perturbation-sim B1: finite-gate the spectral-gap NaN landmine (Davis–Kahan / λ₂→0)#524
Conversation
|
Warning Review limit reached
More reviews will be available in 38 minutes and 7 seconds. Learn how PR review limits work. Your organization has used up its prepaid credits, and credit purchases are no longer available. Enable the review add-on in the billing tab to keep reviews running — you're only billed for reviews past your plan's rate limits ($0.25/file). ⌛ How to resolve this issue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based credits. 🚦 How do rate limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan refill rate. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, the refill rate gradually slows as usage increases. The highest same-day bursts are limited more strictly. Please see our Fair Usage Limits Policy for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (5)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
The spectral half of perturbation-sim divides by a vanishing spectral gap precisely at blackout onset: the Davis-Kahan Fiedler rotation bound sinθ ≤ ‖E‖₂/gap and the mode-instability modifier Δλ·(1/λ₂) both blow up as gap→0 / λ₂→0 — the exact fragmentation regime the model exists to capture. Same NaN class as ndarray entropy_ladder::residue_surprise (PR #221): a subtracted denominator collapsing to 0 → 0/0, or a tiny-positive denominator slipping past a naked `gap > ε` test and producing a silently-wrong finite "bound". Fix (the entropy_ladder pattern): floor the denominator AFTER the subtract-and-min that builds it, relative to λ_max, then divide unconditionally. gap ≤ SPECTRAL_GAP_FLOOR·λ_max ⇒ spectral degeneracy ⇒ FRAGMENTATION_SENTINEL (= +∞, the workspace's documented divergence signal — finiteness-checkable, never NaN). connectivity_loss gated by the floor + clamped to [0,1]; weyl_over_fiedler floor unified, divides by λ₂.abs(), numerator .max(0.0). Sentinel, not swallow: gap=0 is a real result (the Fiedler vector may rotate arbitrarily ⇒ sinθ ≤ ‖E‖₂/0 = +∞). The gate guarantees every spectral output is either a trustworthy finite number or the explicit +∞ divergence signal — never NaN, never a silently-wrong finite. A symmetric ring (degenerate Fiedler mode, λ₂=λ₃) confirmed the sentinel is reachable in normal use, not just a corner case. +4 regression tests (91 → 95 green): NaN-freedom across the precursor regime (near-zero bridge, disconnected graph, degenerate ring, λ₂∈ {0,±1e-14,…}) + sentinel-correctness on fragmentation + normal-regime PARITY (asymmetric weighted path, separated λ₂, non-vacuity asserted) proving the gate touches ONLY the singular path. B2 (witness arc → contract witness_table evaluator): assessed, STOP-and-reported. The contract WitnessTable is a (mailbox_ref, spo_fact_ref) address-resolution primitive, NOT an arc evaluator — no hook exists; wiring would require a NEW contract trait (WitnessArcEvaluator) + a field-carrying type, which is more than additive. Did not force it (witness.rs's own doc names it a separate gated step). Exact surface documented in AGENT_LOG for operator decision. Behaviour-preserving in the normal regime; gated only the degenerate path. clippy -D warnings clean, fmt clean, examples build. Board: EPIPHANIES E-SPECTRAL-GAP-NAN-1 + AGENT_LOG prepended in this commit. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CcpLeEC3XK8Eye53GKBVvi
AdaWorldAPI
force-pushed
the
claude/perturbation-soa-nan-gate-witness
branch
from
398e3d4 to
69a1189
Compare
2 participants
What
Finite-gates the spectral-gap NaN landmine in the
perturbation-simoutage model — "one NaN less" on the SoA substrate, applied to the regime the model exists to capture. Additive, behaviour-preserving in the healthy regime; gates only the degenerate (fragmentation / λ₂→0) path.B1 — spectral NaN audit + finite-gate (shipped)
Audited every division by a spectral gap / λ₂ / eigenvalue-difference / norm across all spectral modules. The crate was already well-defended (
eigen'spseudo_apply/pseudo_inverseuse a relativerel_tol·λ_maxcutoff;lodf/kirchhoff_index/stats/timing/buffer/chaoda/modelare guarded). The real defect was in the Davis–Kahan path: naked, absolute, inconsistent ε-guards that caught0/0but let a tiny-positive noisygapslip through as a silently-wrong finite "bound."Sites gated (floor the denominator after the subtract, relative to
λ_max— theentropy_ladder::residue_surprisepattern, PR #221):perturbation.rsdavis_kahan_boundgapfloored after theminatSPECTRAL_GAP_FLOOR·λ_max;≤ floor ⇒ FRAGMENTATION_SENTINELgap→0= fragmentation ⇒sinθ ≤ ‖E‖/0 = +∞; a finite, finiteness-checkable sentinel that preserves the signal, never NaNperturbation.rsconnectivity_loss.clamp(0,1)0/0= "no connectivity to lose" ⇒ 0; clamp stops a noisyλ₂exceeding the fraction's domainrolling_floor.rsweyl_over_fiedlerSPECTRAL_GAP_FLOOR;/ λ₂.abs().abs()+.max(0)keep the instability ratio non-negativeTests: 91 → 95 (manifest-path;
perturbation-simis an excluded crate). New:blackout_precursor_regime_is_never_nan,disconnected_graph_spectral_outputs_are_nan_free,weyl_over_fiedler_is_nan_free_at_the_precursor(finiteness PASS), andhealthy_grid_davis_kahan_is_finite_and_bounds_rotation(normal-regime parity PASS, non-vacuous). Key discovery: a symmetric ring's Fiedler mode is legitimately doubly-degenerate (λ₂=λ₃ ⇒ gap=0), so the sentinel is reachable in normal use — confirming the gate is signal-preserving, not a corner-case patch. clippy-D warningsclean;fmtclean; examples build.B2 — witness-arc → contract
witness_table: deferred (needs an operator decision)The contract
WitnessTableis an address-resolution primitive (W-slotget/set), not an arc evaluator — there is no additive hook to wirewitness.rs'sparticle == waveParseval oracle into. Doing so requires a new contract trait (WitnessArcEvaluator) on a new field-carrying type — more than additive, crossing the "no new trait/bridge" iron rule (andwitness.rs's own doc calls it "a separate, gated step"). Not forced; the exact surface is documented inAGENT_LOG.mdfor the operator's call.Adjacent finding (flagged, out of scope)
acflow.rs(AC Newton–Raphson) has its own division landmines —Cx::recip ÷ |z|²(zero-impedance line) and Jacobian÷ v[i](NR divergence →v→0). These are voltage-collapse, not spectral-fragmentation, and want a different gate (pivot guard +converged:false). Recorded inE-SPECTRAL-GAP-NAN-1for a future AC-flow pass — not touched here.Board:
EPIPHANIES.mdE-SPECTRAL-GAP-NAN-1+AGENT_LOG.md, same commit.🤖 Generated with Claude Code
Generated by Claude Code