ARGOS · Industry Extraction Program

Industry Extraction Roadmap

A plan to systematically instrument industries with deterministic extraction — because that's the one lever that makes every report and agent in a domain measurably deeper.

v1 draft · prioritization needs Raul's client-demand input

Why

The controlled test showed report depth is set by structured data, not the model — adding one bio-extraction section lifted a report +50%, with the highest-value facts. The model is rented; the extracted graph is owned. So the strategy is simple: instrument industries, one at a time, and the depth compounds across the platform.

report depth = domain's structured disclosure × how much we've extracted

How we prioritize

Order each industry by three factors, with the first dominating:

Client demand / revenue — do we have (or want) big clients there? This should drive the order.
Breadth — how many public names? More companies = more reports = more ROI per unit of extraction.
KPI-richness — does the industry have quantifiable operating metrics worth extracting? (Most do.)

The roadmap

Industry	Status	Key metrics to extract (as time series)
Biotech	Done	Pipeline (assets/phases/targets/indications), FDA & CRL history, de-risking score, phase-advancement base rates. +50% proven.
Mining	Done	NI 43-101 reserves & resources (grade, tonnage, jurisdiction). Next: production & AISC as series.
Cruise	Pilot	Occupancy %, ALBDs (capacity), net yield, net cruise cost ex-fuel, advance bookings, newbuild pipeline, onboard rev/passenger. Client-driven (Carnival).
AI theme	Theme	Cross-sector (semis + software + services), not SIC-routed. AI-attributed revenue mix, data-center capex & capacity (MW), backlog/RPO, customer/supplier concentration, power draw. Novel graph: usage/token & product-adoption (transcript-mined, not 10-K), model/chip landscape (who runs what, open vs. closed), AI-risk taxonomy (model-dependency & local/open-model reliance, regulatory). Highest interest.
REITs	Tier 1	Occupancy, FFO/AFFO, same-store NOI growth, cap rates, lease expirations, debt maturities.
Banks	Tier 1	Net interest margin, deposit growth/mix, loan growth, credit quality (NPLs, charge-offs, reserve coverage), efficiency ratio, CET1.
Energy E&P	Tier 1	Proved/probable reserves, production (BOE/d), decline rates, F&D cost, reserve life, hedge book.
Insurance	Tier 1	Loss ratio, combined ratio, premium growth, prior-year reserve development, investment yield.
Semiconductors	Tier 2	Capacity/utilization, node mix, ASPs, book-to-bill, inventory days.
Airlines	Tier 2	RASM, CASM (ex-fuel), load factor, ASMs, yield, fuel cost/gallon.
Retail	Tier 2	Same-store sales, traffic vs. ticket, store count, inventory turns, e-commerce mix.
Software / SaaS	Tier 2	ARR, net revenue retention, gross/net churn, CAC payback, RPO, rule-of-40.
Telecom	Tier 2	Subscriber net adds, ARPU, churn, capex/sales.

Tiers are a starting guess by breadth + KPI-richness. Client demand should re-order this — a Tier 2 industry with a marquee client jumps the queue (that's exactly why cruise is a pilot).

Industries vs. themes. Most rows are industries — one SIC code, cleanly auto-routed (cruise = 4400, biotech = 2834). AI is a theme: it cuts across semiconductors, software, and services, so no SIC captures it — it routes off a curated universe + mention detection, and its richest signals (token/usage, adoption) live in earnings-call transcripts, not 10-Ks. Themes are their own track (GLP-1, electrification, quantum would follow); AI is the flagship, and interest is highest.

Two rules for every industry team

Time series, not snapshots. Extract each metric across 5–8 periods and pre-compute YoY / deltas / CAGR in code. Models narrate trends well but compute them badly — hand them clean trended data, never make them do the arithmetic.
Structured events, not raw news. A classified event timeline (M&A, guidance changes, management turnover, litigation, regulatory actions — plus industry-specific types) is deterministic and narratable. Build it once as a shared layer + a thin per-industry taxonomy; don't reprocess raw news per team.

The playbook (how we stand up an industry)

Charter — name the KPIs and event types for the industry (this doc's rows are the seed).
Extract — dig the filings, compute each KPI as a time series, resolve entities to the graph.
Expose — ship MCP tools mirroring the bio pattern (e.g. one tool that returns the metric series for a ticker).
Wire — add one report section to the builder, SIC-routed to that industry (a one-liner).
Retest — regenerate the report and re-score data density vs. the generic baseline. Target the same ~+50% lift the bio layer delivered. The scoring harness already exists.

Prepared by inferno · v1 draft for review. Sequencing here is by breadth & KPI-richness only — client demand should drive the final order. Companion to the extraction thesis (kee.staticpipe.com/notes/extraction-moat.html).