Metric definitions

Current metric identifier: delivered-metrics-v2. Old artifacts keep their recorded version. A zero measurement and an unavailable measurement are different states.

Quantity	Definition and eligibility
Success rate	Independently valid solved trials / all planned trials. Never divide only by completed runs.
Sum of costs / makespan	First-goal arrival action counts; t=0 costs zero. Waiting at an absorbing goal does not inflate cost. Valid solved runs only for cost comparison.
Recorded action count	All recorded transitions, separately retained for failed prefixes and diagnostics.
Individual detour	Cost above the sum of static individual BFS shortest paths, divided by that sum. It is not paper Eq. 2 or a joint optimality gap.
Conditional Eq. 2 form	Per-instance excess over the shortest valid solution among the explicitly selected decentralized methods, averaged on the declared common-solved cohort. The available reference set must be named. A general two-method export labels its selected reference and does not imply an optimal baseline.
Certified optimality gap	Requires a qualified optimal solution for the exact same instance and semantics. Missing baseline means unavailable. A lower candidate cost than the alleged optimum is an error to investigate.
Information sharing	For each ordered sender/recipient pair, intersection of actually delivered broadcast coordinates with the final executed spatial set, divided by final spatial-set size, then averaged across all ordered pairs. Repeated locations collapse. Single-agent value is zero.
All-message disclosure	Same spatial definition including negotiation offers. Kept distinct from broadcast-only sharing.
Space-time disclosure	Uses absolute (tick, x, y) states for actual broadcasts/offers; revisits remain distinct. This is an explicitly different metric.
Messages / bytes / cells	Actual delivery receipts. Byte count is the canonical JSON message payload, not transport overhead or a guessed network packet size.
Token exchanges	Actual transferred amount on successful atomic settlements. Zero-token agreements add no transferred volume.
Token balances / Gini	Signed before/after receipts and complete roster. No clipping of negative balances or assumption that the initiator paid. Old session-only logs cannot recover balances reliably.
Safety interventions	Actual changes to desired movement, including an explicit semantic-violation flag. An intervention count alone does not establish physical validity.
Runtime	Solver wall time, validation/replay, queue and process wall time are separate. Process timeout includes spawn overhead. No cross-hardware or failure-conditioned speed-superiority inference.

The legacy reconstructed sharing proxy is not computed for new experiments. It remains readable in historical bundles when present, explicitly labeled as a proxy. Offered utility/heat is null if the strategy did not provide it; response decision components are nested under their actual actor and are not relabeled as the proposer's utility. Telemetry events carry schema, sequence, phase, and available source/attempt/session linkage.

Online communication and settlement accounting runs independently of trace retention. Metrics-only retains executed paths and counters but omits event/frame histories. Events mode records event/replay data; full trace additionally includes local observations, plans and commitments. Trace overflow is visible in its receipt. Disk telemetry uses a bounded queue and checked gzip chunks: a completed prefix can be read after a crash, but queued/unpublished events may be lost. This is not an exact solver resume log.

Per-tick negotiation counts are retained in metrics-only runs as negotiations_by_tick, alongside total and successful session counts. Keys are simulation ticks (t=0 included when a session occurred). A missing key within a recorded run means zero sessions at that tick; an unavailable curve for an uninstrumented solver is omitted, not fabricated as zeros. Cross-run curves must declare their solved cohort, denominator and handling of finished runs.

Event-derived diagnostic summaries

Negotiation success in an event report is conditional on recorded bilateral sessions, not the experiment-wide solution success rate. With no recorded sessions, the rate is null and available is false. The wait ratio is likewise unavailable without recorded moves. Token summaries require a roster or actual balance receipts; absent balances and ambiguous legacy session records remain unavailable. The CLI and HTML report display these values as Unavailable; a measured zero stays zero. Token statistics reconstructed without a full roster describe only the recorded participants (roster_complete: false).