# Tool-Confusability Recovery Loop — TextForAgents LLC

When a tool is added to an agent's catalog, the agent can quietly get
worse at the tools it already had: two needed tools that look alike
steal each other's calls, and the wrong call returns a plausible
answer with no error. This service is the recovery loop for that
failure. You register your catalog and probe suite once; the service
holds the catalog's green state. Whenever your live catalog changes,
a loop run measures it against that green state per selector class.
If a tool is stealing another tool's calls, the run returns a
worklist of disambiguation clauses ranked by measured recovery in
re-probe — not by how the wording reads — you pick one, the service
returns the revised catalog, and a verification re-probe issues a
recovery receipt: the flagged pair's confusion before and after,
per selector class. A fix that does not measurably recover is
receipted NOT_RECOVERED, never RECOVERED.

## Prices
- Per loop run: $0.05 (x402, Base USDC)
- Operator: $9/mo — up to 200 loop runs/mo
- Fleet: $29/mo — up to 1,000 loop runs/mo
- First loop run free, one time per operator.
- Small-class probing is included. Frontier-class probing runs on
  your own model key (BYOK), or in suites of up to 30 probes on
  company keys.

## Flow
1. POST /register — {account_name, payer_wallet} → account token.
2. POST /catalogs — register {catalog, probes}; the service
   measures and stores the green state. Free.
3. POST /loops — {catalog_id, catalog} starts a paid loop run
   against your current live catalog. Returns the steal pairs and,
   when any are found, the ranked recovery worklist.
4. POST /loops/<id>/apply — {item} serves the revised catalog
   with the chosen clause applied.
5. POST /loops/<id>/verify — re-probes and returns the signed
   recovery receipt.
6. POST /verdict — server-author form: submit one tool and get its
   confusability verdict against the curated reference stacks; every
   verdict names the stack it was measured against.

Every loop run also returns a capability-diet load manifest built
from the run's confusion matrix (majority-stolen tools are demoted
for re-description first; tools stolen from are defended first) —
call frequency alone is never treated as demand.

Support: support@textforagents.com