13. Market-overlap assessment

• Status: accepted
• Date: 2026-09-30

Context

Step 7 adds the ConflictClassifier, with anthropic, client, and off modes, plus the assess_market_conflicts tool, the saas_naming_session prompt, and check --market.

Decisions

The official SDK, over core's fixed-origin fetch

The anthropic classifier uses @anthropic-ai/sdk (0.130.0), not hand-written HTTP. The SDK is built with fetch set to createOriginFetch(["https://api.anthropic.com"]), so every request still leaves through core/net, restricted to that origin (invariant 2), and responses are size-capped. Only GET and POST with string bodies are forwarded; anything else throws. The SDK handles headers, versioning, retries (2), and typed errors.

It calls beta.messages.parse with structured output from zodOutputFormat. Each response is checked twice: first against the loose output schema, then per item against a strict one (2 to 4 reasons of at most 300 characters, a known level, a domain that was asked about, and no duplicates).

A site is left unassessed, never given a guessed level, when:

• the request fails (network, 401, 429, 5xx);
• the stop reason isn't end_turn (for example refusal or max_tokens);
• the output doesn't parse or validate;
• or the model returns no valid item for that site.

Model, effort, and fallback come from config

• Model: assessment.model in config, defaulting to claude-opus-5-5, the current default model.
• Effort: assessment.effort, default medium, set explicitly because Opus 5.5 can't disable thinking and its effort default differs from earlier Opus models.
• Refusal fallback: server-side refusal fallback is on by default (fallbacks: "default" with beta server-side-fallback-2026-07-01). If a safety classifier declines a request, the API retries it on a model chosen by refusal category, inside the same call. assessment.refusalFallback: false turns it off.
• Classifier id: anthropic:<model>. It appears in assessedBy and in the cache key, so changing the model never serves another model's judgment.

Prompt-injection handling

• Only extracted fields are sent: title, descriptions, JSON-LD types and name, final URL, content confidence, and untrustedSiteText, which is already cleaned and capped at 600 characters.
• Data blocks can't be broken out of: each site is JSON inside <site> tags, with < and > escaped to < and >, so no site can close its block, open another, or forge the <market> block. A test injects exactly that and counts the blocks.
• The system prompt treats site content as data: it says everything inside <site> was written by the site's owner, that it's data and never instructions, and that claims about how to classify it are to be ignored.
• Reasons are checked, not trusted: they must be about the domains asked about. Anything else is dropped.

What gets judged

Only sites with occupancy unassessed, meaning a real site answered. Parked, for-sale, and no-site domains already have a deterministic answer, and a for-sale lander isn't a competitor. Up to 10 sites go in one request. Judgments are cached for 24 hours per domain, market, and classifier. The market text is normalized for the key and hashed out of it.

Modes

Surface	Default	Why
titlesearch mcp	client	Per the brief: Claude judges from the evidence. The tool description tells it how, and to give reasons.
check --market	anthropic if ANTHROPIC_API_KEY is set, else evidence only with a notice	There's no client model in a terminal.
Either	assessment.mode in config overrides	Setting anthropic without a key is a config error, not a silent downgrade.

In the CLI, client behaves like off.

ANTHROPIC_API_KEY is read only from the environment (invariant 6) and is registered with the redacting logger before anything else runs. A live check with an invalid key reached the real API through the fixed-origin fetch, got a 401, left the site unassessed, and logged no part of the key.

Surfaces

• assess_market_conflicts takes one name, a market description of up to 1000 characters, and extensions. Its description depends on the mode, and it always ends with the trademark reminder.
• saas_naming_session takes product and audience and walks the five steps in the brief.
• Both are registered only when a presence probe is configured.

Not verified

No live, successful classification has been run: this environment had no API key. The request shape is checked against the SDK's own serialization, and responses against realistic Messages API bodies. The first live run should compare a few judgments with a human's.

Consequences

The Anthropic SDK is a dependency of packages/assess only. Core still depends on nothing but Zod.