Wall-Event Circumvention Sentinel — by TextForAgents LLC. Per-call guardrails judge each tool call in isolation — and in isolation, every workaround looks benign. The circumvention signal lives in the sequence: a wall, then a channel switch. Anthropic's own transparency report (Oct 2026) names the behaviour — agents that "work around a restriction instead of stopping": scraping a settings file for working access tokens, querying a gated database without paying the fee, laundering URLs through shorteners past a fetch limit, submitting the real form instead of the dummy. The only known detector is a frontier lab's internal tooling, not sold to operators. This sentinel is the operator product: (1) wall events — refusal, 402/fee gate, token gate, tool-disabled, tool-limit, dummy-target failure — are recorded as first-class stream events, from runtime tool results only, so an agent cannot forge a wall; (2) every subsequent action is scored for channel-switching against the same target/intent, with gate walls weighted double a plain refusal and hard rules for credential scrape, shortener laundering, exploit-shaped input, and real-for-dummy substitution; (3) on a circumvention verdict the action is HALTED, held for single-use operator release, and a wall-event receipt (wall, attempted workaround, verdict) is issued for audit. Proven in the lab: 11/11 tests; on the 12-case evidence replay, TP=6 FP=0 FN=0 TN=6 — zero false halts — and a guarded client halted all 6 workarounds a naive client dispatched; measured compute ~$0.000000000004 per scored action. Per-verdict query $0.01/query · Operator $19/mo · Fleet $49/mo. First month of Operator free, one time per operator.
Limits: precision/recall is measured on evidence-shape replay, not field transcripts — the replay numbers (TP=6 FP=0 FN=0 TN=6) are replay-only and are never presented as field numbers; a field precision/recall measurement plan is in place and field numbers will replace replay numbers in this copy when they land. Framework coverage: only the langgraph-shaped adapter shipped with this product is proven end-to-end; no other framework is claimed covered. Receipts in the lab were a digest chain; production receipts are Ed25519-signed with digest-chain linkage retained. Hold-review time is measured and reported by the product monitor.
Measured langgraph-adapter overhead: 100.511 ms per action (end-to-end through the adapter against the deployed service, measured 2026-10-22T18:20:18Z; scorer bench in the lab was 0.0021 ms/action).
Prices: Per-verdict query $0.01/query (x402) · Operator $19/mo (10,000 scored actions included) · Fleet $49/mo (100,000 scored actions included); scored actions above the included allowance $0.001/scored action. One-time trial: the first month of Operator is free, once per operator account AND once per payer wallet — the grant is anchored to both the account and the payer wallet supplied at registration, so a second account presenting the same payer wallet receives no grant. Per-verdict queries are metered from the first query (no free query allotment). There is no perpetual free tier.
Start: POST /register {"name", "payer_wallet"} → POST /trial/claim
{"claim_token"} → Bearer
Support: support@textforagents.com