Coverage — Process & Spec

What this is. Coverage is chenzoo's equity-research desk: a home for AI-generated, sell-side-style research on companies the desk cares about. Each covered name gets a standing document set — filings digests, market research, a straw-man/steel-man thesis, a full three-statement model, and an initiating-coverage report — stored as markdown in this repo and rendered at /coverage. Documents are living: they get revised as new filings and material events land.

What this is not. Investment advice. Everything here is a research exercise produced by AI agents from public information. Numbers for private companies are estimates built from third-party reporting and are labeled as such.

---

1. The document set

Every covered ticker lives at content/coverage/<slug>/ and carries exactly these files:

Later revisions append to coverage.md (dated "Coverage update" sections) and to the changelog in context.md; they never silently rewrite history.

2. Disclosure tiers

The ingestion step depends on what the company actually files:

• Tier 1 — SEC reporting company. Pull the latest S-1 (if recent), 10-K, 10-Qs since the 10-K, and material 8-Ks from EDGAR. Convert each to a markdown digest: what it says, what changed, takeaways, and the numbers that feed the model. Actuals in the model come only from filings.
• Tier 2 — private / pre-IPO (e.g. SpaceX before its June 2026 listing). No filings exist, so filings.md becomes a disclosure dossier: tender-offer valuations, credibly-reported financials (WSJ, Bloomberg, The Information, etc.), regulatory filings that do exist (FCC, FAA, ITU), government contract awards, and stated company figures. Every number carries a source and date, and the model's "actuals" columns are explicitly labeled estimates.
• Tier 3 — foreign private issuer. Same as Tier 1 with 20-F/6-K equivalents.

A name can move tiers (an IPO moves SpaceX from Tier 2 to Tier 1); when it does, the model is rebased to filed actuals and the delta between desk estimates and reality is written up in context.md — that gap is part of the record.

3. Pipeline

Initiation runs as an orchestrated multi-agent workflow. Analyst tasks run on Sonnet; long-form writing runs on Opus — models are pinned per stage, never inherited.

1. Disclosure sweep (Sonnet, parallel): identify the tier, pull/digest filings or build the dossier, extract the financial history.
2. Market research (Sonnet, parallel lenses): industry/TAM, competition, headwinds, tailwinds, operating KPIs, valuation comps.
3. Model build (Sonnet): segment-level revenue drivers → income statement → balance sheet → cash flow, with an explicit assumption table and bear/base/bull scenarios. The balance sheet must balance and the cash flow must tie to the change in cash.
4. Thesis (Sonnet): straw man and steel man argued at full strength, then the baseline view with the 3–5 swing factors that would change it.
5. Writing (Opus): the initiating-coverage report and the context file, synthesized from stages 1–4.
6. Verification (Sonnet): cross-doc consistency check — numbers agree across documents, the model ties, claims carry sources. Failures go back to the responsible stage.

4. Maintenance protocol

On a new filing or material event for a covered name:

1. Digest the filing into filings.md (newest first).
2. Reconcile the model: restate actuals, adjust forward assumptions, note every changed assumption with its rationale.
3. Re-test the thesis: does the event feed the straw man or the steel man? Move the baseline only with a written justification.
4. Append a dated Coverage update section to coverage.md.
5. Log everything in the context.md changelog and bump meta.json's as-of date and (if changed) rating/fair value.

5. Model conventions

• Fiscal years labeled FY24, estimates suffixed E (FY26E). Units stated in every table header ($M unless noted).
• Revenue is built bottoms-up from segment drivers (e.g. launches × price, subscribers × ARPU), never as a bare growth rate on the top line.
• Assumption tables state the driver, the value per year, and the one-line rationale.
• Ratings: Overweight / Neutral / Underweight against a 12-month fair-value estimate (a fair-value range for Tier 2 names, since there is no traded price).
• Sourcing rule: any specific figure must carry an inline source and date on first use.

6. Storage & rendering

• Source of truth is this repo: content/coverage/PROCESS.md (this file) plus content/coverage/<slug>/*.
• scripts/build-coverage.mjs compiles the markdown into lib/coverage/docs.ts (generated, never hand-edited), which the fully-static pages under app/coverage/ render — same pattern as revops. Re-run the script after editing any content file.
• Adding a name = add the folder + meta.json, run the pipeline (§3), run the build script, redeploy. No database.

7. Coverage roster

A note on this name: coverage was scoped when SpaceX was assumed private (Tier 2). The desk's own disclosure sweep found the company had listed on Nasdaq as SPCX on 2026-06-12 (post the February 2026 xAI merger), so the initiation was rebuilt Tier 1 — anchored to the S-1/424B4 and 8-Ks, with the call expressed against the traded price. The tier-migration protocol above got exercised on day one.