All projects matching “hallucination”

Compile 1-8 documents into a token-budgeted context pack, then measure exact retention of 1-32 caller-labelled evidence spans. Returns the pack, source-linked passages, input and output hashes, original and compiled token estimates, tokensSaved, source coverage, and requiredEvidenceRetentionPercent. Retention is exact span matching only: not accuracy, answer quality, verification, or hallucination

Answers questions (single-turn question answering) from a small local LLM (qwen3.5 2B). Best for general knowledge, definitions and simple reasoning. POST JSON {"input": string}, up to 3000 characters; returns {answer, model, elapsed_ms}. Answers are brief (2-3 sentences). Typically 5-15 s. Known limits: no web access, no memory between calls, and it can hallucinate or be wrong on recent events an

Tokenizes each claim and its supplied evidence, removes a small declared English stop-word list, and reports the fraction of unique claim tokens present in evidence. Coverage below a supplied threshold is flagged. This is lexical grounding screening only: overlap does not prove truth, low overlap does not prove hallucination, and the service does not retrieve or validate evidence.

Map a natural-language drug, diagnosis, or lab phrase to a current RxNorm, ICD-10-CM, or LOINC code. Use when an agent must not hallucinate or invent a billing/pharmacy code.

Vulnerability scan for a list of packages in one call via OSV.dev: GET ?packages=lodash@4.17.15,pypi:requests==2.19.0,go:github.com/gin-gonic/gin@v1.9.0 (or POST a lockfile). Up to 200 across npm, PyPI, Go, crates.io, Maven and more: per-package advisories, severity, fix versions and upgrade targets; hallucinated and malicious packages flagged. JSON.

Call before an agent runs npm install. Checks one npm package against the registry, OSV advisories, install scripts, typosquat and combosquat names and repository provenance, and flags hallucinated or slopsquatted names. Answers safe, caution, block or does_not_exist; every reason is sourced and labelled fact or heuristic.

Verify up to 50 citations in one call, one flat price. Body: {"citations": ["<string>", ...]} - same shapes and verdicts as the single route. Returns results[] (one per citation), the verified/not_found/suspicious counts, verification_time_ms, hallucination_risk (LOW all verified, MEDIUM 1-2 not found, HIGH 3+). Deterministic, no LLM in the path. suspicious = a name_match mismatch: a DIFFERENT cas

Checks whether a token contract address is real, matches its claimed symbol, and whether its supply figures match an independent source (CoinGecko) — before trading, coding, or reporting against it. Catches hallucinated/wrong addresses, address-poisoning impersonation, and tokenomics hallucination.

Verify an SEC filing: given a filer's CIK and a claimed accession number, does that filing actually exist, and does its form type/filing date match what's claimed? Checked against SEC EDGAR's submissions API. Catches a fabricated or misattributed filing citation, including a real-looking accession number paired with an invented form type or date.

Research cards from KARP, SoulDriver's multi-model research engine: GPT, Gemini, Grok and Claude debate a topic over many rounds and publish a conclusion. Each card gives the conclusion, key insight, warnings, dated figures, consensus status, open questions, research date and staleness. See what is covered first with the free POST /v1/coverage; verify time-sensitive points with POST /v1/check. Cat