instagram · gap map · 2026-07-23
After winding down the Instagram study, we listed hard-to-parse sources that might still feed a trader stack. Then we checked the live argos-dev catalog (~134 tools) so we don’t rebuild what the org already ships — and so we can see where a real add-on still lives.
Summary
The brainstorm still holds as a direction — filings + permits + physical sensors + ugly government text — but several of the obvious next bets are already ARGOS tools (especially SEC agreements, SEDAR mining, and SEC-extracted power / large-load demand). The moat worth building is not “more EDGAR” in the abstract.
Use MCP
Filings, 13F, insider tables, SEDAR mining, mars private markets, SEC-derived power, FDA bio, agreements.
Extend domain
Form 4 footnotes, 13D purpose text, ISO queues next to sec_power_*, NRC next to outages.
Greenfield
AIS / canals, WARN, Beige Book NLP, permits/claims, PACER, customs/BoL, VIIRS→asset join, call transcripts.
Already covered
These map cleanly to live tools. “We should scrape X” is mostly wrong here.
| Brainstorm idea | Status | What exists (examples) |
|---|---|---|
| SEC / 8-K / exhibits / material agreements | have | agreements_* · semantic_search · get_sentences · sections_* |
| Form 4 / insider activity | have | insider_* · insiders_of_company (tables, not free-text footnotes) |
| 13F / activist 13D–G | have | thirteenf_* · search_activist_positions |
| SEDAR / mining / NI 43-101-ish | have | sedar_* · mining_company_reserves |
| Funding / M&A from news | have | mars_* · search_news · mars_search_articles |
| Power / AI load / capacity (from SEC) | have | sec_power_demand_forward · capacity · supply_curve · outages · rate_cases · gas_views · gazetteer |
| FDA enforcement | have | bio_warning_letters · bio_inspections · CRL / pipeline tools |
| Macro releases + FRED vintages | have | macro_calendar · macro_series |
| Earnings dates / surprises | have | upcoming_earnings · earnings_surprise_history (not full transcripts) |
Implication for old experiments
D · 8-K exhibits is largely productized via agreements_* — remaining edge is thinner
(latency, coverage holes, non-EX-10 exhibits). A · interconnection is only partially covered:
filer-disclosed load/capacity in SEC text is not the same as raw ISO queue tables.
Partial
| Brainstorm | MCP has | Still missing |
|---|---|---|
| Power / AI interconnect | sec_power_demand_forward, capacity, outages, rate cases… | ISO interconnection queues (PJM/ERCOT/CAISO), FERC docket body as events |
| Mining permits / claims | SEDAR mining reports + reserves | EPA / BLM / provincial claims / air permits (non-filing geography) |
| Insider “plan language” | Form 4 tables / signals | Footnote free text (10b5-1 wording changes) |
| Activist narratives | Positions / 13D–G presence | Purpose-of-transaction prose as NLP events |
| Nuclear / outages | sec_power_outages (filer-disclosed) | NRC event notifications as a separate feed |
| Satellite (VIIRS) | Nothing in MCP — local stack just starting | Hotspot/geofence → plant/mine → ticker join |
discover_tools queries for AIS/canal, WARN, PACER, VIIRS, Beige Book mostly fall through to news, SEDAR,
or power-SEC tools — i.e. no dedicated catalog surface for those sources.
White space
Ranked by complement to existing MCP + hard parse + joinable to entities you already resolve.
Zero MCP footprint. Complements VIIRS (“sky”) with flow (“ships”). Clean event table: chokepoint, wait days, restriction, commodity class → crude / LNG / bulk names.
Why first: pure white space + pairs with work already started on remote sensing.
MCP already answers “what did utilities disclose about large load?” Gap is “what is in the queue /
status-changed this week?” — different latency and false-positive profile. Honest upgrade of experiment A,
not a duplicate of sec_power_*.
Why: power/AI demand is already a product theme; primary queues close the hole.
macro_calendar is release dates; FRED is series. No Beige Book district stress, no state WARN PDFs.
High text-parse edge; joins to sectors you already screen.
Why: underrated vs headline CPI; low overlap with ARGOS.
Not in ARGOS. Edge is hotspot / geofence → plant / mine → ticker, using MCP for identity
(resolve_*, sec_power_gazetteer, mining — not rebuilding filings.
Why: you already started VIIRS; the moat is the join, not the pixels alone.
| # | Idea | Note |
|---|---|---|
| 5 | EPA / state permits / mineral claims | Physical twin of mars funding for resources; MCP mining is filing-side only |
| 6 | PACER / securities litigation | No litigation domain in discover; early event risk vs Benzinga |
| 7 | Customs / BoL / trade stats | Entity resolution pairs with mars company resolve |
| 8 | Form 4 footnote + 13D narrative NLP | Thin extension on existing insider / activist tools |
| 9 | EIA / USDA / CFTC free text | MCP doesn’t expose commodity bulletin NLP |
| 10 | Earnings call / hearing transcripts | Dates/surprises exist; full claim extract does not |
Generic social / influencer text (validated weak on Instagram). Another news wire (Benzinga + mars wires already there). Full rebuild of 8-K agreements, SEDAR mining, 13F, or FDA bio.
Reprioritized
1 · Interconnection + FERC
Keep near the top — but as ISO primary data, not SEC re-extraction.
partial → build queues
2 · AIS / canal / ports
Still tier-1. Pure white space next to VIIRS.
build
3 · Beige Book + WARN + IRPs
Beige Book + WARN still open. IRPs partly in sec_power / rate cases — prefer WARN/Beige first.
build WARN/Beige
4 · Permit / claim / NI 43-101
Split: consume sedar_* for reports; build permits/claims as new domain.
split
5 · 8-K exhibits + Form 4 footnotes
Exhibits → consume agreements. Footnotes → small add-on.
mostly have
+ · VIIRS join
Promote into top tier — ARGOS won’t do the pixel→entity layer for you.
build join
Bottom line
Don’t rebuild EDGAR/SEDAR/mars/power-SEC. Build the physical world and the primary queues that sit beside those products:
Framing test (unchanged)
For each source: is the hard part parsing or access? Can it join to a ticker / asset / commodity within hours–days?
Does it change before Benzinga-class headlines? Can you store a clean event table
(time, entity, event_type, confidence, raw_span)? Instagram mostly failed the third test.
VIIRS, permits, queues, AIS, and exhibit footnotes can pass it — but exhibits are largely done already.
tools/list, discover_tools, sample tools/call).
Source notes also in repo docs/alternative-data-brainstorm.md.
Reports index: kee.staticpipe.com
· this page: /instagram/2026-07-23-alt-data-vs-argos-mcp.html