Prevent browsers from caching HTML responses. HTML pages reference versioned CSS/JS via ?v=N query strings, but the cache-bust only works if

Override to accumulate state across many requests.

Collect an official OpenAI protocol baseline from live API responses.

Validate a raw OpenAI response against the protocol template.

Weighted by per-protocol detection count — tiebreaker for ranking.

Pick the most likely-to-just-work model from a relay's whitelist. Tries exact alias match first, then snapshot-suffix match (e.g. our preferr

Send one minimal /v1/messages request to verify connectivity & response shape. M1 verification command. Prints the raw response fields so you can

Bayesian-weighted score for ranking. Pulls few-sample relays toward a neutral prior (50) so a single fluky 100% doesn't beat a consist

Dynamic sitemap — static product pages + reports + per-domain pages. `lastmod` is included for every URL so search engines know when to revis

Queue a detection job and return the job id immediately. Long-context probe is opt-in. Two tiers: - ``include_long_context`` (standard): 32

Even if a single-sample relay's ranking_score is high, it should NOT outrank a multi-sample relay (≥2). is_ranked tier comes first.

200k tier on a 200k-context model must be probed (not skipped) by clamping the haystack to leave room for the question. Catches the bug where

Sanity check: detector calls messages_create (not chat_completions), sends max_tokens (not max_completion_tokens), and extracts content[].text.

Unknown / legacy aliases don't crash; fall back to anthropic which is the historical default for relay-detector.

The fundamental sort claim: a 100-test consistent 99% beats a 1-test 100%. This catches the regression where someone "fixes" the sort to use o

Anthropic-backend impersonation must be flagged as critical (not just major), so that ProtocolDetector fails the relay even when every required