How it works

The short version

The Brain turns the research you add into one read per asset class — bullish, bearish or mixed — and keeps the evidence for and against it one click away.

1. 1. You add research

   Paste a note in the browser, or save it to the Obsidian vault inbox.

2. 2. Claims become signals

   Claude splits the note into up to 8 signals, each with an asset class and a direction from −5 to +5, and writes a short decision summary.

3. 3. Signals form themes

   Each note’s signals are grouped by asset class into themes, scored for conviction (0–10) and polarity (−5 to +5).

4. 4. Themes become a read

   Each asset class gets a verdict from the conviction behind its bullish and bearish themes, with both sides shown.

The pages

Six pages in the sidebar, plus two ways to open a single research note.

• Today /brain

  The decision brief: the market regime, a one-line headline and evidence agreement; this week’s numbers; Actions to review; What changed in the last 7 days; a card per asset class; and the newest research. Export brief downloads it as Markdown.

• Themes & signals /brain/verdicts

  One read per asset class. Pick one to see its agreement, its bullish and bearish weight, a decision takeaway, and the supporting evidence beside the counter-evidence. Tabs narrow the list to Changed recently (new evidence in 7 days) or High conflict.

• Compare /brain/compare

  Two notes side by side, asset class by asset class: where they agree, where they differ, what only one covers, and the catalysts both mention. It opens on the two newest notes.

• Catalysts /brain/calendar

  Events named in your research, under Upcoming, Needs date review or By type; open one to see the notes behind it. Sync calendar downloads the upcoming ones.

• Research library /brain/timeline

  Every note, newest first, with its verdict, signal count and age. Filter by text, asset class, verdict or date added.

• Add research /brain/paste

  Paste a note, or drop in a .txt or .md file, for analysis — see Adding research.

• Evidence drawer opens over any page

  Click a research note anywhere — a row, an evidence card, a search result — and it opens in a side panel: the decision, then Signals, Source text and Themes tabs. Ctrl- or ⌘-click opens the full page in a new tab instead; Esc closes the panel.

• Full note page /brain/paste/<hash>

  The same note as a page, adding the asset classes it touches, its content hash, a Compare button and Older / Newer links (← and → keys).

How to read the numbers

Claude scores each signal once, when its note is added. Everything built on top — theme scores, verdicts, agreement, the regime — follows the fixed rules below.

Verdicts

Research noteBULLISHMIXEDBEARISH

The average conviction delta of its signals: +1 or more reads bullish, −1 or less bearish, anything in between mixed. Compare applies the same rule to each asset class within a note.

Theme

Set by its polarity, with the same ±1 thresholds. A theme in between is neutral: it is listed under “Neutral / no clear direction” and left out of the weights and agreement below.

Asset class

Add up the conviction (0–10) of its bullish themes and of its bearish themes — its bullish and bearish weight. If only one side has themes, that side wins. If both do, the heavier side wins unless the two are less than 3 points apart, which reads mixed: 16.0 vs 12.0 is bullish, 14.5 vs 12.0 is mixed. No directional theme at all also reads mixed.

The evidence behind a read

Supporting · opposing

Themes on the same side as the asset class’s verdict, and against it. A mixed read has no side to support, so it shows bullish and bearish counts instead — never “N sources agree” when they don’t.

Agreement

The dominant side’s share of the directional weight: the larger of the bullish and bearish weights ÷ both together. 50% is an even split, 100% one-sided; 13.0 vs 7.0 is 65%.

High conflictHIGH CONFLICT

Both sides present and agreement under 60%. The lean is still shown, but treat it as provisional. Contested reads on Today counts these.

BreadthHIGH BREADTH

How many distinct research notes stand behind an asset class’s themes: High is 5 or more, Medium 3–4, Low 2 or fewer. Shown in place of High conflict when a read isn’t contested.

On Today

Evidence agreement

Agreement across every asset class at once: all the dominant-side weight added up ÷ all the directional weight.

RegimeMIXED / DEFENSIVE

Counts only the risk assets — Indian equities, global equities and crypto. More bearish reads than bullish: RISK-OFF, or MIXED / DEFENSIVE when any asset class is bullish. More bullish than bearish: RISK-ON, or MIXED / RISK-ON when any asset class is bearish. A tie: MIXED.

Actions to reviewCONSIDERREDUCEWATCH

Ranked by priority = total conviction weight × freshness: ×1 when the newest theme is at most 7 days old, ×0.75 up to 30 days, ×0.5 after that. The five highest are shown. CONSIDER marks a bullish read that isn’t contested, REDUCE a bearish one, WATCH anything contested or mixed.

What changed

The last 7 days, up to six items in this order: verdict flips (the read recomputed without this week’s themes, by the same rule), agreement moves of 5 points or more, first reads (every theme in the asset class is new), catalysts mentioned again, and new research.

This week’s figures

New research and Signals extracted count the last 7 days; the small +/− figure compares them with the 7 days before.

Catalysts

Catalysts are found by keyword matching, not by Claude: the Brain looks for the wording below, and for dates, in each note’s decision summary and source text.

• Listed when the decision summaries of at least two of the 60 most recent notes mention it; the eight most-mentioned are kept. A date that two summaries share, with no event named, is listed as “Unnamed event on …”.
• Dates are read only when written as a full month name and a day — “June 17” or “June 17, 2026”, not “Jun 17” or “17 June”. Any such date in a note that names the catalyst counts for it.
• Without a year, a date is read against the day its note was added and placed in the year that puts it nearest that day, so “January 15” in a November note means next January, while “June 17” in a September recap stays in June.
• Several dates: the next one from today on is used; if all have passed, the most recent.
• Past dates are never shown as upcoming. The catalyst moves to Needs date review with a PAST DATE tag, beside catalysts whose notes name no date (NO DATE).
• Sync calendar downloads the upcoming catalysts as all-day events (brain-catalysts.ics), one per catalyst, with its notes in the description. Past and undated catalysts are left out, and the button is disabled when nothing is upcoming.

Adding research

There are two ways in. Both run the same analysis.

In the browser

• Source is required — the analyst, bank or channel. Title is optional; the note is saved as “Source - Title”.
• The research: 100 to 14,000 characters, the part Claude reads, so split anything longer. You can also drop in a .txt or .md file; a source: line in its front matter fills in Source.
• Save draft keeps your text in this browser; it is cleared once the note is added.
• Analyze research usually takes 30–60 seconds: Claude extracts up to 8 signals, the themes are regrouped, Claude writes the decision summary, and the Obsidian vault export is rebuilt. If it fails before anything is saved, the page says why and keeps your text.

From Obsidian

• Save a .md file in the vault inbox on the server, /home/ubuntu/brain_vault/inbox/. A watcher checks it every minute and sends each note through the same analysis — no browser needed.
• source: in the front matter names the note (there is no title field); without it, the file name is used.
• Bodies under 100 characters are skipped, and only the first 14,000 characters are sent. README.md is ignored.
• Handled files move to inbox/processed/ with a timestamp — even when the analysis reported a problem, so if a note never shows up, check the dashboard log.

Duplicates

Each note is identified by a hash of its text (SHA-256). Adding the same text again, by either route, adds no signals and recomputes nothing — the page says so and links to the note. An edited copy counts as a new note, so its evidence would be counted twice.

Keyboard shortcuts

Single-key shortcuts pause while you type in a field or the drawer is open. For the “go to” keys, press g, then the letter within about a second. ? shows the list on any page.

Single-key shortcuts on (/, ?, g + letter). Turn off if you use speech input; Ctrl K and Esc always work. Saved in this browser.

Glossary

Research note

One piece of research you added: its text, Claude’s decision summary and the signals taken from it. Also called a paste.

Signal

A single claim from a note, tagged with an asset class, a conviction delta, a strength and a confidence. A note yields up to 8.

Theme

Related signals from one note about one asset class, grouped under a title with a conviction and a polarity. A note that argues strongly both ways can get a bullish and a bearish theme in the same asset class.

Asset class

The market a claim is about, judged by whose price it concerns: Indian equities, Global equities, Bonds & rates, Gold & commodities, Crypto, Real estate, or Options & volatility.

Conviction delta

A signal’s direction and size, from −5 (strongly bearish) to +5 (strongly bullish).

Strength

How strong a signal is, from 1 (weak) to 10 (strong).

Confidence

Low, Medium or High: Claude’s confidence in the signal. It weights the theme scores — ×0.6, ×1.0 and ×1.4.

Conviction

A theme’s score from 0 to 10: the average strength of its signals, weighted by confidence and recency. Shown in words as very high (8+), high (6+), moderate (4+), weak (2+) or very weak. An asset class’s bullish and bearish weights add these up.

Polarity

A theme’s direction from −5 to +5: the average conviction delta of its signals, weighted the same way.

For the technically curious

• API referenceThe JSON endpoints and downloads behind these pages.
• Prompt reviewEach note’s input beside the signals and decision the two prompts produced.
• Classic viewThe previous single-column layout, kept as it was.
• Design systemThe page shell on its own; the component gallery is still a placeholder.

Every number on these pages comes from dashboard/brain_views.py — pure functions, covered by tests/test_brain_views.py. The routes live in dashboard/brain_app.py.