Named-hardware real-data soak evidence — 2026-07-28

This ledger records the first completed one-hour and eight-hour executions of the v4 soak contract on named hardware. It is evidence for the Phase 3 resource-measurement gate, not a universal performance claim. The compact machine-readable record is real-data-soak-2026-07-28.json.

Fixed identities

Item	Recorded identity
Hardware	sasaki-laptop-i5-1145G7-32GiB-jazzy-native; Intel i5-1145G7, 8 logical CPUs, 32 GiB, x86_64
OS / ROS	Ubuntu 24.04.4 LTS, kernel 6.17.0-35-generic, ROS 2 Jazzy
Product and harness	clean commit 0ec55575ffc16eb008e9f24bd6c6f24700bf2f8a, product 0.6.0
Map profile	pointcloud_gnss_smoke
Bag metadata	SHA-256 1b9be8f18b0cbbb369c3f38549ac4c3bb7f8bb895ef417c0fbe57a0a926d2f71
Bag storage	istanbul_bag4_0.db3, 1,061,437,440 bytes, SHA-256 e833a6ae89d5f0ccf0ee8d4bdde3232e7c49d04a8342db452b4ad743ef43b185

Both reports recorded the same clean software, harness and input identities. The full reports were validated against soak-report-v4.schema.json.

Results

Profile	Result	Iterations	Wall time	Longest iteration	Peak RSS	Output	Minimum free	Drops
one-hour	PASS	86 / 86	3,638.645 s	43.164 s / 300 s	209.320 MiB / 1,024 MiB	1,889,888,063 B / 5 GiB	85,689,724,928 B / 60 GiB	0 / 0
eight-hour	PASS	671 / 671	28,834.115 s	45.732 s / 300 s	210.871 MiB / 1,024 MiB	14,553,359,036 B / 30 GiB	70,752,927,744 B / 40 GiB	0 / 0

All eight terminal checks were true for both runs: all_iterations_succeeded, dropped_inputs_within_budget, free_space_within_budget, iteration_duration_within_budget, output_size_within_budget, peak_rss_within_budget, provenance_recorded, and target_duration_reached.

The one-hour report SHA-256 is 25a677d0944fec42bc63421c67a3a055993afd6d44a1991dfe5b49b973f35a3e. The eight-hour report SHA-256 is 67aace0d8521178357bdc3264acdeca57e35c414ecf9dd516b29cf9612eb1361. After the eight-hour run, no process scoped to the evidence worktree remained and the detached evidence worktree was clean.

Reproduction template

Use the exact input represented by the hashes above, a clean checkout of the recorded commit, and budgets appropriate for the target machine:

python3 scripts/run_map_soak.py /data/bags/istanbul_bag4_uncompressed \
  --output-root /data/soak/one-hour \
  --soak-profile one-hour \
  --hardware-label '<non-secret-machine-label>' \
  --map-profile pointcloud_gnss_smoke \
  --max-peak-rss-mib 1024 \
  --max-output-gib 5 \
  --max-dropped-inputs 0 \
  --max-iteration-secs 300 \
  --min-free-space-gib 60 \
  --telemetry-interval-secs 30

python3 scripts/run_map_soak.py /data/bags/istanbul_bag4_uncompressed \
  --output-root /data/soak/eight-hour \
  --soak-profile eight-hour \
  --hardware-label '<non-secret-machine-label>' \
  --map-profile pointcloud_gnss_smoke \
  --max-peak-rss-mib 1024 \
  --max-output-gib 30 \
  --max-dropped-inputs 0 \
  --max-iteration-secs 300 \
  --min-free-space-gib 40 \
  --telemetry-interval-secs 30

Do not compare a rerun to these numbers unless its input hashes, product commit, map profile and resource budgets are recorded. A passing soak proves bounded operational behavior for this configuration; map geometry quality remains covered by the separate real-data and release gates.