[qmohsu]

Eight audit findings, each re-measured before it was fixed. No dataset byte moved.

1. The −0.34 mechanism belonged to data that does not ship

report_what_a_repeat_survey_buys explained MAP's loss to 1-NN by an anti-correlation between the spatial mismatch and the per-visit σ, and explicitly called the estimation-noise diagnosis "wrong". Measured on the shipped multisample survey that correlation is +0.0156 — zero, because the fast fading is one constant for the whole building.

The estimation-noise diagnosis is right, and the demo now runs the proof instead of asserting it. A new oracle row substitutes the true constant σ into the fitted model:

  min_std     sigma range    MAP!=NN   max wt  eff RPs      MAP       NN
      0.5   0.50 -  3.02     44/200   0.9873     1.03   3.72 m   3.16 m
      ...
   oracle   1.50 -  1.50      0/200   0.9788     1.04   3.16 m   3.16 m

The anti-correlation survives as a labelled conditional about a non-shipped variant. Its dose-response column needed dense and sparse repeat surveys that do not exist, so it is gone; the spatial-mismatch column is shipped-verifiable and re-measured at 0.54 / 1.43 / 2.98 dB.

2. Two stale constants

"2.27 m quantisation floor" matches nothing measurable. sqrt(2 s²/12) = 2.041 for s = 5; sampling 200k uniform positions gives 2.043 rms and 1.91 mean; clean-map NN achieves 2.176. RMS is now the named statistic, because everything compared against it is an RMSE, and the mean is labelled as the mean in the dense/sparse survey-effort tables. The "3.39 m" headline row re-measures to 3.01 m (20 seeds: 2.95–3.44).

3. The sparse grid is 6×6

arange(0, 55, 10) has six nodes. 36 RPs/floor and 108 total, not 25/75.

4. example_pattern_recognition demonstrated nothing

diag(Z'Z) is 3.6e5–4.3e5, so every swept λ (0, 0.1, 1, 10) was under 3e-5 of the data term — train RMSE printed 6.09 four times under "Regularization prevents overfitting", with train below test at every λ.

Now: λ ∈ {0, 10, 100, 1e3, 1e4} (measured, the knob bites near 1e2); symlog x so λ=0 is in the panels; and a training-set-size sweep supplying the overfitting the 85-RP split lacks — at 9 training points for 8 features λ=0 fits exactly and scores 167 m, λ=100 returns 10.5 m.

5. Coincident series erased each other

NN ≡ MAP and PostMean ≡ PostMean(k=10), so panels B/F/I showed four series for six legend entries. One shared stroke and marker table now covers B, E, F and I; Key Insight 6 names both causes with both gaps computed.

6. example_probabilistic never stated its degeneracy

Prints sigma_summary() per model, captions the flat MAP line, dashes apart the three identical MAP CDFs, and the x=y reference spans both axes instead of being the zero-length segment the legend advertised.

7. Two test files asserted nothing

Seven test_-named functions with zero asserts, printing [OK]/[FAIL] and returning True; pytest passed all seven regardless. Now 13 assertions, each verified by mutation. One overturned the old file's own printed claim: widening an RP's σ widens its basin rather than reducing its influence — the if guarding that claim had silently never fired.

8. Minors

2.86x → 1.09x and 30x → 60.5x (operations, sourced); scenario labels naming 4 dB shadowing where the unlabelled term is 1.5 dB fast fading; panel G's n=1 category; the σ=0 row in the classification figure; test_shadowing_field's "never exceeded 0.63" measured at 0.658.

Test plan

• tests/core/fingerprinting tests/ch5_fingerprinting — 287 passed
• tests/test_datasets_reproduce_from_their_recipe.py — 43 passed, no data byte moved
• tests/docs — 880 passed, 1 skipped
• validators, semantic, lint, mypy, black, doctest, repo-conventions — 845 passed in the consolidated run
• all six ch5 examples run; four figures regenerated and opened
• MAP≡NN invariants: single-sample 0/200 differ, multisample 44/200 (22%) at min_std 0.5, effective RP count 1.03–1.79 (posterior non-delta)

🤖 Generated with Claude Code