Torus

The entity's neighbourhood: companies within `hops` links (suppliers, customers, competitors...), the links among them, and theses about them.

Lease the oldest research session waiting to be distilled into the graph. Returns its transcript digest, what was already recorded (don't duplicate it), the sources the agent looked at, and existing theses on its entities. Until `complete_distillation`, new records are attributed to that session.

Lease the next onboarded company to map from its annual report. Until `complete_mapping`, links you record are marked as curator links.

Start covering companies: add them, watch their filings, suggest starter packs, and queue a mapping job for each (run by the company-mapper subagent, which reads their annual reports and records suppliers, customers and competitors).

Write the interactive graph explorer HTML file (under TORUS_HOME/explorer.html) and open it in the user's browser.

Prior theses, claims and rejected hypotheses about entities and/or a question. Call this before starting research so you build on earlier work.

List what the most recent research sessions recorded (claims, theses, relations, sources, ruled-out ideas), newest session first, with each session's id and whether each record can still be undone. Read-only. Use it when the user asks what Torus saved, then pass the session id and records to `undo_records` to revert the unwanted ones.

Save one finding about one or more entities, tied to its sources. Record the source first with `record_evidence` and pass its id. A fact or metric without supporting evidence_ids is rejected, as is an inference with neither evidence nor derived_from. Returns the new claim with its id, which theses (`record_thesis`) and relations (`record_relation`) can cite. To fix an existing claim use `supersede_claim`, not a second record_claim.

Record a source. Re-recording the same URI returns the existing record.

Record an idea that was checked and ruled out, so it isn't re-explored.

Link two entities, e.g. TSMC SUPPLIES NVIDIA. Re-recording an existing link updates it (confidence, backing claim, dates) instead of duplicating it.

Save a new investment thesis and the claims it rests on. Record the claims first with `record_claim` and pass their ids. Torus uses the assumptions, subject and triggers to decide which new filings and news threaten the thesis and raise alerts. Returns the thesis with its id and status (active). To change an existing thesis use `update_thesis`; to log an idea you rejected use `record_rejected_hypothesis`.

Judge an alert after reading the event and the thesis's claims; the thesis is updated accordingly and the change logged with the alert.

Withdraw a claim that turned out to be wrong (kept in history).

How tracked theses have done against their benchmarks, their outcomes, and calibration: did the user's confidence match how often they were right?

Keyword search over current claims, evidence and theses.

Add a curated starter pack of companies and supplier/competitor links. Never overrides links research recorded.

Record whether a thesis turned out right, wrong or unclear (overrides the automatic outcome; used for calibration).

Correct a claim: the old version stays in history, the new one is current.

Write a thesis up as a memo (Markdown and a standalone HTML page): what it rests on with sources and quotes, triggers, events and verdicts, ruled-out ideas and history. Saved under TORUS_HOME/memos.

Where a thesis is weakest: thin or expiring assumptions, strong supplier or customer dependencies it never mentions, triggers that can't fire on events, and opposing views already in the graph. Input for /torus:challenge.

History of one entity, oldest first: claims recorded, corrected or retracted about it, theses on it created or changed, and events (filings, news) that name it. Each change has a time, kind, object id and one-line summary. Use it to see how the view on a company evolved; use `what_changed` for the whole graph and `get_entity` for its current links.

Put a thesis on the scorecard: its return against a benchmark from an entry price and date, in the direction of its stance.

Revert records the user doesn't want: claims are retracted (corrections go back to the earlier version), theses the session created are archived, links retired, ruled-out ideas withdrawn. History keeps everything.

Stop watching a company you added with `watch`. Removes it from the explicit watchlist only: a company that is the subject or a trigger entity of an open thesis is still followed. Nothing in the graph is deleted. Returns the updated watch list, as `list_watches` does.

Revise a thesis: status, confidence, the claims it rests on, and its triggers, in one atomic step. Everything is kept in its history. Also tunes its alerts: a threshold of its own, and mutes (e.g. no alerts for a supplier's routine 10-Qs).

Create an entity (or merge into the one with the same canonical_key).

Watch a company: its new SEC filings become events (on the next `torus watch run`). Subjects of open theses are watched automatically.

Everything recorded or revised across the graph since a point in time.

What if this happened? Which theses would it reach, through which links, and would it raise alerts. Records nothing.