Principal-drift ledger for standing instructions

Every guardrail audits the agent; nobody audits the principal. Operators forget their own standing rules, contradict them in a hurry, issue orders that silently override a safety policy set last month — or, in multi-human teams, give a shared fleet incompatible instructions — and the failure is recorded as an agent failure. This product pins the principal's directives the way a lockfile pins tools: standing instructions compiled into signed, versioned, scoped entries in an append-only, hash-chained ledger, with supersedes-links forming an explicit override graph, so "the new order wins" is a recorded, signed event, not a silent overwrite. Every new order is diffed at intake before the agent acts — CONSISTENT / CONTRADICTS(entry) / APPROVAL-REQUIRED / EXPIRED-RULE-INVOKED — a legitimate override produces a signed override receipt instead of an alarm, and every executed action carries an attestation of the ledger version it was checked under, so a post-incident review can separate agent disobedience from principal drift. Proven in the lab: 26/26 tests, verdict precision 15/15, 70% of a house-shaped rule stack compiling deterministically at 0.637 ms per diff with zero inference, multi-principal conflicts resolved deterministically by principal rank.

Prices

• Operator $9/mo per operator — one ledger, all agents, up to 10,000 intake diffs/mo; first ledger up to 25 entries free for the first month, one time per operator.
• Team $29/mo per operator — multi-principal ledgers, override-graph views, up to 100,000 diffs/mo.
• Per-diff API $0.001/diff (x402, Base USDC) — one intake verdict + attestation, agent self-check form.

Start

POST /register {"operator": "your-name"} returns a per-customer token (shown once). POST /trial claims the one-time free first ledger. POST /ledgers creates a ledger; POST /ledgers/{id}/compile compiles a standing rule (rules that cannot compile faithfully are stored signed as residual, never silently lossy); POST /ledgers/{id}/diff diffs an order at intake; POST /ledgers/{id}/execute re-diffs server-side and, only on CONSISTENT or a signed override receipt, executes with an attestation naming the ledger version in force. POST /subscribe {"plan": "operator"|"team"} with a verified x402 payment starts a subscription; without a plan, diffs are metered per-diff at $0.001 via x402. MCP JSON-RPC at POST /mcp (tools: ledger_create, ledger_compile, ledger_diff, ledger_execute, ledger_verify, ledger_version). Machine docs: /openapi.json, /llms.txt, /.well-known/x402, signing keys at /verify-key, terms at /terms, privacy at /privacy.

Limits (stated plainly)

• Diffing is deterministic compiled matching only — no inference runs on the diff path, ever; an inference-backed variant is blocked at config.
• The compiled share is protected; judgment/prose rules that cannot compile (30% of the lab's house-shaped fixture) are stored as signed residual entries and receive NO verdict protection: an order depending on them returns CONFIRMATION-REQUIRED naming the residual rule(s), and the confirmation outcome is appended as a signed ledger event.
• Approval intent compiles as APPROVAL-REQUIRED, never as a prohibition; time/scope qualifiers (weekday morning, after 22:00) are represented in the compiled predicate — a qualifier the compiler cannot represent makes the rule residual instead of silently lossy.
• Demand and the principal-caused share of fleet incidents are unmeasured; the attestation proves which rule set was in force, not that a principal's rules were themselves correct.
• Trial identity is normalized (case/whitespace-insensitive): a second account for the same operator does not re-grant the trial.
• Lost-token recovery is a manual step via support@textforagents.com.

Support: support@textforagents.com · Provider: TextForAgents LLC