Hunt

Hunt — the tells dossier

Hunt is the evidence-gathering half of the Hunt ⊣ Pitch adjunction (../forge/references/meta-meta-map.md). It maps Prospect → Gaps: extract the provable weaknesses (tells) of a prospect so that every later offer line traces to a demonstrated gap. A tell without evidence is a hypothesis; hypotheses are quarantined and never presented to a buyer. The dossier feeds synthesis only — never a client artifact directly (one-directional flow, ../forge/references/evidence-rules.md §6).

Binding law: ../forge/references/evidence-rules.md outranks speed and polish. Every client-relevant observation carries a URL, an access date, and a reproduction step. No fabricated observations, ever.

H1 · Typed contract

Fill these at intake; UNKNOWN is legal, never invented.

ENTITY:             str            # legal/trading name as found — flag UNCONFIRMED until human-verified
ENTITY_URL:         url|UNKNOWN    # canonical site
VERTICAL:           str            # industry/segment — drives which lanes matter
GEOGRAPHY:          str|UNKNOWN    # service footprint — gates competitor set + compliance jurisdiction
COMPETITORS_KNOWN:  list[str]      # named rivals supplied at intake (may be empty → derive)
WARMTH:             {cold|warm|hot}# from RELATIONSHIP_WARMTH — tunes tone + wedge aggressiveness
CAPABILITY_MENU:    list[str]      # what we sell — the monetizability target set
DELIVERY_DOMAIN:    str|UNKNOWN    # inherited from factory slot when run inside a pipeline
LANE_PRIORS:        ref|UNKNOWN    # shared vertical priors (FACTORY/atoms.md tell-patterns for this vertical)

Derived artifact (typed): TELLS_DOSSIER — a file at 01-research/tells-dossier.md with the exact schema below, plus rows appended to 01-research/UNKNOWNS.md and 01-research/NEEDS.md.

H2 · The dossier schema (exact — section order fixed)

01-research/tells-dossier.md MUST contain these sections in this order:

1. Entity block — ENTITY (with UNCONFIRMED tag if unverified), ENTITY_URL, VERTICAL, GEOGRAPHY, competitor set used (named + linked), hunt date, lanes swept, context budget spent.

2. Live-wedge list — 2–3 wedges (see references/wedge-scripts.md), each: wedge-id · the tell it shows · seconds-to-show · setup line · show-steps · landing sentence · source URL+date. At least one wedge must be showable in <60 seconds. If none exist, an explicit NO LIVE WEDGE — [reason] line (rare; usually a shallow funnel sweep — recheck).

3. Tells table — one row per evidenced tell, columns exactly:

   tell-id	lane	observation	evidence-URL	accessed	repro-step	screenshot-able	severity(H/M/L)	provability(H/M/L)	monetizability(H/M/L)	capability-map	composite

   Minimum: sweep all six lanes; ≥1 evidenced tell per applicable lane or an explicit lane clear — [what was checked] note. Every row's evidence-URL + repro-step must let a fresh agent re-observe the tell. Scores follow references/scoring-rubric.md.

4. Hypotheses (quarantine) — suspected-but-unevidenced weaknesses, columns: hypothesis · why suspected · triage route attempted · why unresolved · DO-NOT-SHOW. Nothing here may migrate to the tells table or any artifact until evidenced.

5. NEEDS / UNKNOWNS handoff — the residuals that survived the auto-triage ladder, in the evidence-rules format: | Unknown | Why it matters | Triage route | Status |, mirrored into 01-research/UNKNOWNS.md. Items that would materially improve the bid also append to 01-research/NEEDS.md tagged auto-resolvable / requires-human / requires-prospect.

Evidence-record format (every cell in the observation → repro chain): observation · URL · accessed yyyy-mm-dd · reproduction step · screenshot-able y/n. Full per-lane record specs in references/probe-playbook.md.

Acceptance checklist (pass/fail — all must pass)

• All six lanes swept; each has ≥1 tell or an explicit lane clear note.
• Zero rows in the tells table lack URL + accessed date + repro step.
• ≥2 live wedges, at least one <60s, OR a documented NO LIVE WEDGE reason.
• Every tell maps to a CAPABILITY_MENU item, or is marked phase-2 / partner territory.
• Competitor comparisons name real, linked competitors.
• Hypotheses quarantined; none leaked into the tells table.
• Entity carries UNCONFIRMED unless human-verified.
• Dossier survives a leak test: factual, respectful, no mockery.

H3 · Loop — Frame → Do → Review → Capture

• Frame (retrieval before probing). Grep FACTORY/INDEX.md by VERTICAL for the run's priors — specifically tell-pattern atoms (weakness signatures + where they hid + probe hit-rate) and triage-route atoms (public sources that resolved unknowns) — refresh any that are OVERDUE (../mint-atoms/references/refresh-policy.md) before use, and record every atom loaded in the ## atoms consulted block of 00-intake-digest.md (one line: ATOM-seq · type · claim — why loaded; when run standalone with no digest, log the same block in the dossier's Entity section). This is the same retrieval contract run-factory Stage 0 honors (mint-atoms ledger-protocol.md); the block MUST exist even if empty ((none — no match for <vertical>)). Then pick lane order by prior hit-rate for VERTICAL, set a context budget (default ~15 fetches/prospect; raise for compliance-heavy verticals), and derive the competitor set if COMPETITORS_KNOWN is empty.
• Do. Run the six-lane sweep per references/probe-playbook.md, batching probes (below). Log every observation as an evidence record as you go — never from memory.
• Review. Score with references/scoring-rubric.md; run the wedge-selection algorithm; quarantine hypotheses; run the auto-triage ladder on unknowns before parking any for a human; run the acceptance checklist.
• Capture. Mint atoms to FACTORY/atoms.md (via mint-atoms): tell-patterns (weakness signature + where it hid + which probe found it + hit-rate note) and triage-routes (the public source that resolved an unknown). A hunt that mints no atom did not complete.

H4 · Scale — hunting N prospects in parallel

• Isolation vs. shared priors. Per-prospect evidence is strictly isolated — one prospect's observations never cross-contaminate another's dossier (each writes its own 01-research/tells-dossier.md). Only vertical priors (tell-patterns, competitor sets, jurisdiction rules) are shared read-only across prospects in the same vertical.
• Probe batching. Within a prospect, batch same-source probes (all directory lookups, all review-platform pulls) into one pass to amortize fetches. Across prospects sharing a vertical, resolve shared lookups once (e.g. the vertical's regulator database index, competitor-of-record list) and cache to a run-local 01-research/_vertical-priors.md; dedupe repeated lookups against it.
• Fan-out. One subagent per prospect, context-complete prompt (typed slots + shared priors ref + lane order). Subagents write their own dossier + return a ≤400-word summary; they never score across prospects. The orchestrator ranks prospects by top-composite tell.
• Context budget. Default ~15 fetches/prospect; a batch of N prospects gets a global ceiling — when hit, finish scoring what is gathered and log the unswept lanes as NEEDS rather than fabricating.
• Retry policy. Failed/blocked fetch → retry once; on second failure, do NOT infer content — log the tell as a hypothesis with triage route attempted: fetch failed ×2 and move on. Never let a dead link become an assumed weakness.

H5–H8 · Encoded judgment, safety nets, invariants

Depth lives in three references (read the one you need):

• references/probe-playbook.md — per-lane probes, sources, what to measure/record, ethical + legal boundaries.
• references/scoring-rubric.md — anchored H/M/L for severity/provability/monetizability, composite rule, wedge-selection algorithm.
• references/wedge-scripts.md — example-agnostic live-wedge script patterns + handoff into the offer bijection.

Checkable predicates (safety nets): every tell has URL + date + repro step (else it is a hypothesis); ≥2 live wedges or a documented absence; hypotheses physically quarantined in their own section; entity UNCONFIRMED until human-verified; compliance findings dated (they decay fast).

Invariants restated where load-bearing (../forge/references/evidence-rules.md, factory-schema.md §5): evidence-or-unknown (never invent); auto-triage ladder before any unknown reaches a human (prospect materials → search → public stats → labeled inference → residual); one-directional flow (dossier → synthesis → artifacts, never dossier → artifact); respectful/factual tone; entity IDs stay UNCONFIRMED.

H9 · DONE + refusals

DONE when: all six lanes swept (tells or lane clear notes); acceptance checklist passes; wedges selected or absence documented; hypotheses quarantined; unknowns triaged with residuals in NEEDS/UNKNOWNS; ≥1 atom minted; entity confirmation status stated.

Refuse to: score or table a tell with no evidence record (it goes to quarantine); run deceptive recon — no pretexting, no fake purchase inquiries, no posing as a customer where prohibited, no accessing non-public systems (references/probe-playbook.md boundaries); test response times except via legitimately public channels; present the dossier as client-facing (it feeds synthesis only); mock the prospect.

H10 · Self-improvement hooks

Each hunt records: which probes yielded high-composite tells vs. dead ends (feeds lane-order priors), which triage routes resolved unknowns (feeds the ladder), which wedges landed or misfired in the meeting (negative wedge-patterns are atoms too). Feed corrections through mint-atoms; when a probe repeatedly finds nothing for a vertical, demote it in that vertical's LANE_PRIORS. Generalization check: a new tell-pattern must hold for ≥2 prospects in the vertical before it becomes a shared prior.