RELIABILITY + WORKING ROUTE CONTRACT · V0.1
Check first. Recover safely.
Agent WEX indexes minimized outcomes for exact public tool-compatibility cells. It lets an agent inspect recent reliability before a call and request a bounded recovery route after a failure. It does not build, host, execute, or authorize agents.
The exact cell
A preflight query fixes the tool registry and ID, client ID, environment class, authentication mode, and operation. The current tool and client versions identify the route being considered. Candidate recovery routes may vary the tool version, client version, and resolution kind inside that same bounded cell.
Free aggregate preflight
Preflight reads the latest accepted outcome per signed node in the requested evidence window. It returns current success and failure counts, freshness, low-cardinality failure classes, heuristic confidence, and possible outage or regression alerts. The aggregate assessment is free. It executes nothing, grants no authority, and does not automatically intercept calls; the agent or runtime invokes it before a fragile or expensive tool path.
The support unit
Receipts are collapsed through provenance root, signing node, participant, and controller group. For the public Working Route decision, support is counted at the controller_group cut: more keys, runtimes, devices, or VMs under one controller do not create more network support. Counts at every resolution remain visible, and none proves independent ownership.
Independence depends on the question
The same three machines can be useful replication for a machine-specific compatibility question while remaining one controller for operator consensus. Agent WEX preserves multi-resolution counts; it does not silently choose a different boundary. Any other assessment must explicitly state its decision, relevant failure domain, independence cut, and support threshold. Changing the cut never grants action authority.
The output
A returned route is marked evidenceStatus: unverified-network-evidence, controllerIndependenceVerified: false, executionTruthVerified: false, gateRequired: true, and authorityGranted: false. The caller must apply its own policy and authorization.
Failure becomes exchange value
Signup is free and starts with zero credits. An accepted first additive outcome from a signed node earns one or two access credits based on freshness; a fresh accepted failure currently earns two. The failure also updates the demand and reliability map. Repeated claims do not earn again. Credits only track contribution and access; they never affect evidence weight.
Unlock only when useful
If preflight finds a supported alternative, the caller may deliberately request --unlock, which uses one earned credit. If no supported route exists, every credit remains available and the uncovered cell becomes visible as a gap for future evidence. A released route always carries gateRequired: true.
Measured impact, bounded claim
After trying an unlocked route, a node may report a bounded outcome and numeric estimates for attempts, tokens, or latency avoided. There is no free-text feedback field. These totals are self-reported estimates used to evaluate product utility; they are not verified savings or credit inputs.
Current scope
The preview can compare current and alternative versions and resolutions inside an exact compatibility cell. Broader fleet routing across different tools, providers, authentication methods, runtimes, or operations requires cross-cell coverage and policy that the preview does not yet claim. Aggregate vendor intelligence becomes more useful as coverage grows.