Upload the draft disclosure script, investor presentation, press release or transcript for the call you are preparing for. Prediction is then conditioned on the numbers management is about to read out — which is the only way to anticipate arithmetic follow-ups on a bridge that has never been disclosed before. Nothing is written to the archive.

Pin an answer, a source passage, or a predicted question from anywhere in the workspace — they collect here so the brief for this call is one place, not scattered across chat history and search results. Pinned items persist on the server, so they survive a reload.

Each test quarter is scored separately against a training cutoff that excludes it and everything after it — predicting Q1FY27 trains through Q4FY26, predicting Q4FY26 trains through Q3FY26. That is the production setting: before a real call you do have every prior call. Scores are never averaged into a single headline, and the spread between quarters is reported as a first-class number.

The Overall layer ranks candidate topics with five named weights (anomaly, disclosure, momentum, base rate, drill-flag). This never edits those defaults directly — it runs the SAME held-out quarters twice, current weights vs your candidate, and only reports PROMOTE if the candidate doesn't regress recall or fall below the existing promotion-gate floors.

Reads this run's topic_ranked_low misses (the one failure category a weight change can fix — see Error Researcher above), works out which single weight would close the most of them within a safety cap, and backtests that ONE candidate the same way the Framework Loop above does, plus a bootstrap confidence signal on the recall delta. Every proposal is logged below with status pending review until a human approves or rejects it. Approving never edits this app's production weights by itself — that stays a deliberate step an engineer takes by hand afterward.

Topic recall answers "was the bucket on the brief". This answers "did we anticipate what they actually asked" — each analyst's real question is split into individual concerns and judged 1.0 / 0.5 / 0.0 against our framed question text. A topic can be a perfect hit while the actual question is about something else entirely; this is the number that would have caught that. Costs real LLM calls across every held-out analyst in both test quarters — only run this deliberately.

Preview parses the PDF and shows what would be extracted — quarter, call date, analysts and firms, and the extraction-quality monitor's fuzzy-match flags against the canonical registry — writing nothing. Commit saves it into the archive, recompiles the dataset and graph, and refreshes every view.