Company Monitoring provider policy and 500-company cost package

• Review date: 2026-08-05
• Protocol: cm_eval_v1
• Runtime decision: blocked

This review freezes the policy and price inputs used by the Company Monitoring Stage 0 decision. It is not legal advice or a durable provider guarantee. Provider terms, prices, declared use cases, and model endpoints must be refreshed before any paid runtime is enabled.

Exa

The Exa adapter remains dark while COMPANY_MONITORING_ROLLOUT_FLAGS.exaProvider in convex/config/productCatalog.ts is false.

• Use only Exa's official API. No browser scraping or credential sharing.
• The current Search price is $7 per 1,000 requests for up to ten results, plus $1 per 1,000 additional results above ten. Each requested content type is $1 per 1,000 pages. The default Search rate limit is 10 queries per second.
• Search responses expose result-level crawl status and cost metadata. A capped, partial, or failed result must remain partial and cannot refresh coverage as adequate quiet.
• Exa's service does not confer rights to republish every indexed page. Store only permitted excerpts and metadata and preserve source-specific restrictions.

Sources: Exa pricing, Search API, Contents API, and rate limits.

X

The X adapter remains dark while COMPANY_MONITORING_ROLLOUT_FLAGS.xProvider in convex/config/productCatalog.ts is false. Activation must retain the content-handling controls below.

• Use only the official X API. Recent search covers seven days and returns up to 100 Posts per request; full-archive search is a separate pay-per-use or Enterprise surface.
• Current pay-per-use pricing is $0.005 per Post read and $0.010 per User read. Pay-per-use plans have a two-million-Post monthly read cap.
• X requires the declared use case to remain current. Its agreement requires an Enterprise plan when use grows beyond commercial prototyping, initial integration, or a limited number of end users. API credentials and purchased credits do not activate the Company Monitoring rollout flag.
• Offline X Content must track deletion, edit, protection, suspension, and withholding. Applicable removals must occur as soon as reasonably possible and within 24 hours of a request. Batch compliance may be used for bounded audits; high-volume compliance streams require Enterprise access.
• X Content may not train an AI or machine-learning model. Raw X text, handles, or provider URLs cannot enter customer alerts, logs, or committed evaluation fixtures.

Before activation, runtime enforcement must cover offline edit, deletion, protection, and withholding handling plus the model-training prohibition.

Sources: X pricing, usage and billing, search access, Developer Policy, Developer Agreement, and Batch Compliance.

Model routing

Status: the dedicated runtime contract is implemented; empirical use remains blocked by the frozen Stage 0 decision.

The cost model uses OpenRouter's deepseek/deepseek-v4-flash price snapshot: $0.09 per million input tokens and $0.18 per million output tokens, plus the 5.5% credit-purchase fee. Reasoning is disabled for this extraction/classification shape.

Every Company Monitoring request must:

• pin and record the model and provider-policy version;
• send provider.zdr: true or an equivalent enforced organization guardrail;
• disallow providers that train on prompts;
• retain no prompts or completions in OpenRouter activity logging; and
• fail closed when no eligible zero-data-retention endpoint exists.

The provider result must separately mark the route approved and prove all four runtime controls: ZDR, no prompt training, disabled reasoning, and pinned model and provider routing. Declaring those requirements in policy without runtime enforcement remains blocked.

OpenRouter states that its own prompts are not retained unless prompt logging is enabled, supports request-level zero-data-retention routing, and conservatively marks endpoints with unknown policy as retaining and training. The dedicated Company Monitoring client now sends provider.only with one configured route, allow_fallbacks: false, require_parameters: true, data_collection: deny, and zdr: true. It sends temperature: 0, disables reasoning, requests router metadata, and fails closed unless the response proves the requested model, direct single-attempt route, selected endpoint, expected resolved provider, no transformation pipeline, and request cost. Live admission requires both COMPANY_MONITORING_CLASSIFIER_PROVIDER_ROUTE and COMPANY_MONITORING_CLASSIFIER_RESOLVED_PROVIDER; a key and model alone no longer enable admission classification. Admission receipts include the requested route and the attested resolved provider in their model-version identity.

These request controls do not override the checked-in protocol. While Stage 0 is STOP, the worker's checked-in classifier runtime gate remains false, and the offline prediction command exits before it loads credentials or sends a paid request. Provisioning the key, model, and route therefore cannot activate live classification. Account-level activity-logging state also remains an external attestation; no score can infer it from a successful model response.

Sources: OpenRouter zero data retention, provider logging, pricing, and DeepSeek V4 Flash pricing.

Production-shaped monthly model

The model covers exactly one account-level shared-discovery workload of 500 companies for a 30-day month. It is a budget envelope, not measured capacity or a promise of provider yield.

Component	Frozen assumption	Monthly cost
Exa Search	12 searches/day, 25 results/search; $0.007 base plus 15 additional results at $0.001 each	$7.92
Exa content	One content type for 25 pages across 12 searches/day at $0.001/page	$9.00
X Post reads	250 Posts/day at $0.005/Post	$37.50
X User reads	500 User reads/month at $0.010/User	$5.00
Model	250 candidates/day, 4,000 input and 1,000 output tokens each, plus 5.5% fee	$4.27275
Allocated infrastructure	Convex, Railway, storage, and telemetry envelope	$25.00
Subtotal	Before contingency	$88.69275
Contingency	25%	$22.1731875
Total	500 companies	$110.8659375

The frozen ceiling is $125 per account-month, or $0.25 per monitored company. The modeled total is $0.221731875 per company and therefore passes the arithmetic gate. This does not override the Stage 0 stop. Before paid beta, the fourteen-day tracer must replace every volume assumption with measured requests, returned resources, caps, tokens, storage, and retry cost. Any changed provider price or workload shape requires a new cost-package version and product-owner decision. In particular, changing the portfolio size from 500 or modeling more than one account cannot reuse this package.

Frozen requirements and mutable runtime evidence

Provider access, retention, compliance, and model-routing requirements are part of the frozen approved-threshold projection. Runtime status, evidence digests, and enforcement flags live in a separate mutable result record. That separation permits honest evidence to arrive without changing the approved threshold digest while preventing a status-only promotion.