The model's own error estimate

The artifact stores enough to say how uncertain it is at a parameter point, not just to produce a waveform. Written 2026-09-02, when it first covered all seven modes over the whole waveform; updated the same day with the owner's decision on the two-route blend (checklist B5), below.

What you get

times, errors = model.get_waveform_error_estimate(mass_ratio, spin1z, spin2z)
delta_A, delta_phi = errors[(2, 2)]

The name and the argument spelling already matched the public surface settled by checklist A7, so nothing about this call changed there; the times it returns are the same geometric, peak-aligned samples get_td_waveform_modes_geometric returns, and waveform-evaluation.md covers those.

Seven modes, 15775 samples each, over the full inspiral, merger and ringdown. delta_A is a 1-sigma amplitude uncertainty in the units of the mode content; delta_phi is delta_A / |c_lm| in radians, and is meaningful ONLY where the mode stands well above its own envelope.

It is a band, not a number. Three stages, all per-sample:

predict_var     (14, 25)     variance of each Chebyshev coefficient
                             14 elements by 25 coefficients
error_envelope  (N,)         1 sigma at every sample
mode_errors     two (N,)     amplitude and phase

The envelope propagates the coefficient variances through the same Chebyshev evaluation the waveform uses, env(Phi)^2 = sum_j T_j(x(Phi))^2 Var[C_j] on the element containing each sample, so the band is wide where the fit is poorly determined and narrow where it is not.

What it reports, measured

Relative amplitude uncertainty at q = 4, chi = (0.3, -0.2), on the released artifact:

It widens by about 86x from inspiral to ringdown on (2,2), which is the expected direction.

Where it comes from, and one trap

The surrogate is a Gaussian process regression, so fitting returns a coefficient AND its variance. Those stored variances are what is reported: the fit's own statement of how well it pinned each coefficient down, which grows away from training simulations.

The merger-ringdown envelope must NOT come from error_envelope. Under the SWSpH representation the regressed coefficients are the detrended spheroidal residual, so error_envelope returns the uncertainty of a DIFFERENT quantity, before re-winding, mixing and the remnant-map term. Wrong, not merely uncalibrated, and it would look plausible. spherical_plunge_spheroidal_ringdown_error_envelope is the correct source and the dispatch is now guarded, so an unrecognised arm raises. The released arm is SWSpH.

The two envelopes are joined with the SAME blend window the waveform uses, INSPIRAL_MERGER_RINGDOWN_BLEND, named once for both joins.

remnant_sigma defaults to (1.2e-4, 5.3e-5), sourced exactly to loo_Mf_rms = 1.1987e-04 and loo_chif_rms = 5.30597e-05 in paper_data/swsph_remnant_map.npz. Three qualifications: the M_f number excludes SXS:BBH:1124, which would make it 2.7e-4; they are RMS rather than maxima; and that file's dphi_tot_lm columns are quoted at t_end = 55.5 M while the arm reaches +124.5, and that term grows linearly, so those columns are 2.24x low.

Limits

Calibration, not blindness. The merger-ringdown arm's error is parameter-space regression essentially in full, e_gpr 3.44e-03 against e_own 2.55e-05, a 115x ratio, which is exactly the kind of error a GPR variance is meant to describe. The open question is whether the band is correctly SIZED. One measurement bears on it: the in-sample full-IMR mismatch of 7.90e-07 on (2,2) implies about 1.26e-03 relative error while the envelope reports 4.77e-04, so it is a factor 2.6 OPTIMISTIC, in sample. Out of sample it will be worse by an unmeasured amount. See [[error-aware-likelihood]], whose limit 1 carries the retraction of an earlier and wrong framing of this.

The timing term is absent, and it is the largest known omission. m omega sigma_tau EXCEEDS the reported content phase uncertainty on every mode, by 1.7x to 28.5x, and by 28.5x on (2,2). Folding it in is blocked on measurement rather than code: the tau to Phi(t) map is an inversion, so any folded term is a linearisation with unmeasured second order, the clock and the content are fitted from the same simulations so they are not independent, and folding it moves four published numbers by more than an order of magnitude.

Only the diagonal coefficient variances are propagated, so coefficient correlations are discarded.

The reported uncertainty is that of the route carrying the weight, not of the combination that is served. Owner decision 2026-09-02, checklist B5: the estimate is NOT propagated through the two-route blend, this is documented rather than changed, and improvements ride a later release. What it means: on the action-angle inspiral every mode is served as c = w_D cD + w_R cR, an inverse-variance combination of the direct route and the PN-ratio route with one scalar weight per route per mode (aa_api._predict_pieces_in_arm), and what this method reports on the inspiral is the DIRECT route's envelope over the direct route's amplitude. Measured at q = 4, chi = (0.3, -0.2):

So on six modes the served content is the direct route to one percent or better and the band describes what is served; on the dominant mode the ratio route carries 15 percent and the band describes 85 percent of it.

Why it is not simply propagated, because a reader will ask: the two routes are NOT independent. Both are fitted from the same 260 simulations, the ratio being the same coorbital coefficient divided by its PN weight, so the textbook inverse-variance result vD vR / (vD + vR) understates, and it would NARROW a band already measured a factor 2.6 optimistic in sample (above). Three further reasons are recorded: propagating moves the four published odd-m numbers by 0.2 to 5.7 percent; the AdA arm has no route split, so an AA-only propagation would make the two arms' estimates differ in kind; and the weights would have to be recomputed outside the serving path, a second copy free to drift. The fully correlated limit (w_D sqrt(vD) + w_R sqrt(vR))^2 is a defensible upper bound and is the interim if one is ever wanted. The proper fix, for a later release, is not to propagate at all: measure the BLENDED content's own leave-one-out residual and calibrate one variance for it, exactly as the fourteen existing calibration factors were produced (commit aad8d7c0), which makes the correlation question disappear rather than answering it. It costs a leave-one-out campaign.

The envelope is DISCONTINUOUS at the fit's own element boundaries. Found 2026-09-02 while building the paper figure, and new: earlier work established continuity at the ARM JOIN, which is a different thing. Adjacent samples differ by up to a factor 4.11 on (2,2) at t = +2.17 M, beside the merger-ringdown element break at +2.0625 M, and every large jump sits within one grid spacing (0.25 M) of a break, while the served amplitude over the same region changes by at most 0.00977 in log10, which is only the ringdown decay. The cause is structural: the estimate is assembled per element and nothing constrains it to agree across a boundary. Consequences for a caller: differentiating the envelope, or reading a single point near a break, gives an answer that depends on which side of the boundary the sample landed.

The presentation hides the peak. docs/USAGE.md quotes the MEDIAN of these arrays. For (3,3) the median is 2.362e-05 and the maximum 4.756e-04, twenty times larger. The envelope-max column carries it, but the median reads as the answer.

Dynamics uncertainty, which is separate and opt-in

model.get_dynamics_error_estimate(mass_ratio, spin1z, spin2z)

reaches the E, J, k, tau and xpc variances, which are fitted and stored in the artifact and which nothing could reach before 2026-09-02. It is a separate method because the product is different: no mode index, different units, no merger-ringdown counterpart. It costs 4.9 ms against 109 ms for a waveform, and nothing if you do not call it.

Also true

predict_var warns that a point is outside the training hull at q = 8. The stored log-q bound is 2.0794264247 against log(8) = 2.0794415417, so the largest trained mass ratio is 7.999879 and the 8 in NRHJSur3dq8 is 1.51e-05 outside the model's own hull. Owner decision 2026-09-02: leave it and document it, which this paragraph and artifacts-and-resolution.md do; the three options considered are recorded with the model. model.in_hull is the authority, and q = 7.9998 is inside.

Where the code is