Skip to main content

​
Commercial Intelligence v1 — consolidation and evidence

Status: merged and deployed. Live warehouse, stability and bounded Web Evidence controls recorded here as of this branch; risks and permits search shipped afterward as two further warehouse operations — see semantic-api.md for the current operation count and routes. This document is the integration ledger for the semantic market-data API, the domain research protocols, targeted Web Evidence, the customer Workbench and their evaluation program. It records what is physically implemented, what has been demonstrated live, and what remains transitional. Architecture is not considered proven merely because hermetic tests pass.

​
One product path

Mixed questions compose domain protocols; they do not create a fourth physical domain. Independent objectives may execute concurrently over the same shared belt. Dependencies are explicit — for example, a supplier search conditioned on a selected opportunity follows that opportunity search. The universal model-directed loop remains an exceptional escape for an intent the router cannot classify, not the normal path. The warehouse operations (nine as of semantic-api.md, including risk and permit search) remain the only market-data tools. The model does not receive SQL. search_web_evidence is a separate, bounded verification capability and can run only after warehouse evidence identifies a concrete entity, claim or coverage gap. It is not supplier discovery, enrichment or exhaustive web search. The post-protocol gate is a structured decision, not an unconstrained tool turn. DeepSeek receives the question, non-sensitive counts and organization names from a closed projection, then emits search|skip plus the five request fields the capability accepts. The platform validates and executes that request. An explicit positive request cannot be silently skipped; an explicit refusal keeps the web budget untouched.

​
Domain protocols

​
Resource and report policy

  • The call count is a high safety fuse, not the research strategy.
  • Each warehouse observation records elapsed cost and useful yield.
  • Three consecutive calls without new evidence, a new detail or a distinct bounded zero-result probe stop exploration as stagnated.
  • Repairable report excess is normalized before semantic validation. One bounded synthesis repair receives the concrete failure. If synthesis still fails, the platform emits an honest partial or unsupported report and keeps the complete audit; useful rows are never discarded for presentation shape.
  • Evidence, coverage, provenance, timestamps, cost and citation URLs are platform-owned. Model-authored facts and recommendations may only refer to evidence IDs that exist in the audit.

​
Customer surface

/w/:slug/research is the canonical integration candidate. It uses the existing Chat and Work Session runtime, the shared market-data contracts and live /capabilities. It renders the three domains, cursor pagination, details, semantic failures, reports, evidence, coverage, warnings and a collapsed audit. It has no fixture fallback. /w/:slug/intelligence remains reachable and fixture-driven. It is not a second supported product; it is retained until live parity, accessibility and reachability checks are recorded and a human approves deletion.

​
Evaluation gates

  1. Retrieval: 30 Supplier, 20 Opportunity and 15 History candidate cases. Materialization records a live corpus basis but remains unsigned until a human supplies expected stable IDs and reviewedBy.
  2. Trajectory: live Supplier, Opportunity, History and Mixed questions, plus web-needed, web-unneeded, insufficient-coverage and monetary currency/scope cases.
  3. Stability: five runs per frozen configuration, reported by domain with verdict agreement and top-candidate Jaccard. Retrieval, strategy and synthesis variance are separate failures.
  4. Ownership: every failure is one or more of retrieval_miss, strategy_miss, provider_fail, contract_reject, synthesis_error or render_fail.
Hermetic tests demonstrate contracts and control flow. Live controls now prove one necessary and one unnecessary Web Evidence trajectory, including platform- owned clickable citations; they do not substitute for a retrieval gold set or customer-visible browser QA.

​
Reachability-based deletion ledger

The ledger was rechecked after the live completion slice on 2026-08-13. No additional legacy surface is safe to delete: /research has contract/build parity and live report artifacts. A local Workbench-contract smoke authenticated through the same workspace boundary and exercised live capabilities, coverage, Supplier, Opportunity and History routes against staging: all five returned the shared market-data/domain/1 contract with governed coverage; Supplier and History also returned their required semantic warnings. Visual accessibility and narrow-viewport parity remain unproven because no controllable browser was available in the verification environment. /intelligence therefore remains transitional. No deletion is justified by filename, age or architectural preference. A customer-facing or billing-reachable path requires a runtime/reachability proof and explicit human approval.

​
Explicitly deferred

No Dynamic Workflow persistence, multi-agent supervisor, sandbox/code mode, enrichment, FindAll, new Source Packs, dense retrieval, entity-resolution graph, automatic outreach or weakening of semantic invariants is part of this consolidation. Each needs its own measured trigger.

​
Durable governance

Draft Topics:
  • commercial-intelligence-domain-protocols
  • targeted-web-evidence-policy
  • commercial-research-workbench
  • commercial-intelligence-evaluation-program
Execution board: Commercial Intelligence v1 — consolidation and evidence. The board leaves human gold review, legacy deletion, merge and deployment as explicit human decisions. Live N=5 evidence and limitations: commercial-intelligence-v1-live-eval.md.