Fallback policies

Does loop avoidance improve fallback completion, and does a Grid detour justify its cost?

DERIVED statistics and accounting MODELLED network outcomes

Pre-registered hypothesis

Greedy-with-visited reduces stuck/loop losses relative to pure greedy; Grid detour reduces those losses further or improves stretch, but may add hops. Continue/drop can win on cost or delivered-packet latency and must remain visible.

Frozen preregistration, identified and verified against this result’s manifest.

Verdict

Mixed: greedy-with-visited reduces the stuck/loop loss rate against pure greedy in 2 of 2 shells; the Grid detour shows no extra benefit, and continue wins triggered completion in C5.

Stuck/loop loss rate, greedy_visited minus greedy_pure (paired Student-t 95%): C5 -0.222 [-0.365, -0.0796], Holm-adjusted p 0.00323; C2 -0.306 [-0.464, -0.147], Holm-adjusted p 0.000777.
Grid detour minus greedy_visited: C5 0.0556 [-0.023, 0.134], Holm-adjusted p 0.32; C2 0 [0, 0], Holm-adjusted p 1 (identical in every run).
Continue completes more triggered packets than greedy_visited in C5: continue 1 vs greedy_visited 0.833.

Every number in this verdict is computed at build time from the verified summary.csv and the registered paired comparisons in paired_contrasts.csv (SHA-256 a221c1c49cefa741f6fa970fde7fa067f767f2de94bdd19c1c57c3638af03f3c, checked against the manifest).

C5 orbital inputs are REAL public data; C2 orbits and positions are MODELLED. Topology, probes, faults and network behaviour are MODELLED. Starlink’s operational routing is UNKNOWN; Starlink does not use T-CRH.

Charts and data

Download CSV · Download manifest · Download paired contrasts

Source: results/E05/20261011T185503Z-confirmatory/summary.csv. SHA-256: 4df763a2e9a0ea8ba5cfcc5a31826cca04a73c9f91e457d5239b0b0ee3976df3.

This first view shows the 5 rows of the initial selection, out of 12. With scripting, the page loads the same summary.csv offered for download and checks its SHA-256 against the manifest before other selections are enabled.

ConstellationConditionChart metric

Triggered-packet delivery ratio (proportion). Shaded bands show recorded 95% confidence intervals over runs, not packets.

Data table

Search this selection Show all CSV columns

Showing 1–5 of 5 matching rows. The download contains every original row and column.

Values from summary.csv. Empty fields are shown as undefined; infinite effective latency remains infinite; a p-value stored as 0 is shown as below numerical resolution.
const	condition	scheme	n	complete	triggered	triggered_pdr_mean	triggered_pdr_lo	triggered_pdr_hi	stuck_rate_mean	stuck_rate_lo	stuck_rate_hi	triggered_stretch_mean_mean	triggered_extra_hops_mean_mean	infinite_tail_runs	stretch_min_exact	stretch_min_adjusted	stretch_min_raw
C2	trigger	continue	36	True	14400	1.0	1.0	1.0	0.0	0.0	0.0	1.0000032216042933	0.5	0	1.0	1.0	0.9999608004211481
C2	trigger	drop	36	True	14400	0.0	0.0	0.0	0.0	0.0	0.0	undefined	undefined	36	1.0	1.0	0.9999608004211481
C2	trigger	greedy_pure	36	True	14400	0.6944444444444444	0.5363746271745987	0.8525142617142901	0.3055555555555556	0.14748573828570985	0.4636253728254013	undefined	undefined	11	1.0	1.0	0.9999608004211481
C2	trigger	greedy_visited	36	True	14400	1.0	1.0	1.0	0.0	0.0	0.0	1.1955543015079344	1.8055555555555556	0	1.0	1.0	0.9999608004211481
C2	trigger	grid_detour	36	True	14400	1.0	1.0	1.0	0.0	0.0	0.0	1.1955543015079344	1.8055555555555556	0	1.0	1.0	0.9999608004211481

Where ChronoRoute loses

Every listed point disadvantage is retained, including privileged-oracle advantages. Where a registered paired comparison exists for the same cell, its paired Student-t 95% interval and Holm-adjusted p-value are quoted; otherwise marginal intervals alone do not establish a significant loss. Ties in the predicted stuck-loss reduction are listed as unsupported advantages.

Continue completing more triggered packets than a fallback policy is a real loss for the fallback, not an artefact.

A Grid detour that does not reduce stuck losses below greedy-with-visited is listed as an unsupported advantage.

Descriptive comparisons from the same summary; source row numbers count data rows, excluding the header.
Cell	Metric	Policy	ChronoRoute	Comparator	Comparator value	Interpretation	Source rows
C2 / trigger	triggered_stretch_mean_mean	greedy_visited	1.1955543015079344	continue	1.0000032216042933	Point disadvantage; adjusted paired significance unavailable	5, 2
C2 / trigger	triggered_extra_hops_mean_mean	greedy_visited	1.8055555555555556	continue	0.5	Point disadvantage; adjusted paired significance unavailable	5, 2
C2 / trigger	triggered_stretch_mean_mean	grid_detour	1.1955543015079344	continue	1.0000032216042933	Point disadvantage; adjusted paired significance unavailable	6, 2
C2 / trigger	triggered_extra_hops_mean_mean	grid_detour	1.8055555555555556	continue	0.5	Point disadvantage; adjusted paired significance unavailable	6, 2
C2 / trigger	stuck_rate_mean	grid_detour	0.0	greedy_visited	0.0	Predicted reduction unsupported (tie) Not significant after adjustment; paired difference 0 [0, 0] (Student-t 95%), Holm-adjusted p 1	6, 5
C5 / trigger	triggered_pdr_mean	greedy_visited	0.8333333333333334	continue	1.0	Point disadvantage; adjusted paired significance unavailable	11, 8
C5 / trigger	stuck_rate_mean	greedy_visited	0.16666666666666666	continue	0.0	Point disadvantage; adjusted paired significance unavailable	11, 8
C5 / trigger	stuck_rate_mean	greedy_visited	0.16666666666666666	drop	0.0	Point disadvantage; adjusted paired significance unavailable	11, 9
C5 / trigger	triggered_pdr_mean	grid_detour	0.7777777777777778	continue	1.0	Point disadvantage; adjusted paired significance unavailable	12, 8
C5 / trigger	stuck_rate_mean	grid_detour	0.2222222222222222	continue	0.0	Point disadvantage; adjusted paired significance unavailable	12, 8
C5 / trigger	stuck_rate_mean	grid_detour	0.2222222222222222	drop	0.0	Point disadvantage; adjusted paired significance unavailable	12, 9
C5 / trigger	triggered_pdr_mean	grid_detour	0.7777777777777778	greedy_visited	0.8333333333333334	Not significant after adjustment; paired difference -0.0556 [-0.134, 0.023] (Student-t 95%), Holm-adjusted p 0.32	12, 11
C5 / trigger	stuck_rate_mean	grid_detour	0.2222222222222222	greedy_visited	0.16666666666666666	Not significant after adjustment; paired difference 0.0556 [-0.023, 0.134] (Student-t 95%), Holm-adjusted p 0.32	12, 11

Limitations

The intervention forces the checker decision at an on-path node without adding service delay; these are policy outcomes, not false-positive data.
Triggered-only completion and cost prevent untriggered traffic from hiding policy effects.
Losses have infinite effective latency. Undefined trigger denominators are not zero.
Chart intervals are marginal per-cell intervals. The verdict quotes the registered paired comparisons from paired_contrasts.csv: paired Student-t 95% intervals (primary) with Holm-adjusted p-values.
Continue never reroutes, so it can complete more triggered packets than a fallback when the planned path still works; it remains a visible competitor.
Raw stretch below unity can result from step crossing; use the recorded exact-domain and tolerance-adjusted checks.

These pages use only summary.csv and manifest.json, plus paired_contrasts.csv for the verdict and the paired comparisons quoted above; each file is checked against the manifest SHA-256. They do not infer missing comparisons from figures, reports, raw traces or other result files.

How this was made

Recorded producer

E05 confirmatory campaign (1.0.0.dev0)

Manifest commit

a18cffb6da450ea4449acf46f455b474fee2dcbf; dirty: false

Run

9610fa72ee94fe54

Time

2026-10-11T18:56:03.141804Z to 2026-10-11T19:01:20.883292Z; 317.741477 seconds of recorded host wall time (REAL)

Snapshot

20261010T1358Z-celestrak; SHA-256: c37e70812155c2d1d9eeb3e96681303c00c9fca91287939b21bf5313ce83aecd

Seeds

1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035

Command

venv/bin/python -B experiments/policy_campaign.py --out results/E05/20261011T185503Z-confirmatory --workers 8 --prior results/E04/20261011T185503Z-collected --experiment E05

Config SHA-256

6cdfbf1c65becdefa1499f461e68ccf145296cd2082f768a94c1de85e093d95b

Labels

isl_topology_grid4: MODELLED
link_rate_and_queues: MODELLED
orbital_elements: REAL
routes_and_budgets: DERIVED
traffic_probes: MODELLED

C2 override: C2 elements and positions MODELLED

Manifest

chronoroute.manifest/1; 450 recorded output files (5088698 bytes); 8 recorded input hashes. Download the full manifest (84875 bytes, SHA-256 c93abef7351d9ff2db99c9b66fc8d047c58b693cea7c989acc1596876826f3a7) for the configuration, versions, host and every input and output hash.

summary.csv: 17090 bytes, SHA-256 4df763a2e9a0ea8ba5cfcc5a31826cca04a73c9f91e457d5239b0b0ee3976df3
paired_contrasts.csv: 4356 bytes, SHA-256 a221c1c49cefa741f6fa970fde7fa067f767f2de94bdd19c1c57c3638af03f3c
Versions: chronoroute 1.0.0.dev0; matplotlib 3.11.1; networkx 3.6.1; numpy 2.5.3; python 3.14.4; sgp4 2.27
Host (REAL): cpu_model: 13th Gen Intel(R) Core(TM) i9-13900H; logical_cores: 20; machine: x86_64; platform: Linux-6.18.33.2-microsoft-standard-WSL2-x86_64-with-glibc2.43; ram_bytes: 8131358720

For collected results, the manifest commit identifies the collector. Input hashes preserve the producer chain in the downloaded manifest; this page does not load those additional files.