N/A status
A field that cannot apply to the plan, such as a safe harbor formula in a plan that is not safe harbor, is now marked N/A with its own count instead of being counted as certified. The reason and the quoted election are in the source panel, and the field exports as N/A.
Excel, PDF and JSON exports
Results now download as an Excel workbook, a real PDF report, an expanded CSV and a complete JSON record. Every field has a stable key. Two checkboxes control whether quoted sources and flagged fields are included; certified-only remains the default.
Saved plans
Each extraction is saved in this browser with its PDF, so you can reopen it from the Console without running it again. Saved plans stay on this device and can be deleted one at a time or all at once.
Source panel
Source now opens a panel on the right instead of a pop-up. It shows the extracted value, how it was read, the quoted text in its surrounding provision, the choices offered on a checkbox election, the PDF page itself, and the other fields in the same group. Ambiguous, conflicting and uncertain fields open the same panel from a Review button, with the reason they were not certified. Their values are still not exported.
Light mode, waivz.ai look
The console now matches the redesigned waivz.ai: white canvas, blue sidebar, NOX dividers and status lights. The old empty "waiting for upload" space is replaced by a working console.
Current models
Claude Fable 5.2 (default), Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra. Document chat runs on the Claude model that did the extraction, or on Fable 5.1 after a GPT run. Gemini has been removed.
Prompt caching
The PDF is cached on Claude for an hour after the first pass, so the Pass 2 groups and every chat turn read it from cache instead of re-sending it. The CACHE pill on the Results page shows how many tokens were read from cache.
Still here
Two-pass extraction, vendor detection, source evidence on every certified field, the Copy and HTML table exports and multi-turn document chat are unchanged.