MBTA departure-event mesh¶

Status: Stage 1 carrier retained. The later population, task, and topology stages completed without replacing it.

Decision¶

Retain the ragged departure-event mesh as the source carrier. A 30- or 60-second clock preserves the 575 target identities in this slice, but spends most of its cells on absence. A five-minute clock is compact but merges 90 events and therefore requires an aggregation policy. Clocks remain derived, matched controls rather than source truth.

source rows + Schedule
          |
          v
  physical departures       597 records
      /          \
     v            v
  `run` 534    `headway` 575
      \          /
       v        v
   causal prefix(cutoff)

clock projection (derived only)
  30 s  -> lossless, 89.1% empty
  60 s  -> lossless, 78.2% empty
 300 s  -> 90 identities merged

This is a representation decision, not a forecast, topology-value, MBTA service-quality, online-causality, or public API claim.

Contract¶

One event key is:

(service_date, vehicle_id, parent_station,
 direction_id, departure_timestamp)

The departure timestamp is the following adjacent Schedule call's observed Vehicle Position movement timestamp. Source trip instance, stop identity, sequence, following call, Schedule departure, and source headway remain aliases on the physical event. Thus every retained event maps back to its exact source row and both applicable Schedule calls.

run(a, b) connects consecutive observed station departures in one trip and vehicle trace. headway(a, b) connects consecutive physical departures in one parent-station, trunk-route, and direction lane. Both relations name their endpoints and elapsed seconds and require time(a) < time(b).

prefix(c) is the event-induced subgraph whose timestamps are strictly below c. It cannot contain an edge incident to an excluded target because relation retention is derived only from the retained event-key set.

Evidence¶

The clean, locked run was:

uv run --locked python -m experiments.run gtfs_event_mesh

The envelope binds the structural comparison to PyG 5c6461b2305ad068a6d61165b3c55852a11aaa41, PyG Temporal fe555bc30ee197755c4b58a89407033a5f383415, and TSL aa5f313e000d192bdec270748b8d01df5912e58e. No framework is imported at runtime.

Exact mesh¶

All 575 internally derivable source headways reproduce exactly. The other 33 labels need an event beyond the retained cut and remain boundary evidence; none is imputed. A midpoint witness retains 297 events and 536 relations, excludes 300 events, and identifies 30 crossing relations that the prefix omits.

The lowering groups only trip traces and station-direction lanes, sorts within those groups, and emits adjacent relations. It never constructs station-pair, event-pair, or full station-time arrays.

Clock comparison¶

The comparison uses the same 575 exact target events and 22 lanes over the declared two-hour interval. One scalar cell can retain at most one event identity without a sidecar.

The small clocks do not lose identity in this witness, but neither is simpler source truth: each introduces an arbitrary cadence, missing-cell semantics, and several times more storage than the target events. The five-minute clock cannot be lossless without a ragged sidecar, which recreates the event carrier.

Structural comparison¶

These are useful contract comparators, not architectural owners. The Stage 1 host records stay experiment-local, and src/tinymesh/, dependencies, and exports remain unchanged.

Malformed witnesses¶

Tests prove deterministic lowering and exact Schedule/source round trips. A duplicate trip alias collapses only when its physical key agrees. Conflicting alias labels, unresolved or contradictory Schedule identity, non-forward time, self edges, duplicate relations, contradictory relation kinds, and unknown endpoints fail with the relevant identity. Checksum or artifact drift fails in the retained loader before lowering. Prefix tests exclude the target and every future incident relation.

Stage 2 consequence¶

Acquire and retain event facts, not pre-binned snapshots. Stage 2 should first estimate a capped 28-day, all-rapid-transit extraction, then retain the smallest population that covers ordinary and disrupted service without weakening observational lineage. Its manifest must preserve:

• immutable raw checksums, license, source revision, active Schedule, route, service date, trip instance, vehicle, stop sequence, parent station, direction, trunk route, and movement timestamp;
• exact Schedule-call identity as a separate mask: absence may disable plan-derived features or baselines but cannot erase an observed movement;
• source headway only as an audit label, plus explicit generation and ingestion clocks when the source actually provides them;
• enough leading and trailing context to derive departures and headways at the evaluation boundary without imputation;
• duplicate physical aliases and conflicts as measured outcomes; and
• event counts, lane rates, gaps, boundary loss, regime coverage, and byte cost before Stage 3 chooses windows, splits, clocks, or models.

If availability clocks cannot be recovered, the eventual claim remains retrospective event-time forecasting. The 30- and 60-second projections may be reintroduced in Stage 3 as matched controls, after the causal task and aggregation policy are frozen.

Limits¶

This witness is one line, one weekday, and two hours. Empty-cell and collision rates may change materially by line, service regime, and duration. The rare duplicate-alias case is synthetic here, though the retained seven-day source audit found one real duplicate. The experiment proves carrier mechanics and source-label equivalence only; it does not establish forecast sufficiency.