Company Signals

End users are marketing agencies doing cold outreach. Sonar surfaces companies worth contacting because of a signal, then scores each lead by how strongly that signal correlates with marketing budget or a need the agency can address. Every search discovers signals live — they’re perishable, so Sonar never trusts a cached one. The global intelligence cache only spares re-enriching the slow-moving company facts: the first org to learn a company pays once, and every later search skips its enrichment for a fraction of the cost.

The whole run as a closed loop. A brief leaves the frontend, the API opens a run on the Claude Agents Platform, which discovers and resolves signals then fans out per company to the enrichment agent - the agent reads the intel cache or runs the waterfall and writes firmographics back. Results return along two paths: per-field patch frames stream live to fill the table (emit → run-SSE → apply) while resolved cells persist durably to the database. A later search skips enrichment for any company already cached. Follow the arrows, or click any stage to trace it. The frame schema and how the frontend applies each patch are in section 09.

The detailed breakdown of one Sonar search as a tiered map - one colour per repo, top to bottom along the data path. Search signals from the public data sources; once found, the run fans out per company to the enrichment agent (cache-first); each enrichment writes back to the global intel DB with provenance; then per-field frames stream to the live table. Hover any container for what it owns, or trace one company down the happy path.

Sources are a registry: each self-registers a schema, handler, category, availability check, and reputation weight at import; the discovery agent’s toolset is rebuilt per call, so a new provider appears with zero orchestration change. Discovery today is all layer 2 — Exa, Parallel, and a web_search fallback, all pull. The roadmap moves weight down the stack: free primary documents and webhook event feeds surface the event, search verifies and gap-fills it.

1. L0
   Free primaryfilings · wires · firehose

   Primary documents at source latency, $0. Beats every paid provider on US funding and leadership latency.

2. L1
   Structured event feedswebhooks · pre-categorized

   Events arrive pushed and typed — no polling, no extraction. The biggest gap in the current stack.

3. L2
   Neural searchExa · Parallel — in prod

   Semantic discovery, plus verification and gap-fill of everything the layers above surface. Today this layer is the whole stack.

4. L3
   Web fallbackweb_search — last resort

   Long-tail grounding. Cited and verified before storage, stored low-reputation.

Which source sees which signal first? Click a signal type to rank its best sources, click a row for detail, and toggle sources in and out to price the stack.

v1 seed weights, tuned to a marketing agency’s value: signals that imply fresh budget or a fillable gap rank highest. Pick the agency’s offer and watch the weights re-rank — the offer conditions which signals matter.

One real-world event surfaces from many sources — a Series A lands in SEC, Crunchbase, a trade outlet, and the company blog. That is not noise to discard; it is corroboration. Sonar folds the sources onto one canonical signal and keeps every source as evidence, so confidence and provenance survive dedup.

Signals decide who matters; enrichment fills in the durable company facts behind the lead. On a company-facts miss the run cascades down the tiers in fixed cost order — the model only decides escalate-or-stop per tier against a completeness target, never reorders them. Every tier has the same exit: the moment the target is met, the fields land and the cascade stops. Click any tier to inspect it.

• deterministic
• paid tier · model decides stop / escalate
• miss · cascade down
• target met · fields land

The lead score is computed in code — no override-rule prompt, no LLM-emitted number. Compose a few signals below and watch the score move; each weight here is a tunable constant in the real config.

Curve is illustrative of the deterministic formula — exact constants are tuned at build from eval data. The point: the number is reproducible code, not model judgement.

A global canonical store keyed by domain, read first in the waterfall and written back after every run. Public data (firmographics, signals) is shared across orgs — enrich once, reuse everywhere — while subjective judgement (score, fit, notes) stays in a per-org overlay behind RLS.

The run already owns an SSE channel for step progress; save-back reuses it. As cells resolve, the orchestrator emits tiny per-field frames on that same stream and the frontend applies each one to a live table — no polling, no re-fetch. Every frame is keyed by run_company_id (the workflow_run_companies row), so the table is just a reactive map the frames patch in place.

The two agentic nodes run under hard ceilings. Discovery stops at the first limit; enrichment’s order is fixed and the model only decides whether to escalate a tier.