Thesis Tracker

Thesis Tracker

Two verdicts per update, because they move apart: the company thesis says whether the business is doing what we underwrote, and the security call says whether the stock is a decision we can act on today. A business can improve while the stock gets worse, since expectations rerate faster than evidence arrives, and a weakening business does not license a trim when we hold no price, no valuation frame and no position context. One status column hides both cases.

Evidence labels, source tiers, staleness, the readiness posture and the intake limits: .agents/skills/research-conventions/SKILL.md, read before the first deliverable.

The tracker is append-only. Prior pillars, thresholds and ledger rows stay as written; a row that no longer holds is marked superseded or stale with its date, or contradicted by a named later row, and the replacement is a new row. Drift is only visible against a history that was not edited, and catching drift is what the tracker is for.

Step 1: Take the mode, then frame the thesis and its pillars

The mode comes from what the user supplies, and its default runs without asking.

Three to five pillars, each a claim that evidence could kill. A claim nothing could disconfirm is a preference, not a pillar.

Every pillar carries these fields:

Provenance. Every pillar, threshold and action trigger is labelled inherited (written in the original underwriting), draft (our proposal, awaiting confirmation) or approved (an agreed monitoring rule). A pillar or level we introduce stays draft until the user confirms it, so an analyst-chosen exit price reads as a proposal rather than a mandate rule. This tracks approval, not support: a pillar drawn from a filing is fact on the evidence labels and still draft until someone signs off on it.

Blank shell. With no thesis supplied, deliver the empty structure and ask for the minimum inputs: ticker, direction, horizon, the two or three claims the position rests on, and any thresholds the user already runs. Fill nothing from memory, and say plainly which fields are waiting.

Done when the mode is named, every pillar carries all thirteen fields or names the field as missing, every pillar and threshold carries a provenance label, and no field holds a figure without a source.

Step 2: Append to the evidence ledger

One row per data point, appended, never rewritten:

Signal scale, used as a qualitative discipline: strongly confirming, confirming, mildly confirming, neutral, mildly weakening, weakening, strongly weakening, plus mixed, invalidating and untested. Read the distribution of signals across a pillar and state the direction in words. Summing them into a composite score invents precision the evidence does not carry.

Evidence-quality parse. A headline beat becomes evidence only once it is decomposed: volume against price against mix, cost actions, tax rate, share count and buyback, currency, one-time items, KPI definition quality, the shape of guidance, revisions beyond the next quarter, cash conversion. Name which component carried the beat and whether it recurs. A beat that came from tax and share count leaves every operating pillar untested.

Management credibility. Commentary earns weight when it is quantified, consistent with what was said last quarter, specific about the mechanism, and candid about what went wrong. Vague optimism, a changed KPI definition and selective disclosure earn none. Write which of the two you are looking at.

Accounting red flags. Each one lowers evidence quality on the pillar it touches and opens a dated follow-up: a KPI definition or disclosure change, non-GAAP adjustments growing as a share of earnings, a revenue-recognition change, receivables or inventory building faster than sales, cash conversion falling away from reported earnings, a spike in capitalised costs, a segment restatement, an auditor or CFO departure, related-party transactions, and "one-time" charges that recur.

Action. The verb comes from the closed vocabulary in .agents/skills/research-conventions/references/judgment.md, which also states the inputs each verb needs before it is available. The tracker adds no verbs: on a short leg the reader's word for exit is cover. Its two workflow outcomes, update model when the change lands in a line item and escalate when a threshold in step 7 fires, are steps the tracker takes rather than position actions, and they sit beside the verb rather than in its slot.

Done when every new data point is one ledger row naming its pillar, its signal and its evidence quality, prior rows are unchanged, and every action verb has its inputs in hand.

Step 3: Status the pillars and reconcile the aggregate

Each pillar takes exactly one of eight values:

Reconciliation. The aggregate follows the core pillars. One core pillar at impaired with an aggregate more benign than watch requires an evidenced override, written as one sentence naming what offsets it and the evidence behind the offset. Two or more core pillars at impaired set the aggregate to impaired. A core pillar at broken or changed sets the aggregate to the same. This is the anti-drift mechanism: without it the aggregate sits at intact for a year while the pillars underneath it rot.

The scorecard the reader sees is one row per pillar, in this shape:

Scoring honesty. Weighted pillar scores and conviction charts appear only when the user's own method defines the weights. Absent that, the aggregate is one of the eight words plus the sentence that justifies it.

Done when every pillar carries one of the eight values, the aggregate is reconciled or the override sentence is written, and no numeric conviction score appears that the user did not define.