sweep

The mechanised version of the vault's core mandate: connect optically insignificant datapoints to macro-level shifts. Corporate language is a leading indicator — managements adopt a term (capacity auction, asset light, inventory digestion, agentic) before the economics it names show up in reported numbers, and when several companies along one value chain adopt it within a few months, a schelling point is forming ([G-3]). This skill dates that curve from primary sources and names the adopters, producing an immutable Research note.

What the tool can and cannot do (probe-validated 2026-09-09):

• search_documents returns totalCount + company/event/document pointers per date window — but no snippets (text: null). The sweep therefore has a hard two-stage structure: search finds WHERE, summaries/deep reads establish HOW the term is used. No usage claim is ever made from a hit-list alone.
• The documentTypes filter is loosely honoured (transcript queries return slide hits) — always read each hit's actual documentType.
• Query syntax is fragile in compounds: an exact phrase AND'd with an extra term can silently 0-hit while each part alone returns hundreds. Keep queries to a single term or a single backtick-quoted phrase; express restrictions via the filter parameter (GICS, country, speaker.role, companyId), never via extra query words. A 0-hit result on a term you know exists means the query is malformed — simplify and retry once before trusting it.
• Terms have prior language communities. Validated live: "capacity auction" ran 45 hits in Jun-Sep 2025 vs 21 in Jun-Sep 2026 — a naive read says the term is dying, but the 2025 base is utilities (RWE, Constellation, Alliant: electricity capacity auctions), and the real signal is Nebius entering that community in 2026 with a new meaning. Raw counts track the blended series; the adoption read requires segmenting the curve by GICS and tracking community composition, not volume.

Arguments

/sweep "<term>" plus optional flags:

• --since YYYY-MM-DD — curve start (default 24 months back)
• --gics "<GICS name>" — segment/restrict to an industry (repeatable up to 2; e.g. "Information Technology" at level 1, "Semiconductor Materials and Equipment" at level 3/4)
• --country XX — ISO country filter
• --role CEO|CFO|Analyst — transcript speaker-role filter (who says it matters: CEO adoption ≠ analyst questioning)
• --types transcripts,slides,reports — default transcripts,slides
• --portfolio — restrict to the Vault Active watchlist companyIds (from .data/quartr_watchlist.json)

Multi-word input without quotes → treat as an exact phrase. Multiple unrelated terms → ask the user to run separate sweeps (comparability of one curve is the product; blended queries are uninterpretable).

Step 0: Pre-flight

1. Lock: vault-wide per _shared/preflight.md Procedure 1, timeout 10 min (writes one Research note; reads theses + graph).
2. Quartr probe (hard): per _shared/quartr.md §1 load search_documents,get_document_summary,get_event_summary,list_companies. Unavailable → abort (❌ /sweep has no fallback source).
3. Mental models (analytical tier — this skill renders judgements): read [[Generalist - Overview]] always; the matching Industry - / Lens - file(s) when --gics/the term's obvious domain puts a covered sector in play. Apply the READING PROTOCOL: an adoption curve that flatters an existing thesis is a disconfirmation trigger, not a victory lap.
4. Read _graph.md (adjacency primer per _shared/graph-primer.md) and .data/quartr_ids.json (build the reverse map companyId→vault ticker) in the same batch.

Step 1: Adoption curve (bounded window scan)

Bucket the lookback into windows: monthly for ≤12 months, quarterly beyond (cap: 14 window queries). Per window:

mcp__quartr__search_documents {query: "`<term>`", documentTypes: [...], startDate, endDate, limit: 1}

totalCount is the datapoint; limit: 1 keeps responses tiny. When --gics/--country/--role is set, pass it via filter (e.g. gics.level1:=["Information Technology"], speaker.role:=[CEO]). When the term shows cross-industry usage (Step 2 reveals a mixed community), re-run the LAST 6 windows with the target-industry filter to produce the segmented curve (cap: +6 queries) — the segmented series is the one the note leads with; the blended series is context.

Curve shape readings (hypotheses, not verdicts): sustained acceleration = diffusion; a step-function = one event echoing (check whether hits cluster on one company's reporting week); flat-high = established vocabulary, no inflection signal; decline in blended + rise in segmented = community handoff (the validated capacity-auction case).

Step 2: Who — the language community

One discovery query on the most recent window at limit: 100 (+1 page if totalCount > 100 and the run is unfiltered). Group hits by company; record documentType (verified per hit), event, date. Then:

• First-mention dating for the most thesis-relevant companies (vault names, graph-adjacent names, repeated hitters): per company, filter: "companyId:=[<id>]" with a wide date range, order implied by windowing — bisect with at most 2 windows per company to date their earliest use. Cap: 6 companies.
• Role composition when --role unset but transcripts dominate: one extra query with speaker.role:=[CEO] on the latest window — CEO-mouth adoption is the strong form of the signal.

Step 3: Vault cross-reference

Map hit companyIds → vault tickers (reverse ID map; unmapped hit companies are looked up by name against thesis/watchlist names before being declared outside-vault). Then against _graph.md:

• Adjacency propagation flag (the payoff): ≥2 companies that are graph-adjacent (same cluster, shared supplier/customer edges) adopting the term within a 2-quarter span → flag as a candidate schelling point, naming the chain and the implicated theses.
• Term present in hit companies but ABSENT from the vault's own thesis/sector vocabulary (grep the implicated notes) → vocabulary gap: the market is naming something the vault hasn't framed yet.
• Inverse check: the vault uses the term but the corpus barely does → the vault may be early (good) or idiosyncratic (check which, via Step 4 context).

Step 4: Context verification (the two-stage rule)

For the top ≤5 most load-bearing hits (by thesis-relevance, then recency): get_event_summary (events) or get_document_summary (documents) to establish how the term is actually used — new meaning vs incumbent meaning, offensive vs defensive framing, management-initiated vs analyst-prompted. A hit whose summary is null and whose usage matters → note it as unverified and recommend /transcript TICKER --event <id>; never guess. No figure or usage claim enters the Evidence section without this layer or a deep read behind it.

Step 5: Write the Research note

Research/YYYY-MM-DD - Sweep [term] - data.md — immutable record per Research-note conventions. Frontmatter: date, tags: [research, data, sweep + implicated tickers/sector tags], status: active, source: Quartr search_documents corpus sweep — query "<term>", windows <start>..<end>, source_type: data, publish omitted (Research notes don't carry it).

Body (all 4 required sections):

• ## Thesis Delta — which theses the curve touches and how (hypothesis language; trigger touches flagged per _shared/trigger-touch.md discipline — flag, never auto-action)
• ## Summary — the curve's shape and the community's composition shift, led by the segmented series; the mechanism the term names; what adoption timing implies
• ## Evidence — the curve table (window | blended count | segmented count), the adopter table (company | vault link/outside | first mention | role | verified context), each verified usage tagged [Quartr] with its deep link
• ## Contradiction Check — the strongest mundane explanation (reporting-calendar clustering, one-event echo, incumbent-industry base rate, Quartr coverage growth inflating late-period counts — state this one explicitly every time, since corpus size is not constant over the lookback)
• ## Source Excerpts — the ≤5 verified usages, quoted short with links

Wikilink implicated theses/sectors. No thesis Log writes (multi-thesis fanout is /sync's job — the note's links make it discoverable); remind the user to run /graph last.

Step 6: Release lock + report

Stats line (windows, total hits, community size, vault overlap), the curve in miniature, the adjacency-propagation flags, the deep-read queue, and suggested follow-ups (/transcript --event, /deepen, /monitor if the term deserves a standing alert).

Design constraints (xxx DO NOT VIOLATE xxx)

1. Counts are hypotheses. Every curve reading ships with its mundane alternative (Contradiction Check). Corpus-growth inflation is stated in every note.
2. Community composition beats volume. A blended count is never presented without the who — the validated failure mode is reading a community handoff as a decline.
3. Two-stage always. Hits carry no snippets; usage claims require the summary/deep-read layer. No exceptions for "obvious" terms.
4. Simple queries only. One term or one exact phrase; restrictions go in filter. On unexpected 0-hits, simplify once, then report the anomaly rather than working around it with compound guesses.
5. Bounded budget. ≤14 curve + ≤6 segmented + ≤2 discovery + ≤12 first-mention + ≤5 summary calls per run. A sweep that wants more is two sweeps.
6. No thesis writes. One immutable Research note; propagation stays with /sync, conviction with /status.