Deep research with a hypothesis ledger

What you'll get

Three capabilities that, together, turn research from a one-off answer into a compounding asset. Deep research produces a cited, fact-checked report on a question. The hypothesis ledger preserves what you concluded and why, the wrong conclusions included. The watchlist keeps scanning the topics you care about in the gaps between your requests. Used together, each piece of research makes the next one cheaper.

The steps

1. Ask a researchable question.

/marg:route deep research: is the Indian D2C skincare market actually growing, or is it funding-driven noise?

Marg fans the question across sources, reads them, checks the claims against each other rather than trusting any single one, and synthesizes a cited report. A vague question gets narrowed with you before any of that starts, because "research skincare" can only produce a survey, while a sharp question produces a verdict you can act on.

2. Log what you concluded.

/marg:research-log record: hypothesis that mid-market D2C brands churn from agencies within 6 months

A ledger entry states the hypothesis in falsifiable form, naming both what would confirm it and what would disconfirm it, along with the evidence tier behind it, the sources, and a confidence band. From there the entry earns a verdict over time: validated, invalidated, or nuanced.

3. Keep the invalidated ones. This is the counterintuitive habit, and it is the one that pays. When a hypothesis dies, the ledger records why it died, and that recorded failure is reusable knowledge rather than a dead end. Six months on, when the same idea walks back into a strategy meeting, the ledger answers it in seconds with the original evidence still attached, so you stop re-litigating settled questions.

4. Put standing topics on watch.

/marg:research-watch watch: AI agent pricing models in SaaS

Watched topics get swept on a cadence, and an anomaly, a topic drawing far more attention than its apparent significance should, gets flagged with context and marked unverified until something confirms it. You can wire the sweep onto a schedule, but nothing runs on its own unless you set that up.

What comes back

Reports cite their sources, and every load-bearing claim carries an evidence tier, so you can see at a glance whether a conclusion rests on what people said in a survey or on what they did with their budgets. That distinction is the whole point. Stated intent and revealed behavior routinely diverge, and the tiers keep you honest about which one you are actually reading before you bet on it.

Variations

• Quick lookups: not every question needs the full pipeline. Ordinary ones route to a lighter researcher automatically, and the deep machinery only engages when the stakes or the ambiguity earn it.
• Ledger reviews: /marg:research-log with no arguments rolls the ledger up: verdict counts, which research angle produced the most validated hypotheses, and any open entry that has gone stale.

If something goes wrong

• The report hedges everything: the question allowed it to. Attach a decision the research has to serve ("we enter this market, or not, by July") and rerun, because research aimed at a real decision cannot hide in qualifiers.
• Two sources contradict each other: that is the system working, not failing. The report surfaces the conflict with both citations instead of averaging it into mush, and the verdict says which side the weight of evidence favors.
• A watched topic never produces anomalies: it is either genuinely quiet or its baseline is set too high. Ask for the watch entry and loosen the threshold.