ADR: Scout scope boundary — acquire + extract, never interpret

• Date: 2026-06-12
• Status: Accepted (decided with Arijit in session; supersedes the implicit "Scout = target's own web presence" framing)
• Context: Scout had absorbed 9 intelligence use cases as stubs, raising the question of whether it was swallowing PRISM and the algolia-search-audit sub-skills. A boundary drawn by topic (company, hiring, finance) or by source domain (the target's own site) proved hazy — 10-Ks live on EDGAR, interviews on third-party sites, competitor signals on competitor sites, yet all are clearly "gathering."

Decision

The boundary is drawn by kind of work, not by topic or source domain:

One-line rule: Scout gathers and structures evidence; it never has an opinion.

Sub-decisions

1. Licensed/vendor data APIs stay out of Scout (SimilarWeb, BuiltWith, Yahoo Finance, Crunchbase). Scout's doors are open-web only — including public APIs that are merely doors to public documents (SEC EDGAR, RSS/Atom, sitemaps). Vendor-analytics MCPs remain in the skills/PRISM layer.
2. Document intelligence is IN, as a first-class use case: filings & transcripts — EDGAR + IR discovery, 10-K/10-Q section extraction (MD&A, risk factors), earnings-call and interview transcript acquisition, verbatim quote records. This is the largest scope addition vs. today and is explicitly extraction, not interpretation.
3. Quote topic-tagging is allowed, additive-only: deterministic keyword/topic tags (e.g. search, digital, personalization) may be attached to extracted quotes. Tags never filter at source — all quotes are always returned; consumers may ignore tags. Significance ranking is interpretation and stays out.
4. Scout is target-agnostic: "competitor intelligence" is the same acquisition contracts run against more targets (RunRequest.targets). Choosing the competitor list is interpretation and happens upstream.
5. The algolia-intel-* skills are committed consumers: as each Scout use case goes live, the matching skill's fetch logic (WebFetch/Apify) is replaced by a Scout call. Skills keep their reasoning prompts. This is a split of each skill (fetch half → Scout contract, brain half → stays), not a port.

Consequences for the use-case registry

• company, careers, news, investor, docs, website-quality, products — confirmed, scoped as acquisition contracts (fields mined from the matching algolia-intel-* skill specs, web-acquirable subset only).
• NEW: filings (filings & transcripts) — added per sub-decision 2.
• research — narrowed to "fetch + extract from given URLs"; open-ended research synthesis is out of scope.
• social — URL-verification mode only (ToS posture) until reviewed.
• prism — redefined as the prospect bundle: a composition that runs the acquisition contracts across a target list (prospect + competitors). No synthesis inside it.
• locations — remains descoped.

What Scout will never contain

Vertical classification, signal scoring, Algolia value mapping, business cases, playbooks, outreach campaigns, search-quality judgment (query relevance rating), competitor selection, any LLM-generated narrative about what evidence means.

Open (explicitly deferred)

• Monetization/productization direction (evidence-engine vs crawl-to-index vs agent backend) — the boundary above keeps all three open; decide when there is a reason to.
• Whether Scout's embedded live browser later hosts the interactive search-audit testing flow (judgment-laden; currently out).