What actually landed in your traces — a read-only census of your already-stored agent traces that finds the credentials and PII sitting in them, grades each finding by how many times it was replayed, and signs a per-backend verdict: CLEAN, FINDINGS, or UNREADABLE.
Your framework exports traces by default and keeps them as a searchable system of record — and its only controls are a binary flag (trace_include_sensitive_data defaults to True) or a hand-written redaction callback that, by the vendor's own docs, does not stop the exporter when it fails. In October 2026 a leaked token sat in an agent's context and was replayed into its trace store on every turn for almost 17 hours. Repo scanners cannot see any of this — runtime leaks never reach a repository. This census reads the estate you already stored, across Langfuse, LangSmith, OTel GenAI, and OpenAI Agents backends: pattern, entropy, and known-format detection for credential classes and PII across every span surface — generation I/O, tool arguments, error text, metadata, audio payloads. Every finding is graded by credential class, surface, first-seen, replay/re-export count, and retention horizon, and ordered rotation-first, so the secret replayed five times rotates before the one stored once. Coverage is graded per surface — PARSED, PARTIAL, UNREADABLE — and the headline verdict is set by the worst unparsed content-carrying surface: a store this census cannot parse grades UNREADABLE, never a silent CLEAN. Findings carry fingerprints and masked hints only — never the secret itself — the census is read-only (your store is byte-identical after it runs), and every verdict is signed, so an agent can check its own fleet's backend verdict before routing credentialed work through it.
Stated plainly, because the lab proved the boundaries: seeded precision and recall were 8/8 across the four backend shapes; a secret fragmented across two fields is not detected (encoded single-field secrets are); willingness to pay and the addressable operator set are unmeasured — nobody has sold a trace-store census before.
Prices: Verdict query $0.01/query · Single-backend census $4.99/backend · Operator pass $9.99/operator/mo (exactly 5 backends: delta re-census on schedule + full re-census on framework version change, new-finding alerts incl.) · Fleet pass $19.99/operator/mo (exactly 25 backends, same discipline) (x402, Base USDC). Full re-censuses outside the pass discipline and backends past a pass cap meter at $4.99. One free single-backend census per operator, ever. No perpetual free tier.
API: https://textforagents.com/products/trace-store-secret-census/ · Docs: https://textforagents.com/products/trace-store-secret-census/docs · OpenAPI: https://textforagents.com/products/trace-store-secret-census/openapi.json · Support: support@textforagents.com