radar-sweep

name	radar-sweep
description	The market-radar's sensing pass — sweep recent public signal (HN + Reddit, free, recency-windowed) on a company or theme and bucket every hit against the contract's watch signals (distress, ownership, momentum, valuation, AI-pressure). Use to scout a candidate or refresh the pipeline.

Thesis-driven: the hunt (band, regions, sectors, signals, weights, dedup sources) lives in a thesis file — theses/bending-spoons.yaml is the default and mirrors Bending Spoons' contract. To scout for another acquirer: copy that file OUTSIDE this repo, tune every field (band, regions, sector hints, signal keywords, weights, event queries, dedup csvs), and add --thesis <your file> to every command. The engine is fully generic; the thesis is the strategy — keep yours private. Start any session with python3 ${CLAUDE_SKILL_DIR}/scripts/radar_sweep.py --preflight (thesis + sources status).

Run the deterministic sensing pass, then judge — never the reverse. The hunt is COHERENT WITH THE SUBJECT: it buys formerly-public companies, carve-outs from listed parents, and private digital brands — so the radar works in three tiers, structured-first:

Tier 1 — STRUCTURED (EDGAR, free): the mechanical gates and ownership events for anything US-listed (including EU companies listed in the US, which file 20-F/IFRS — read natively):

• --universe CY2024 — the mechanical top of the funnel (US): every US filer's annual revenue that year (union of 3 XBRL revenue concepts, deduped by CIK, owned/pipeline names removed), filtered to the $50M–$5B band. ~2,500 names in-band for CY2024 — Bending Spoons' own funnel scale (~L2617); narrow it with --gates (sector + region) per candidate
• --universe-eu CY2024 [--countries IT,DE,FR] — the same net over the EU-listed universe, from the ESEF repository (filings.xbrl.org, free, keyless): enumerates every EU-regulated-market filing for the period + countries, fetches IFRS revenue per filing (capped by --limit, default 60 — raise it or narrow --countries to go deeper; enumeration is always complete), band-filters in native currency (indicative — mostly EUR, not USD-converted), removes owned/pipeline. Deep-dive any name with --company. This closes the EU half of the funnel the US frames endpoint can't see.
• --events --days 30 — recent 8-Ks with sale/divestiture language, digital-SIC filtered
• --gates "TKR1, TKR2" — per candidate: revenue band · HQ EU/NA · digital SIC → PASS/FAIL [derived]

Tier 2 — CHATTER (HN + Reddit, free): brand/theme sweeps bucketed on the watch signals, --rank for maturity scores. Corroboration and early smoke — never the primary evidence.

Tier 3 — PRESS (WebSearch, use it seriously): for private companies and EU-listed names EDGAR cannot see (most of Bending Spoons' >1,000-target universe is private — this tier IS the "manual analysis" Bending Spoons itself performs). Recipes — run 2+ and triangulate, date every claim:

• ownership: "<company>" ("explores sale" OR "strategic alternatives" OR "in talks to sell")
• parent divestiture: "<parent>" (divest OR carve-out OR "non-core") <asset>
• distress: site:layoffs.fyi <company> · "<company>" (layoffs OR restructuring)
• momentum: "<company>" (subscribers OR MAU OR users) (decline OR churn OR fell)
• valuation: "<company>" (valuation OR "down round" OR writedown)
• EU targets, local press: "<company>" (vendita OR cession OR Übernahme OR "te koop") Everything from tiers 2-3 is [to-validate — press only]; tier 1 values are [derived] from EDGAR.

Tier 1 extensions (live): --company "Legal Name" — the EU-listed + unlisted chain: GLEIF (identity, LEI, HQ country — free, keyless) → ESEF/filings.xbrl.org (structured IFRS revenue via xBRL-JSON — free, keyless; NOTE: repository coverage varies by country — FR/IT/NL/Nordics are good, DE is partial — so "no filing found" never means "not listed") → honest n.d. for private companies (no LEI/no ESEF: that slice is chatter + press + judgment; UK privates via Companies House once CH_API_KEY is set). --loyalty "Brand" — the brand-strength card: committed base (App Store rating volume/avg) · brand interest (Wikipedia trend %) · community (subreddit size where reachable) — proxies, monitorable by snapshot+diff; retention/NRR proper is non-public and enters at diligence. --forms --days 5 — ownership-change filings from the EDGAR daily index (SC 13D activist stakes · SC 13E-3 going-private · 25/15 delisting), sector-filtered per thesis. Tier 2 extensions (live): --interest "Brand" — Wikipedia-pageviews momentum (window vs previous window, % delta); --appstore "App" — rating volume/average as a committed-base proxy (iTunes, keyless).

Source roadmap (researched and endpoint-verified; not yet integrated — in value order):

source	gives	access	effort
filings.xbrl.org (ESEF)	structured IFRS revenue for ALL EU-listed cos (~25k filings, 28 countries)	free, no key, filings.xbrl.org/api	M (XBRL parsing)
UK Companies House	private-UK accounts (turnover where filed) + PSC ownership	free key, 600 req/5min	M
GLEIF LEI	HQ/legal form + parent-child ownership graph; the cross-source join key	free, no key, api.gleif.org/api/v1	S
WARN Firehose	50-state layoff notices, daily, bulk CSV/JSON	free (check ToS) — or self-host biglocalnews/warn-scraper	M
GitHub API	repo-activity decline for dev-tool targets	free, 5k req/h with token	M
Tranco (+ CrUX/Cloudflare Radar)	domain-rank trend for web-SaaS targets	free (Tranco keyless; CrUX/Radar free keys)	S-M
npm / PyPI stats	dev-adoption momentum	free, no key	S
Pappers (FR) / BRREG (NO) / PRH (FI)	private-company data per country	free tiers vary	S-M
growth rankings (FT1000, Deloitte Fast 500 EMEA, Inc 5000, Sifted leaderboards)	annual PRIVATE-company universe lists with revenue/growth indications; the signal is the DELTA — names that stop appearing = growth over, maturation begins	public annual lists (download/scrape once a year; Sifted partly paywalled)	S (annual diff)
Skip (verified weak): layoffs.fyi (fragile scrape + attribution), Polymarket (thin M&A coverage),
Google Play / Google Trends (no reliable free automation), SimilarWeb/data.ai/Sensor Tower (paid only).

Setup (once per shell, required by SEC EDGAR): export RADAR_CONTACT="Name email@domain" — the structured tier (--gates/--events/--universe) returns 403 without a real contact.

1. Run the sweep (from the repo root): python3 ${CLAUDE_SKILL_DIR}/scripts/radar_sweep.py "<company or theme>" --days 30 Several candidates at once → --rank "A, B, C": one maturity score per candidate (formula declared in the script header, [decision]), owned/pipeline names excluded automatically. Widen with --days 90 for slow-moving targets.
2. Read the brief: hits are bucketed by the market-radar contract's WATCH SIGNALS (capabilities/actions/market-radar.md) and ranked by engagement. Everything in it is press/social → [to-validate — press only]; never present a hit as a fact.
3. JUDGE (or spawn the market-radar agent with the brief in its prompt — it has no execution tools, so give it the brief; it drafts the propose-params): does the signal say buyable now? Which of the nine Target gates does the candidate plausibly pass?
4. If yes: propose via the engine — python3 -c "import sys; sys.path.insert(0,'mcp'); import engine, json; print(json.dumps(engine.propose('source', PARAMS), indent=2))" and STOP: admission into the pipeline is the human gate (engine.apply with the user's name).

WebSearch remains a complementary source for press the sweep can't reach — same [to-validate — press only] rule.