[qmohsu]

Chapter 4's core math was verified clean by the ION GNSS+ pre-conference audit
(information equivalence 2.2e-16 at all 400 points, CRLB attained, image-source
re-derived). Every defect below is a demo, figure or README honesty defect, and
core/ is untouched.

The through-line: converged is a step-tolerance stop

In these solvers info["converged"] is set by norm(delta) < tol at
tol = 1e-6 m. The residual test above it cannot fire at all on noisy data —
four ranges over two unknowns leave a residual no solve can drive to zero. So
the honest reading of converged=False is "needed more than max_iters
steps", not "failed". Four demos filtered their samples on it.

Site	Reported	Honest	Why
example_toa_positioning Ex. 6	WLS improvement 23.1%	53.4% (theory 50.1%)	LS scored over 120/200 draws, WLS over 200/200. The 80 dropped were the worse half: RMSE 0.416 m vs 0.181 m for survivors, 0.298 m over all 200.
example_comparison RSS series	RMSE over converged (13/50 at 6 dB)	median over all 50	The console table beside it already reported a median over everything, so figure and table disagreed about the same run.
example_tdoa_positioning Demo 9	RW-LS 0.3163 m over 459 draws	0.3384 m over 500	Three sibling columns were already over 500.
Demo 4	both arms filtered	200/200 either way	No effect — worth recording, because the shape is not always live.

Raising max_iters to 50 gives 200/200 (Ex. 6) and 500/500 (Demo 9) with the
answers moving by 1e-4 m: the extra iterations were polishing, not rescuing.
example_comparison now runs all four methods through core.rf.solve_batch,
so the medians are over every fix and the failures live in their own panel on
solve_batch's four conditions.

Demos contradicting their own numbers

• TDOA Demo 2 drew independent noise per difference, deleting the shared
  reference error. err/noise ran 0.56–0.59; with e_j − e_0 it runs 0.80–0.88.
  Its "ratio constant" self-check passes either way — a statistic that tests
  only proportionality cannot see a wrong variance.
• TDOA Demo 6 seeded both solves at the truth and printed one draw to four
  decimals. From a neutral seed the collinear array's median over 500 is
  1.2687 m, not 0.9724. Seeded on the beacon line all 500 stall at
  5.3852 m — the seed-to-truth distance, identical across draws. That case
  is now a row of its own rather than a number that looks like an error.
• AOA Demo 7 printed "aligned anchors … cause high condition number and
  large PLE errors" above four rows in which the near-collinear array had the
  smallest error, the square the lowest condition number, and
  geometry_warning read "no" throughout. Two defects: the two columns drew
  independent noise, at one draw each. On 500 shared draws all four layouts
  land within 1.37×. The flag was right — it thresholds the bearing spread
  at the agent, which is 19.5° from (5, 7) however collinear the anchors are.
  A second table walks the agent onto the line, where the spread does collapse
  and the cost scales with range: −22.3 m of bias along the shared direction
  and 4.6 m of scatter at 30 m out, against 0.13 m across it.

Figure

ch4_rf_comparison plotted AOA's [0,1,3,5,10] degrees and RSS's
[0,2,4,6,8] dB at TOA's metre positions under an axis reading
"Measurement Noise (m)". x is a level index now, each series carries its own
schedule in the legend, and the CDF and box-plot titles name all four methods'
noise at that level. Panel 1 is labelled median (it was an unlabelled RMSE);
panel 4 is "Fixes solved".

--data defaulted to the same figure basename as the inline run, so it
overwrote the committed comparison with a different picture. It writes
ch4_rf_comparison_dataset now.

ch4_initial_guess_basin no longer reproduced from HEAD — its traced seed
had moved. Regenerated; the classification counts are unchanged
(785/341/896/82/181/522/1340) and only panel 4's trace differs. Proven
idempotent: two further full regenerations leave every byte identical.

READMEs

• Chapter: basin trace (−4.5,−2.0) → 9.4e11 m in 18 iters → (12,12) → 4.43e12 m in 14; 305/196 → 308/182; PLE's "sensitive to aligned beacons"
  rewritten to what Demo 7 measures.
• ch4_rf_2d_square: the stale "TDOA ~0.07 m … most accurate … GDOP lowest"
  sentence, which the same file disowns 240 lines earlier; the geometry and
  TOA-noise tables rebuilt from the shipped data (the old ones put the optimal
  array at 0.8–0.9 GDOP, below the theoretical floor of 1 for four beacons, and
  named a Clustered geometry --geometry has no choice for); "twelve
  times" → eleven (11.95/1.07); 1.75 → 1.76 (10·tan 10°); "20-50%" → the
  measured per-method 10% / −31% / 31%.
• ch4_rf_2d_linear: 111 → 113, which its own two other lines already said.
• ch4_rf_2d_nlos: "much larger than 0.818 m" → the measured 0.635 m,
  which is smaller, plus why (only 2 of 4 beacons biased, so LS splits the
  difference) and the shape test that says it is bias — median 0.615, mean
  0.607, RMSE 0.635 all agree, where noise spreads them.
• example_tdoa_positioning printed "always beneficial" above a −0.3% row.
  Qualified: that row's true gain is under 1% and a ratio of two RMSEs over 200
  draws carries ~5% standard error; over eight other seeds it runs +0.15% to
  +0.94%.

Minors

--compare-geometry's DOP column gained a unit column (AOA's entry is m/rad);
the covariance-matrix figure's half-integer ticks are now one per cell named
d1,0 … d4,0; closed_form_comparison's four boxes are all over 500 draws.

Test plan

• tests/ch4_rf_point_positioning tests/core/rf — 243 passed
• tests/docs (minus notebooks) — 866 passed, 1 skipped
• tests/test_datasets_reproduce_from_their_recipe.py — 43 passed, no
  dataset byte moved
• tests/test_every_figure_has_a_demo_behind_it.py — 200 passed, 8 skipped
• test_repo_conventions / test_lint_debt_only_shrinks /
  test_black_is_actually_run / test_mypy_debt_only_shrinks /
  test_doctests_that_pass_keep_passing / test_python_floor_matches_syntax
  / test_repo_validators / test_semantic_api_debt_only_shrinks — green
• Zero-noise rows still exact: TOA/TDOA/AOA all 0.000 at level 1 in
  example_comparison; TDOA Demo 2's 0.00 row unchanged; TOA Examples 1–5
  byte-identical
• Every changed PNG opened and read

🤖 Generated with Claude Code