git:20261008.1a28e75 to git:20261010.d31599c

33 added, 70 removed. Audit A to A.

---

name: lead-enrichment

- description: Enrich a lead from supplied account signals and produce a consent-aware, evidence-bound outreach recommendation.

+ description: Normalize factual person and company evidence for one confidently matched lead, preserving provenance, freshness, missing fields and conflicts.

---

# Lead Enrichment

- Turn known lead, account, and engagement signals into a reviewable picture of

- fit, confidence, risk, and the narrowest sensible follow-up. Good enrichment

- does not mean filling every field. It means making useful claims only where the

- evidence supports them and making “do not contact” or “learn more first” first-

- class outcomes.

- This is a supplied-signal synthesis skill. It does not scrape the web, query an

- enrichment vendor, update a CRM, or send outreach. Upstream systems own source

- collection; downstream routing and provider skills own action.

- ## When to use it

- Use `lead-enrichment` when a product already has bounded signals and needs a

- consistent, auditable assessment before routing a lead. It is useful for

- combining product activity, declared firmographics, CRM facts, and consent state

- without letting an agent silently invent the missing pieces.

- Do not use it to infer sensitive traits, reconstruct personal profiles, or

- manufacture permission from engagement. A recommendation is not consent and a

- source digest is not proof that Runx independently verified the provider.

- ## How it works

- 1. Supply the known lead fields and typed signals with unique source references,

- upstream SHA-256 digests, claims, and observation times.

- 2. Deterministic admission checks provenance, freshness, duplicates, consent,

- suppression, region, and channel constraints before synthesis.

- 3. Opt-out and do-not-contact signals stop the lane immediately. The model never

- gets to reason its way around them.

- 4. Synthesis builds the lead profile, fit assessment, recommendation, and risk

- flags using only admitted lead fields and signals.

- 5. Finalization rejects invented source references and any language claiming

- outreach permission, CRM mutation, or send completion.

- ## Inputs and result

- - `lead` contains known identity and account fields; unknown values remain

- unknown.

- - `signals` contain stable `source_ref`, `source_digest`, type, claim, and

- `observed_at` fields.

- - `as_of` and `max_age_days` establish a reproducible freshness decision.

- - `constraints` carry consent, suppression, region, and allowed-channel state.

- The result is a source-bound enrichment packet containing the supported profile,

- evidence references, fit and confidence, risk flags, and a non-delivery

- recommendation such as hold, review, nurture, or bounded follow-up. It can feed

- `lead-router`, which must preserve the packet digest and cannot broaden consent.

+ Build factual context for one person or company. Enrichment answers what is

+ known about the subject. `lead-qualify` decides whether those facts meet an ICP;

+ `cold-outreach` prepares copy only after contact eligibility is supplied.

- ## Stop conditions

+ The default `enrich` runner reuses supplied structured evidence. It makes no

+ provider request and spends no enrichment credits. An upstream declared

+ connector operation can supply evidence; Apollo is the planned first lookup

+ provider, whose credential presentation remains a native connector dependency.

+ Do not claim that running this skill performs an Apollo lookup.

- - Stop on opt-out, do-not-contact, prohibited region, or conflicting suppression

- evidence.

- - Return `needs_more_evidence` when the remaining signals are stale, too thin,

- or lack stable provenance.

- - Never infer protected or sensitive traits, personal circumstances, or missing

- firmographic facts.

- - Distinguish observed facts from inferred fit and express confidence honestly.

- - Do not claim a CRM update, enrollment, approval, outreach, or delivery.

+ ## Procedure

- ## Example

+ 1. Bind one subject reference and the professional fields actually needed.

+ 2. Supply source references, native or upstream digests, observation times,

+ explicit subject matches, and factual field/value pairs. A caller-supplied

+ digest is provenance, not independent provider verification.

+ 3. Reuse only fresh sources confidently matched to that exact subject. A weak

+ name match, stale observation or mismatched identity does not fill a field.

+ 4. Preserve conflicting values with their separate sources. Do not silently

+ choose the latest value or invent a merged person from several matches.

+ 5. Return `enriched`, `partial` or `needs_more_evidence`, supported facts,

+ missing fields, conflicts and the admitted source index.

- A lead has a verified company role, two recent product events, and an explicit

- email opt-in. The packet may identify a plausible product fit and recommend a

- bounded review for outreach, citing the three exact source references. It cannot

- infer budget or buying authority. If the same lead has a suppression signal,

- the result is do-not-contact regardless of apparent fit.

+ The runner performs deterministic normalization rather than an extra model

+ turn. Evidence acquisition and entity matching belong to the selected provider;

+ commercial fit belongs to qualification. It does not infer buying intent,

+ budget, consent, sensitive traits or private personal details.

- ## Agent task contract

+ ## Composition

- ### `lead-enrichment-synthesize`

+ Pass `lead_enrichment_packet` into `lead-qualify#qualify` together with the

+ actual ICP criteria. Preserve the native artifact digest. For another lookup,

+ request only missing professional fields with a bounded credit allowance and

+ the same subject identity. Never enable personal-email/phone reveal or a

+ waterfall merely to make an incomplete packet look complete.

- Use only admitted lead fields and signals. Return the supported profile,

- evidence, fit assessment, recommended action, and risk flags. Every evidence

- claim must cite exact admitted source references; label confidence as observed

- or inferred. Never infer sensitive traits or claim consent, CRM mutation,

- enrollment, or send completion.

+ The skill neither updates CRM records nor authorizes messages. Fieldwork owns

+ refresh schedules and continuing Work; connectors own exact provider requests,

+ credential custody and billing evidence.