Pick protocol from explicit --protocol flag or model-name heuristic.

Advertised maximum input context for known models, in tokens. Returns 128_000 as a conservative default for unknown models so we don't accide

Build a DetectorComparison for one detector. Either side may be None (e.g. detector skipped on one run). When the relay skipped a detector th

Probe a relay's /models endpoint. Always returns a structured dict. Never raises — network and protocol errors are returned as ``ok=false`` w

Map a model id to a protocol bucket, or None if unrecognised. Heuristic only — relays sometimes route the same alias to a different backend t

Patch httpx.AsyncClient to use a mock transport for one call.

Force probe_relay's httpx.AsyncClient to use this transport.

Three deterministic needles seeded by relay identity. Each Veridrop run gets its own seed (relay domain + timestamp), so needles are unique p

Non-streaming POST /v1/messages. Returns: (request_body, response_dict, response_headers, latency_ms). Raises AnthropicAPIError on no

Return True if the answer hits expected keywords AND avoids anti keywords.

Return ``(allowed, retry_after_seconds)`` for one hit on this IP. Sliding-window: counts hits in the trailing ``window_s`` seconds. If the co

Lenient match for Gemini OpenAI-compat responses. Relays return the model id in `response.model` in several shapes: - bare alias: `gemini-2.5

Match the user-supplied model ID against known aliases (double-prefix). Accepts both alias (`claude-opus-4-7`) and snapshot (`claude-haiku-4-

Conservative rate-limit detection — false positives here would let a real truncation slip through as "rate_limited", so prefer explicit TPM /

Weighted average over results that are not skipped. Per DESIGN.md §5.2: effective_weight = Σ d.weight for d.status != "skip" tota

Build a haystack of approximately target_tokens with needles embedded at their target positions. `protocol` selects the tokenizer-family-spec

Anthropic Messages API response shape: content blocks + usage.

Run one tier: build haystack, send, score recall. Haystack is sized to leave ~QUESTION_BUFFER tokens of headroom for the question its

For each needle, did the model reproduce its `answer` in the response? Case-insensitive substring match — the answers are unique IDs (e.g. 'A

Adaptive tiers covering low/mid/high of model's context window. For a 1M model returns ~(32k, 500k, 950k); for a 200k model returns ~(32k, 10

Pull model id strings from any of the common response shapes. OpenAI: {"data": [{"id": "..."} ...], "object": "list"} Anthropic: {"data":

Extract assistant message text from a non-stream Chat Completions response.

Normalize a Gemini OpenAI-compat base URL. Preserve explicit prefixes such as `/v1` and `/v1beta/openai`. If the user enters only a host root

Normalize Anthropic-compatible base URLs to the API root. Users often paste OpenAI-style relay roots ending in `/v1`. This client posts to `/

Verdict capped by critical findings. A relay scoring 75 with NO critical issues is genuinely "passed". A relay scoring 75 with critical issue

Rough USD estimate for sending target_tokens of input. Hard-coded per-model price tiers — accurate enough for "is this $0.05 or $5?" guidance

Return a run-level error when upstream quota/billing blocks detection. These are not model-quality failures. Scoring them as ordinary detector

True if any detector reported one or more critical-severity issues. Detectors surface critical issues via either: - `details.critical_issue_c

Canonicalize model ID separators so users typing `claude-sonnet-4.5` match the official `claude-sonnet-4-5` form. Without this the strict pref

Run model-alive preflight; raise HTTPException 422 with structured detail if the model is dead so the frontend can offer a one-click swap to t

Mask API key for display: 'sk-y7xUabc...0h' -> 'sk-y7xU••••••0h'.

Streaming POST /v1/messages, yields (event, elapsed_ms_since_request_start). Caller is responsible for the entire SSE lifecycle. We do NOT ac

Return ``(alive, error_text_or_None)``. ``alive=False`` means the upstream rejected the call — error_text holds the upstream message (truncat

Keep only diagnostic headers; never persist auth/cookie headers.