Closes #151. MATLAB sets resp.ind = ind_real (tuning_response.m:834) -- the RAW individual responses. The Python port returned the control-subtracted ones instead. `blankind` already matched; only `ind` differed. WHY THE DISTINCTION IS NOT COSMETIC These values can be complex. An F1 or F2 response carries a phase that depends on when the stimulus happened to start -- this module's own docstring says so -- and the control is measured in a different window, so the relative phase between a response and its control is arbitrary. Complex subtraction is vector subtraction, so that arbitrary phase moves the answer. For a signal of magnitude 10 against a control of 3: relative phase abs(a - b) 0 deg 7.00 90 deg 10.44 180 deg 13.00 abs(a) - abs(b) is 7.00 throughout. An 86% swing driven purely by a quantity that is not comparable in the first place. On a whole curve it does not merely add noise: with random control phases a suppression of -0.91 came out as a response of 4.67. Nothing needed the difference. The control individuals are returned alongside as `blankind`, and are recorded per point in the stimulus_tuningcurve document, so a caller comparing the two groups -- which is what neural_response_significance does -- compares magnitudes, and that is well defined. `curve` is still built from the subtracted means, as in MATLAB: a tuning curve is about response above baseline, and its mean is taken before the magnitude is. The two coexist in one struct on purpose. THE INDEX PATH IS LEFT ALONE MATLAB's oridirtuning.m:151 does complex-subtract for the vector and fit indices, and this does not change that. The same phase argument reaches it, but whether those indices should use abs(a) - abs(b) is a question about the analysis rather than the port, and it would move published numbers. Recorded in #151 for a decision by someone who knows the science; parity with MATLAB is kept meanwhile. TESTS Three, all of which fail against the previous behaviour: that `ind` holds the raw individuals, that `ind` and `curve` deliberately answer different questions, and that a complex response is not subtracted before its magnitude is taken. The absence of any assertion on `ind` -- the existing tests covered curve, blankresp, spont and spontind -- is what let this ship. ALSO FIXED: A STALE DOCSTRING I SHIPPED ndi.app.oridirtuning's module docstring still announced the index half as "NOT implemented, and deliberately not worked around", which stopped being true in the same PR that implemented it (#145). Rewritten to say where the math actually comes from and why two response structs exist. The comment in calculate_oridir_indexes explaining why it does not call this helper is updated too: the reason is no longer that the helper disagrees with MATLAB, but that the index call needs the subtracted individuals the helper does not return. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CqrNhfGnhKGcQvNiP2LNXK

Commits

Commits on Sep 2, 2026

Return the raw individual responses from tuningcurvedoc2vhlabrespstruct