Repository navigation
Map node/runner telemetry and checker handoff contracts #104
Description
Outcome: the node/platform handoff is agreed. The remaining implementation and acceptance checks are consolidated into the app/verifier (#90), collection/correlation (#44) and qualification (#91) tickets. This mapping ticket is complete.
Reviewed source baseline: Node 7a75438ed94fa0472bb6c70efe5f81f5193ab728 and Go runner 38b527c7e74b3801e5ddd67e51ad52d98576bfe0. These are source-review references, not release pins or deployed acceptance evidence.
The node provides ordinary registration, sessions, request/stream transport and execution diagnostics. The evaluation responder runs as an ordinary sandboxed HTTP app. The platform's trusted verifier checks responses, measures time externally and constructs the report; the platform also owns paid invocation, collection and qualification. The allocator owns placement and deployment decisions. Generic sandbox isolation, resource controls, cancellation and cleanup remain runtime work.
The agreed correlation model uses ordinary runner_id/session_id in reports and logs/events. Metrics remain aggregated by app/runner, with available deployment/pod metadata and time providing context. Session IDs do not become metric labels, and evaluation requires no new Node evaluation-run ID. Collection acceptance includes a worked trace, restarts/replacements, single-shot calls and multiple evaluations within one session. Add a generic field only for a demonstrated correlation gap.
Keep platform errors running or collecting an evaluation distinct from errors attributable to the orchestrator, node or operator. A platform error retains the last successful result until its existing expiry and records the failed attempt separately. It does not refresh the measurement time or extend validity. Target-attributable failures affect the relevant reliability/eligibility and cannot be hidden by an older pass. Insufficient evidence is recorded as unknown attribution.
The initial profile is the supplied CPU/network challenge–response plan. Runs are paid, credentials stay outside the app, and GPU/RAM/storage benchmarks are deferred. Reports keep latency, directional throughput, hash rate and reliability separate and record the verifier and observed routing path; these measurements do not prove physical-machine identity.
Parameters, budgets, freshness defaults, report/attempt identity, retry behavior and the versioned result format are implementation responsibilities in #90. #44 implements collection, safe correlation and deduplication; #91 applies the agreed qualification policy. Existing sandbox acceptance work exercises ordinary execution, streaming/cancellation as applicable, payment and cleanup. Reports and diagnostics exclude credentials, session tokens, private generation seeds and sensitive payloads.
The schema work formerly in #40 is consolidated into #44, and the framework work formerly in #46 into #90. No separate handoff/specification ticket or evaluation-specific node subsystem is needed. Closing this issue records the completed mapping and decisions; implementation and end-to-end evidence remain open in the delivery tickets.
Activity
j0sh commentedon Oct 30, 2026
The handoff is settled. Reports join to session logs/events through ordinary runner/session identity; metrics stay at app/runner granularity, with deployment context and time providing the surrounding view. Session IDs do not become metric labels. The collection work includes concrete traces for restarts, single-shot calls and repeated evaluations in a session.
We also agreed to distinguish platform errors running or collecting an evaluation from errors attributable to the orchestrator, node or operator. A platform error retains the last successful result until its original expiry, without refreshing its timestamp. Target-side failures remain evidence about the endpoint and affect the relevant reliability/eligibility; an old pass must not mask them. Unknown attribution stays explicit.
To reduce overlap, the telemetry schema work in #40 is folded into collection/correlation in #44, and the framework work in #46 into the app/verifier implementation in #90. Qualification policy remains in #91. Those implementation tickets include the remaining acceptance checks.
Closing this mapping ticket now that the decisions and ownership are recorded. This does not claim the implementation or end-to-end tests are complete.