HALOWERK aialignwerk

Endpoints (10 live)

• POST /v1/agent-collusion — Requires the same agents in each supplied round, computes each pair’s observed action agreement, expected agreement from marginal action frequencies and chance-corrected kappa, then flags high positive agreement against a supplied threshold. Coordination can be benign or task-driven; this metric does not prove communication, intent or collusion. (0.005 USDC on Base)
• POST /v1/bias-audit — Calculates each group positive-outcome rate, compares it with an explicit or automatically selected reference group, and reports rate differences and selection-rate ratios. Aggregate disparity metrics do not establish discrimination, fairness, causation or legal compliance and can conceal intersectional or sampling effects. (0.004 USDC on Base)
• POST /v1/confidence-calib — Compares scalar confidence with binary correctness, calculates mean squared Brier loss, partitions confidence into caller-selected equal-width bins and reports expected and maximum calibration error. It does not validate labels, correct class imbalance or prove calibration beyond the supplied sample. (0.003 USDC on Base)
• POST /v1/resource-hog — Divides observed CPU, memory-time, network, tool-call and wall-time usage by caller-supplied budgets, reports exceeded dimensions and ranks observations by the maximum ratio. It does not collect telemetry, infer intent, identify inefficient algorithms or terminate an agent. (0.003 USDC on Base)
• POST /v1/goal-drift — Aligns caller-labeled dimensions, computes cosine similarity and a normalized L1 difference, and flags drift against a supplied cosine threshold. It does not infer goals from behavior, decide which goal is correct or detect deception; results depend entirely on the supplied vector representation. (0.003 USDC on Base)
• POST /v1/jailbreak-defense — Applies a small fixed set of defensive regular-expression categories and returns category names, a bounded risk score and a review recommendation. It does not execute, transform or forward the prompt. Pattern matching is incomplete and can produce false positives; it should be one signal in layered controls, not the sole access decision. (0.004 USDC on Base)
• POST /v1/memory-sanitize — Replaces private-key blocks, credential assignments, email addresses, Luhn-valid payment-card candidates, IPv4 addresses and phone-like digit strings, returning only sanitized text and per-category counts. The request archive stores only input field names and the result archive stores only this method via the endpoint factory. Pattern redaction is incomplete: it cannot identify every secret, name, address or contextual identifier and must not be treated as irreversible anonymization. (0.003 USDC on Base)
• POST /v1/hallucination-rate — 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. (0.004 USDC on Base)
• POST /v1/sycophancy-detect — Compares term-frequency cosine similarity between a leading user position and paired independent/conditioned responses, then reports the positive similarity shift. It is a reproducible surface-form signal, not proof of agreement, truthfulness, motive or sycophancy; paraphrases and legitimate corrections can evade or trigger it. (0.004 USDC on Base)
• POST /v1/capability-eval — Computes weighted pass rates and weighted scores from caller-labeled test outcomes, with deterministic category summaries. It does not run tests, validate labels, measure untested capabilities or establish deployment safety; the result is only as representative as the supplied evaluation set. (0.003 USDC on Base)