Foreign instrument 01 · outcome: broke
TRANSFER BROKE. Asked to place 150 quantisation marks across the full 2024 ABS population raster, canonical Floyd-Steinberg emitted only 40 to 42 across four scan rotations; the locked state-first fixed-count rival emitted 150. Operate the maps, inspect the discarded residual, and test the provenance error the printer cannot see. These are marks, not seats: the experiment includes no electorates, voters, contiguity, communities of interest, constitutional minima, or redistribution law.
Try the count now: choose the declared fallback
Quantisation marks, not seats. This page contains no electorates, contiguity, voter eligibility, communities of interest, constitutional minima, redistribution law, or representation of people as members of Parliament. It tests whether an image algorithm keeps a fixed count when its input comes from a population raster. It is not an Australian electoral proposal.
I · home calibration
Before population enters the apparatus, this page reproduces an in-house synthetic benchmark from The Grey That Isn't There. Every value in the live column is computed now from the canonical 7, 3, 5, 1 kernel. Floyd and Steinberg 1976 supports the algorithm citation, not these benchmark numbers.
Running live calibration
| Native check | In-house expected | Computed live |
|---|---|---|
| Kernel conservation | 1 = 7/16 + 3/16 + 5/16 + 1/16 | computing |
| Flat 0.2, Floyd / threshold coverage | 19.9% / 0.0% | computing |
| Ramp MAE, threshold / Bayer / Atkinson / Floyd | 0.2423 / 0.0083 / 0.0499 / 0.0041 | computing |
| Sphere MAE, same order | 0.2033 / 0.0110 / 0.0655 / 0.0080 | computing |
| Sphere Floyd MAE, sigma 1.2 / 3.5 | 0.0164 / 0.0097 | computing |
| Forty quarters, full carry | 10 marks; first eight 01000100 | computing |
Correction to the scout: these reconstruction values are mean absolute errors, because that is what the cited native verifier computes. Calling them RMSE would be false.
II · the finite edge
Weights that add to one only conserve error while all four future neighbours exist. At an image edge, canonical handling drops the fractions aimed outside. A two-by-two raster is enough to break a fixed count exactly.
Correction to the scout's exact trace. It printed 291/512 and -4027/8192 for the last two decisions. Direct rational propagation with the canonical below weight 5/16 gives 259/512 and -3227/8192. The scout's three-mark output remains right. This page and its verifier test every fraction.
III · foreign target
The full 7,401 by 4,100 ABS rectangle contains 30,344,100 one-kilometre cells. Its positive values are scaled to a total requested ink mass of 150. NoData and valid zero cells stay in the scan as zero levels. Nothing is cropped, wrapped, rescaled to the workbook, or forced at the last pixel.
Loading the frozen full-grid run.
Floyd-Steinberg markdeclared fallback markrequested massNoData shown in grey
The four canonical rotations emit 42, 42, 40, and 41 marks. Their largest pairwise Jaccard distance is 1.000. The failure is not a small rearrangement around a stable fixed count.
IV · declared fallback comparator
The overnight build uses the scout's declared smaller fallback: Hamilton largest remainder fixes state totals from the joined 2024 raster sums, then scalar residual carry chooses locations within each state. This is weaker than a fully validated unequal-probability spatially balanced sampler, so the page does not call it that sampler. It is still a stronger target-native rival than random dots because it preserves both the national count and predeclared state totals.
| Window side | Selected method p99 | Other method p99 | Worst-scan FS / rival | Selected max |
|---|---|---|---|---|
| loading |
The declared fallback emits 150 in every rotation. At 500 km the worst-scan 99th-percentile discrepancy is 22.9200 for Floyd-Steinberg and 1.7913 for the fallback, a ratio of 12.795. The 25 and 100 km ratios are 0.988 and 1.018, both inside the locked 10% tie band. Ocean-only windows make the smaller-scale percentile weak evidence, and the national count failure decides BROKE before any spatial ranking is needed.
The declared fallback emits 150 in every rotation. It is a secondary comparator because the nonzero national count error already decides BROKE.
| State | Requested mass | Selected emitted | Signed error | Rival fixed total |
|---|---|---|---|---|
| loading |
The maximum absolute state quota error across the four Floyd-Steinberg scans is 30.7565 marks. State grouping does not cause the national count failure, but it shows how badly a global image lattice can miss a target-field aggregate.
V · independent target checks
The raster's positive sum is 27,194,097.408977121. The independently fetched ABS workbook says 27,194,369, a raster-minus-workbook gap of -271.59102287888527. Both are ABS products, so this checks extraction and labelling, not independent measurement. The adapter is not rescaled to hide the gap.
The centre-point state join leaves 430,317.240833879 raster residents outside the nine ASGS polygons. That is a consequence of the cell-centre convention, especially around coasts and external areas, and is reported rather than pushed into a nearby state. Two independently written joins agreed on 64 frozen, spread-out sample cells.
State interpretation is mapping-sensitive. An independent maximum-cell-overlap sensitivity reduces the unassigned set from 2,248 cells and 430,317.240833879 modelled residents to 3 cells and 27 residents. The national 40 to 42 count does not use state labels and is unchanged, but the centre-joined state errors must not be read as facts about population representation.
| State or territory | Raster centre-joined | Workbook 2024 | Raster minus workbook |
|---|---|---|---|
| loading |
The independently dated AEC determination is 23 July 2026: 46, 38, 30, 16, 10, 5, 3, 2, total 150. It uses later population statistics, a quota derived from state populations and twice the number of state senators, Tasmania's constitutional minimum, and separate territory rules. It is a label, date, identity, and transcription check. It is not a numerical target for the 30 June 2024 raster.
VI · adapter audit
Every row below destroys target meaning. The imported object is ink density on a finite rectangular lattice. Nothing in the source says that a cell is a constituency, that a diagonal is a political neighbour, or that east should happen before south.
| Source field | Formal object | What the mapping loses |
|---|---|---|
| loading |
Choose any adapter row. Computable ablations were precomputed by the same full-grid scan function as the headline. Undefined removals say so instead of inventing a number. The verifier reruns every computable row from the frozen research data.
loading
Result: loading
Regenerated by this layer's extraction script, build-extract.mjs, run against the pinned sources above with a local image decoder on the path.
The largest input level at H=150 is below 0.2. Binary capacity would hold only through H=761 for this file, from floor(27,194,097.408977121 / 35,702.984375). The displayed run stays at 150.
VII · fault bench
Unmodified apparatus
Engine: PASS, kernel 16/16 and carry pattern intact
Provenance audit: PASS, 2024 label and SHA-256 agree
Both arithmetic and provenance checks are green.
Each button mutates a cloned input and runs the same validators as the clean apparatus. The 6/16 coefficient enters the live benchmark. The flipped carry record enters the trace validator. The provenance fault changes only the displayed period of the pinned 2024 target package to 2023, so quantisation inputs remain unchanged while the package validator catches the false label. No unshipped older grid is claimed.
Checking six shipped target files against frozen SHA-256 values.
T = (110; 0.988, 1.018, 12.795; 30.7565; 1.000).
Components are maximum count error, worst-rotation imported-to-rival p99 ratios at 25, 100, and 500 km, maximum state quota error, and maximum pairwise scan Jaccard distance. Any nonzero first component makes the verdict BROKE.
30,344,100 cells; 22,650,905 NoData; 7,408,120 valid zeros; 285,075 positive cells. H=150, four rotations, and 48,544 / 47,173 / 40,165 aligned windows at 25 / 100 / 500 km.
loading
The source extraction was run sequentially under a 2 GB budget. The browser receives sparse marks, simplified display geometry, and exact 25 km mass blocks, not the full raster or 75 MB boundaries.
H=150, tie to 1, four rotations, NoData as zero, cell-centre state join, 25 km origin lattice, three window sizes, and a 10% tie band are supplied. The grid is modelled population, not exact locations; it publishes no cell intervals. Boundary vintage differs. The completed rival is the declared smaller fallback, not the strongest published spatially balanced sampler.
| Source | HTTP | Bytes | Content type | Licence | Retrieved | SHA-256 |
|---|---|---|---|---|---|---|
| loading |
We searched the open web, OpenAlex, and Crossref on 10 September 2026 using the query families listed in the audit and did not find a published or indexed application of Floyd-Steinberg error diffusion to legislative-seat apportionment or to fixed-count parliamentary marks derived from a population raster. This limited search does not cover unindexed code, art, non-English literature, Scopus, or Web of Science.
VIII · verdict
The result sentence will be selected only after the shipped data and declared rule agree.
The failure says something precise. Floyd-Steinberg preserves local average tone in its native field, but a fixed national total is a different invariant. Here residual diffuses through millions of zero and NoData cells and leaves the finite rectangle. Boundary repairs would define a different algorithm.
Sources
ABS grid, workbook, and boundary data are attributed to the Australian Bureau of Statistics. AEC factual entitlement material is attributed © Commonwealth of Australia 2017. The Gazette transcription names the Federal Register of Legislation as source. Source material is used under Creative Commons Attribution 4.0 International, subject to each source's stated exclusions. Retrieved 10 September 2026. This page uses transformed extracts, simplified display geometry, computed sensitivities, and short factual transcriptions; those changes are not endorsed by the source agencies.