ohmic baseline mode: FUSE+IDA hybrid current with explicit Ip closure and a closure_channel toggle - #42
Open
d-burg wants to merge 20 commits into
Open
ohmic baseline mode: FUSE+IDA hybrid current with explicit Ip closure and a closure_channel toggle#42d-burg wants to merge 20 commits into
d-burg wants to merge 20 commits into
Conversation
…closed on j_ohmic only
…, not the cylindrical l_i proxy; proxy recorded alongside; refuse if FUSE total misses Ip by >2%
…linterp) on a copy_eq snapshot; OFT compute_flux_integral and the cylindrical proxy recorded only
… legacy-ravgs OFT builds; record fsa-convention Ip as cross-check
… total carries Ip (user-approved 2026-08-21)
…e term carried once (was triple-counted); 0.05% post-closure check
…ll Ip integrals on that snapshot (SWB-landed geometry read +31% of Ip on an ohmic-ramp slice)
…j_BS (j_ohmic untouched); the two channels bracket the closure uncertainty
…baseline_mode=ohmic)
…nt; ohmic-channel closure manufactures q_min inflation at high f_BS (see the q_min source comparison)
…cale_jBS as the per-draw spike -- draws were losing (1-bs)/bs of the pedestal bootstrap; zero error only at bs_scale=1
…rium roundtrip is resolution-INDEPENDENT at 1.305%)
…trips the equilibrium own Ip) instead of sign(dot(eq j_phi, cp j_tor)). The heuristic tested a current-direction convention, not the P-prime term sign: on ohmic ramp shots (eq j_phi > 0 while FUSE dd carries Ip < 0) it flipped and cost +1.31%25 of Ip. Ramp roundtrip 1.306%25 -> 0.002%25; flattop unchanged (bit-for-bit).
…iptions Public repo -- shot numbers stay out of code, docs and comments. The measured percentages are kept; only the case identifiers are generalised. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- tests/test_ohmic_closure.py: the affine FSA measure (c counted once, not per component -- the triple-count regression), exact cross-convention agreement on the equilibrium's own GS profile (the identity behind the 0.5% roundtrip gate), pprime_sign behaviour, both closure channels' scale formulas closing Ip exactly, validator acceptance, and defaults. - _validate_workflow now actually refuses generate() in ohmic mode (baseline-only until the draw path's sigma=0 reproduction is verified) -- the PR description promised this guard but it was never implemented; workflow='custom' downgrades it to a warning as usual. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…draw scale distribution's center
OFT applies scale_jBS INSIDE solve_with_bootstrap, so a 1.0 reference
makes the delta composition
bs*SWB0 + bs*SWB_pert - 1.0*SWB0
lose (1-bs)/bs of the pedestal bootstrap on every draw -- zero error only
at bs_scale=1, which is why the original validation passed.
The center is sigma0_reference_scale(jBS_scale_range): the midpoint of
the configured range (run.py has already re-centred it on bl.bs_scale),
or 1.0 when no range is set (draws then all run at scale 1.0). The
per-draw scale_jBS local CANNOT be used here -- it is bound inside the
draw loop, and referencing it before that raised an UnboundLocalError
that the cache block's blanket except swallowed into a silent
cache-disable fallback.
tests/test_delta_cache_reference.py pins the center semantics, the
helper wiring, and the scoping property itself (an AST load-before-store
check over generate_bouquet's locals that fails on the broken form).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…he gate reference, refuse NaN, extract close_ip, plug the jphi_diff leak Adversarial-review findings on the ohmic baseline mode: - P' sign: identically +1 by construction -- the affine c term is built from the ANCHOR's own P' on the anchor's own geometry, so the GS identity fixes its orientation and the data's current-direction convention enters only through the sign of the linear parts. The 'self-consistency probe' this replaces was a tautology (the probe was built with pprime_sign=+1 and reproduced the anchor's Ip to machine precision, so +1 could never lose), able to mask up to ~2c (~6% of Ip) on a genuinely flipped case with no gate able to catch it. - Roundtrip gate now references the anchor's ACHIEVED Ip, not Ip_target: a first pass converging 0.6% off target failed a perfect measure, and a 0.4% solver offset ate most of the 0.5% budget. Ip_anchor is recorded in ip_closure. - NaN can no longer pass the gates: the measure (weights, affine c, roundtrip) is checked finite by name, fsa_current_geometry now validates <1/R^2> and P' finiteness, and close_ip refuses NaN scales. - Closure algebra extracted to utils.close_ip (both channels, the ~0 divisor refusals, and the [0.2, 5] bounds); run.py and the tests now exercise the same shipped formulas instead of parallel re-derivations. - closure_channel is validated before solve_with_bootstrap (and in the workflow validator) instead of after the full SWB iteration sequence. - bl.jphi_diff is cleared after being recorded: generate() passes it through unconditionally, so a workflow='custom' ohmic generate folded it into every draw while the baseline solve omitted it. - sgn is taken from the linear total (the c term could flip the vote on a low-Ip ramp slice); recorded-only would-be scales bind each integrator once and record NaN instead of dividing by ~0. - Comment corrections: the reader's j_inductive is a RESIDUAL (j_tor - j_BS - j_NBI - j_RF), not to_toroidal(j_ohmic) -- verified on postdictive FUSE output, where the sawteeth source nets exactly zero current and is already folded into the diffused j_ohmic, while a near-axis j_non_inductive artifact (~0.4% of Ip at flattop) is the real unmodelled term the closure scale absorbs; the closure-channel default is 'bootstrap', not 'ohmic'. Also cherry-picked: the corrected issue-#44 sigma0 reference fix (sigma0_reference_scale + its tests) replacing the UnboundLocalError form. Fast suite: 414 passed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
d-burg
force-pushed
the
hybrid-ohmic-mode
branch
from
September 4, 2026 19:05
2f69012 to
54a5265
Compare
There was a problem hiding this comment.
🟡 Changes recommended
utils.close_ip has a concrete edge-case correctness bug (possible division by zero when Ip_target_signed is 0) and a bounds semantics mismatch that should be fixed before merging.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
This PR introduces an IMAS-path third baseline mode (jBS_baseline_mode="ohmic") that builds a hybrid current recipe (FUSE ohmic + SWB bootstrap + FUSE fixed components) and closes (I_p) via an explicit affine FSA-measure closure, plus a closure_channel toggle to attribute the deficit to either the bootstrap or ohmic channel.
Changes:
- Add
GenerationConfig.closure_channeland newjBS_baseline_mode="ohmic"workflow/validation behavior. - Implement
utils.close_ip(...)for shared closure algebra and add additional FSA geometry finiteness guards. - Add targeted tests for the affine FSA measure, closure algebra, workflow guard, and the sigma=0 delta-cache reference scaling.
File summaries
| File | Description |
|---|---|
bouquet/utils.py |
Adds close_ip helper and strengthens fsa_current_geometry finiteness checks. |
bouquet/run.py |
Implements the new "ohmic" baseline mode (anchor-geometry snapshot, closure logic, diagnostics) and extends workflow validation. |
bouquet/TokaMaker_interface.py |
Adds sigma0_reference_scale and uses it for the pre-loop SWB cache reference scale. |
bouquet/config.py |
Documents and adds GenerationConfig.closure_channel. |
bouquet/baseline.py |
Adds Baseline.ohm_scale and Baseline.ip_closure bookkeeping fields. |
tests/test_ohmic_closure.py |
New solve-free tests for affine measure, convention identity, closure algebra, and workflow guard behavior. |
tests/test_delta_cache_reference.py |
New tests pinning sigma=0 reference scaling semantics and guarding against the previous scoping regression. |
Review details
- Files reviewed: 7/7 changed files
- Comments generated: 6
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+739
to
+742
| _npsi_env = __import__("os").environ.get("BQ_FSA_NPSI") | ||
| _cap = _cef(mygs, | ||
| npsi=int(_npsi_env) if _npsi_env else max(257, psi_N.size), | ||
| psi_pad=psi_pad, exact_inv_R2=True) |
Comment on lines
+677
to
+678
| Ip_t = abs(float(Ip_target_signed)) | ||
| deficit = float(Ip_target_signed) - float(c_affine) - float(ip_fix) |
Comment on lines
+814
to
+815
| from .utils import (fsa_current_geometry, Ip_fsa_integral, | ||
| Ip_fsa_weights, eq_jphi_profile, close_ip) |
| "(expected 'ohmic' or 'bootstrap')") | ||
| lo, hi = scale_bounds | ||
| for name, s in zip(("ohm_scale", "bs_scale"), scales): | ||
| if not (lo < s < hi): |
Comment on lines
+41
to
+42
| src = inspect.getsource(tmi) | ||
| assert "_scale_ref = sigma0_reference_scale(jBS_scale_range)" in src |
Comment on lines
+19
to
+23
| import warnings | ||
|
|
||
| import numpy as np | ||
| import pytest | ||
| from scipy.integrate import trapezoid |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
jBS_baseline_mode="ohmic": FUSE+IDA hybrid current with explicit Ip closureAdds a third baseline mode for the IMAS path:
with Ip closed exactly on the LCFS-truncated FSA measure (
utils.Ip_fsa_integral,jphi-linterp convention, affine P′ term carried once), evaluated on a
copy_eq()snapshot of the anchor geometry taken before
solve_with_bootstrap(theSWB-landed geometry mis-integrates Ip by tens of percent in high-βp cases).
Closure channel toggle (
GenerationConfig.closure_channel)The hybrid components generally do not sum to Ip; the deficit has to land somewhere:
closure_channel="bootstrap"(default): keepj_inductiveexactly as FUSEdiffused it and rescale
j_BS—lin(ohm) + s_BS·lin(bs) + lin(fix) + c = Ip.Default because it preserves the source's core current: closure on
j_ohmichollows the core and inflates
q_minagainst the measured value whenever therecomputed bootstrap fraction is far above the source's.
closure_channel="ohmic": rescalej_inductiveonly, preserving its diffusedshape —
s·lin(ohm) + lin(bs) + lin(fix) + c = Ip. Useful as the other edge ofthe closure bracket.
The two are the extreme attributions of the same deficit; running both brackets the
closure uncertainty of the recipe. The applied scale (
ohm_scale/bs_scale) is afirst-class per-slice output, with the full integral decomposition recorded in
baseline.ip_closure(roundtrip check, FUSE-total error, per-component linear parts,and recorded-only alternative integrators).
P′-term sign: +1 by construction
The affine c term is built from the anchor's own P′ on the anchor's own
geometry, so the GS identity fixes its orientation; the external profile's
current-direction convention enters only through the sign of the linear parts.
Two failed detectors are retired: the original
sign(dot(eq j_phi, cp j_tor))heuristic (flipped on ohmic-ramp slices, +1.31% of Ip) and the
"self-consistency probe" that replaced it, which adversarial review showed to
be a tautology — the probe was built with
pprime_sign=+1and reproduced theanchor's Ip to machine precision, so +1 could never lose.
Hardening from adversarial review
(recorded as
ip_closure["Ip_anchor"]), notIp_target— a first passconverging 0.6% off target no longer fails a perfect measure.
checked finite by name,
fsa_current_geometryvalidates<1/R^2>/P′finiteness, and
close_iprefuses NaN scales.utils.close_ip(both channels, ~0-divisorrefusals, the [0.2, 5] bounds) — run.py and the tests exercise the same
shipped formulas.
closure_channelis validated beforesolve_with_bootstrap(a typo nolonger burns the full SWB iteration sequence).
bl.jphi_diffis recorded and then cleared in ohmic mode, so aworkflow='custom'generate can no longer fold the dropped anchor intodraws the baseline solve omitted.
j_inductiveis documented as what it is — a residualj_tor − j_BS − j_NBI − j_RF, notto_toroidal(j_ohmic); on postdictiveFUSE output the sawteeth source nets exactly zero current (already folded
into the diffused j_ohmic), while a near-axis
j_non_inductiveartifact(~0.4% of Ip at flattop) is the real unmodelled term the closure scale
absorbs.
Guards
within 0.5 %;
its σ=0 reproduction has been verified —
_validate_workflowraises ongenerate()(baseline-only), with the usualworkflow='custom'downgradeto a warning for the verification work itself.
Tests
tests/test_ohmic_closure.py(solve-free, synthetic FSA geometry): the affinemeasure identity (
I_p[J] = ∫wJ + c, withccounted once, not percomponent — the regression the linear-parts closure exists to prevent), exact
cross-convention agreement of the equilibrium's own GS profile between
jphi-linterpandfsa(the identity behind the 0.5 % roundtrip gate),pprime_signbehaviour, both closure channels' scale formulas closing Ipexactly, validator acceptance/refusal, and config/Baseline defaults.
Diagnostics
BQ_FSA_NPSIenv var: FSA quadrature-resolution knob (diagnostic; theramp-equilibrium roundtrip is resolution-independent).
BQ_CLOSURE_DUMP=<path.npz>: dumps the measure's ingredients so a roundtriperror can be localised in psi_N.
Not included / known limits
jphi_diffFUSE-equilibrium anchor is not applied in ohmic mode (recorded).the source reconstruction's (core hollowing, q0 rising against the measured value)
— that regime is exactly what the
bootstrapchannel + bracketing is for.closure_channel="mse") is split outinto a stacked follow-up PR (
feat/mse-closure-channel) so this PR stays areviewable unit.
🤖 Generated with Claude Code