claims.md - the sole source for anything leaving the machine
Generated 2026-08-19T03:01:03+00:00 from the same build as outputs/summary.md. Every value is interpolated from the canonical dataframe; nothing here is typed by hand. If a number appears in a draft and is not in this file, it is not publishable.
Feed 20260818-20261017.
1. Brisbane's mainland has 190 suburbs and 1.24 million residents
Value — 190 suburbs / 1,242,389 residents
Produced by — load_suburbs, load_mesh_blocks, reconcile_population
Verification — Re-derived by spot check A with a different Excel reader and a shapely STRtree containment join: exact match
Mandatory caveat — Mainland BCC only: the five Moreton Bay island localities are excluded, so this is not the full Brisbane LGA. 2021 Census population against a 2026 timetable; growth suburbs are undercounted.
2. Residents beyond a 400 m walk of any serviced stop [HEADLINE]
Value — 476,945 (38.4%)
Produced by — citywide_network.network_access -> coverage_by_resident
Verification — Spot check C re-derived it with scipy min_only multi-source over appended stop nodes and sjoin_nearest snapping, cross-checked against an exact networkx Dijkstra on 250 sampled origins: agrees to the person
Mandatory caveat — Walking distance over the OSM pedestrian network; 9 suburbs (13,280 residents, 1.1%) fail the snap-quality bar and keep their straight-line figures, which reports MORE coverage than routing them would. ANY service means one scheduled trip a week counts. Proximity, not frequency. 2021 Census population against a 2026 timetable; growth suburbs are undercounted. It reproduces to the person under an independent implementation but is not accurate to the person: a defensible band is about +/-10%. See RED_TEAM.md attacks 2 and 8.
3. The headline under every measurement convention we tested [PUBLISHABLE SENTENCE]
Value — Under every measurement convention we tested, holding the 400 m threshold fixed, the figure lands between 468,290 and 495,185 residents (37.7% to 39.9%). We publish 476,945 (38.4%).
Produced by — conventions.run / conventions.summarise, 81 permutations over four axes (block origin, snap threshold, fallback rule, walk threshold)
Verification — One Dijkstra pass reused by every permutation, so the grid varies only the conventions. The published figure sits at the 68th percentile of the full grid and below the median (482,799) of this band
Mandatory caveat — Quote the band only with the 400 m threshold named as held. Allow the threshold to move +/-5% and the band widens to 427,076-533,840 (34.4%-43.0%), where a two-in-five framing fails at the bottom - at 420 m the honest phrasing is nearer one in three. 400 m is Queensland's own Priority Development Area standard, not ours, which is what makes holding it legitimate.
4. The single convention choice that moves the headline most
Value — the 400 m walk threshold: +/-5% moves it by 80,512 residents (17% of the headline)
Produced by — conventions.summarise, each axis varied with the other three held at their published settings
Verification — Six times the next largest conventional axis (block origin, 13,905). Every other convention in the pipeline is worth under 3% of the headline
Mandatory caveat — Two settings in the grid are bounds, not conventions, and must be described as such: polygon-edge-for-every-block (163,271 at its extreme) measures each resident from the corner of their block nearest a stop, and a 50 m snap threshold cuts the routed set from 176 suburbs to 67, so it reports mostly straight-line distance under another name. The full-grid band (13.1%-43.0%) should never be quoted without them.
5. Ransome's place on the zero-coverage list [FOOTNOTE ONLY - never the illustrative case]
Value — 0% coverage on the published basis, and in 30 of 36 conventional permutations - not all of them
Produced by — conventions.summarise stability pass over the grid; walkability.rule_on_suburbs for the ruling graph
Verification — Re-tested against all three weaknesses at once: its closest block is 477 m on foot (1.19x the threshold), its ruling-graph snap is 66.4% with trustworthy=0, and it is the only entry on the list that changes under any conventional permutation
Mandatory caveat — Three independent weaknesses, the only suburb on the list with more than one that matters. It stays on the list - dropping it would shrink our own finding in our own favour - but it must never be the case cited to illustrate it, and any mention needs the footnote. Cite Karana Downs, Mount Crosby, Upper Brookfield or Kholo instead: 8 of the 10 hold 0% in all 81 permutations. Eagle Farm (5 residents, 424 m) and Port Of Brisbane (3 residents) are also double-flagged but are immaterial by population.
6. The straight-line floor for the same test
Value — 109,732 (8.8%)
Produced by — mesh_block_access -> coverage_by_resident
Verification — Same spot check, same pass
Mandatory caveat — Straight-line, so it is a floor, not an estimate. Never publish it as the headline.
7. Beyond an 800 m walk of any serviced stop
Value — 98,880 (8.0%); floor 17,540 (1.4%)
Produced by — citywide_network.network_access
Verification — Same pass as the headline
Mandatory caveat — Walking distance over the OSM pedestrian network; 9 suburbs (13,280 residents, 1.1%) fail the snap-quality bar and keep their straight-line figures, which reports MORE coverage than routing them would.
8. Beyond a 400 m walk of an HOURLY service
Value — 531,863 (42.8%); floor 151,436
Produced by — stop_frequency (>=1 departure in >=10 of the 12 hours 07:00-19:00) -> network_access
Verification — Classifier asserted on two cases every run: Cultural Centre platforms must pass turn-up, Sumner's 452 stops must fail hourly. Build aborts if either fails
Mandatory caveat — HOURLY is a one-departure-an-hour test. It is NOT the RMIT half-hourly indicator -- see the reconciliation section.
9. Beyond a 400 m walk of a TURN-UP-AND-GO service
Value — 935,399 (75.3%); floor 656,905
Produced by — stop_frequency (>=4 departures in >=10 of the 12 hours)
Verification — Same classifier assertions
Mandatory caveat — Slightly MORE generous than TransLink's '15 minutes or better' branding: 94.7% of qualifying stops average <=15 min but only 29.6% never exceed it (median worst gap 18 min).
10. Beyond a 400 m walk of any Sunday service
Value — 619,584 (49.9%); floor 235,809
Produced by — stop_frequency (any boardable departure on the representative Sunday)
Verification — Representative Sunday chosen as the median-traffic Sunday inside the feed's own validity window
Mandatory caveat — One representative Sunday (20260927), not an average of all.
11. Suburbs where not one resident is within a 400 m walk of a stop
Value — 10 suburbs / 7,272 residents
Produced by — zero_coverage from pct_net_any_400m
Verification — Spot check D re-derived the list and total independently: exact match
Mandatory caveat — 2 of the 10 sit within 1.5x the threshold (Eagle Farm 424 m, 5 residents; Ransome 477 m, 268 residents) and would move if it did. Ransome is also the only entry that fails to hold 0% across the convention grid - see the Ransome row above; cite one of the 8 that hold in every permutation instead. The tail of the list is industrial and reserve blocks holding a handful of people.
12. Suburbs effectively unserved (<=5% of residents within a 400 m walk)
Value — 11 suburbs / 7,605 residents
Produced by — effectively_unserved from pct_net_any_400m
Verification — Same pass as the 0% list
Mandatory caveat — 5% is a chosen threshold, not a standard.
13. Suburbs whose nearest stop belongs to another council
Produced by — lga_of against the ABS 2021 LGA boundaries, from each suburb's largest populated block
Verification — Nearest stop measured from the largest POPULATED block, not from the suburb boundary, so it is a stop residents could actually reach
Mandatory caveat — Computed only for the effectively-unserved set. It is not a citywide cross-border census.
14. Suburbs with no serviced stop inside them or within 400 m [SECONDARY -- never the headline]
Value — 5 populated / 6,960 residents (9 suburbs / 6,963 including the unpopulated)
Produced by — unserved_400m from stops_within_400m
Verification — Spot check D re-derived it by buffering suburb polygons and querying a shapely STRtree of stops: exact match
Mandatory caveat — This is the retired whole-suburb binary. It turns on suburb shape and hides the 469,985 residents beyond the threshold inside suburbs that do have a stop somewhere. Publish only as a secondary measure, explicitly labelled.
15. Residents living in non-private dwellings
Value — 14,498 (1.2%)
Produced by — flag_institutional: no dwellings, or >=8 persons per dwelling
Verification — Re-derived block by block in the label sweep: exact match. ABS classifies prison mesh blocks as Residential, so the category field cannot find them
Mandatory caveat — A heuristic, not an ABS classification. It catches prisons, barracks, halls of residence, hospitals and aged care together.
16. Wacol's coverage counting and excluding institutional residents
Value — 51.3% -> 98.5% (3,526 of 4,273)
Produced by — coverage_by_resident with the institutional flag
Verification — Both figures printed in every worst-covered table
Mandatory caveat — Never cite Wacol as transport poverty. 83% of its residents are in the correctional precinct and immigration detention and are not free to travel.
17. Nathan moves the other way once institutional residents are removed
Value — 60.6% -> 43.7%
Produced by — Same
Verification — Same
Mandatory caveat — The institutional correction does not always flatter: Nathan's university accommodation is the well-served part.
18. Karana Downs' largest residential block to its nearest stop
Value — 4,776 m, in Ipswich
Produced by — story_far_suburbs -> lga_of
Verification — Straight-line from the largest populated block; the walking distance exceeds the 5 km routing limit entirely
Mandatory caveat — Straight-line. The walk is longer, and the stop is in another council's network.
19. Sumner's entire weekday service window
Value — 06:20-09:15 out, 15:22-18:51 back, route 452; no midday, no weekend except NightLink N449
Produced by — story_sumner from stop_times on the representative weekday
Verification — Asserted every run as a classifier verification case: Sumner's stops must fail HOURLY
Mandatory caveat — One representative weekday. Sumner also holds 7 institutional residents, immaterial here.
20. Best and worst covered ward
Value — Central 92.4% / Calamvale 34.5%, 38,496 beyond
Produced by — wards.aggregate from mesh blocks, ranked on the walking column
Verification — Spot check E re-derived the ward table by STRtree point-in-polygon and np.bincount: agrees to the person. Ward sums tile exactly to the mainland total or the build aborts
Mandatory caveat — Walking basis. On the straight-line floor the worst ward is Pullenvale and the best is The Gabba -- the ranking is not basis-independent, so never quote one basis's identity beside the other's number.
21. Residents whose best reachable stop leaves a >=3 h daytime hole
Value — 485,446 (39.5% of the 1,227,891 non-institutional residents; 39.1% of all 1,242,389)
Produced by — sumner_sweep.py, per populated mesh block, gaps anchored to the 07:00-19:00 window edges
Verification — Edge-anchored by necessity: unanchored gaps score Sumner as well served, which is the case the sweep exists to catch
Mandatory caveat — Different denominator from every other headline. It excludes institutional residents. Always name the base.
22. ...of whom have no reachable daytime service at all
Value — 472,339 (38.5% non-institutional; 38.0% of all)
Produced by — Same
Verification — Same
Mandatory caveat — Same denominator caveat.
23. Suburbs with weekday service but no weekend service at all
Value — 3: Nudgee Beach, Pinkenba, Sumner
Produced by — coverage trip visits on the three representative days
Verification — Measured at stops inside the boundary only
Mandatory caveat — Three representative days, not the whole feed. Stops just outside the boundary are not counted for this test.
24. The walking-network ruling on the five contested suburbs
Produced by — walkability.rule_on_suburbs, OSMnx/Overpass graphs, networkx single-source Dijkstra
Verification — An independent stack from the citywide build; 46 of 46 blocks give identical 400 m verdicts
Mandatory caveat — **Anstead and Pullenvale are published at their straight-line values (39.7%, 21.8%) because they fail the citywide snap bar, while the ruling gives 19.6% and 12.5%. Cite one or the other, never both. Ransome's ruling is self-flagged untrustworthy (66.4% snapped on the ruling graph); its 0.0% stands on the citywide evidence, where it snaps 100%.
25. Suburbs that keep straight-line figures because the network is untrustworthy there
Value — 9 suburbs / 13,280 residents (1.1%)
Produced by — citywide_network.suburb_quality, gate pct_well_snapped >= 80
Verification — Re-derived in the label sweep from pct_well_snapped: exact match. Every fallback reports MORE coverage than routing would
Mandatory caveat — The mixed measure biases toward the operator, so beyond-threshold counts stay floors on their own basis.
26. Our 400 m figure sits within six points of Podaris's independent isochrone measurement
Value — 64.9% of dwellings within 400 m against their 59% of households
Produced by — benchmarks.reconcile
Verification — Rebuilt on their unit (dwellings) from the same mesh blocks; their extent is ~1.8x ours and reaches sparser fringe, so they should read slightly worse, and they do
Mandatory caveat — At 800 m and 1.2 km the curves diverge sharply (96.8% against 83%) and we cannot fully attribute it. Quote the 400 m agreement only, and say so.
27. Our any-stop figure sits within four points of RMIT's 2018 network-distance measurement
Value — 64.9% of dwellings against their 61% of residences
Produced by — benchmarks.reconcile
Verification — Same threshold, same distance basis (pedestrian network), same unit (dwellings)
Mandatory caveat — Eight-year vintage gap and a wider study region.
28. Our half-hourly tier does NOT reconcile with RMIT's
Value — 29.7% (strict) / 44.6% (loose) against their 12%
Produced by — benchmarks._halfhourly_stops
Verification — Two readings of their wording, both computed; the stricter is still 2.5x more generous
Mandatory caveat — Unresolved divergence, and it cuts against us. If RMIT is right, more Brisbane residents lack a useful service than this analysis reports. Never claim the two agree.
29. GoThere's dataset is not comparable to ours
Value — 3 of its 6 'transport deserts' are outside Brisbane City Council; its 20 km ring excludes 4 of the 6 populated suburbs at 0% coverage
Produced by — benchmarks.reconcile
Verification — Its method read from its own methodology note: one WorldPop-weighted point per suburb, 800 m straight line, 205 suburbs within 20 km of Brisbane Central
Mandatory caveat — It publishes no population share at any radius. Comparing it to our 800 m figure is a category error; do not do it.
30. Serviced stops
Value — 6,008 inside Brisbane of 12,848 in the SEQ feed
Produced by — serviced_stops
Verification — Spot check B rebuilt the set with a different join order: exact match
Mandatory caveat — Rail platforms count separately. Stops with zero scheduled trips are dropped.
31. Representative days and feed validity
Value — 20260915 / 20261003 / 20260927; feed valid 20260818-20261017
Produced by — pick_representative_days, school_term_dates
Verification — Chosen as the median-traffic date per day type inside the feed's own validity window, weekday restricted to Tue-Thu of a school term; school term derived from a THRESHOLD on school-feed trips, not >0
Mandatory caveat — A single timetable snapshot. Re-running against a later feed will move every frequency figure.
32. No populated suburb without regular service is rescued by a school run
Value — 0 at 400 m; 1 at 800 m (Port Of Brisbane, unpopulated)
Produced by — school_stop_points, school_only_access
Verification — The school feed is a SUPERSET of the main feed: all 108,236 main-feed trips reappear in it, so the dedicated runs are the 3,096-trip set difference
Mandatory caveat — TransLink publishes the school feed as a pilot. Treat its coverage as indicative, not a service guarantee.