Stability and Reproducibility (v3.5.5)

CamillaFIR is designed to avoid the classic failure mode:

small measurement or setting changes -> large unstable filter changes

The system stays predictable because it is deterministic, bounded, and observable.

1) Deterministic pipeline and versioned behavior

Given the same:

  • CamillaFIR version
  • input files
  • settings
  • base fs/taps or comparison-grid settings

CamillaFIR produces repeatable outputs.

The pipeline keeps its behavior deterministic for fixed inputs and settings. Automatic mode may use seeded Optuna-backed search, but the seed is derived deterministically from the measurement/signature context so identical runs remain reproducible. Release v3.5.5 (2026-03-07) keeps this deterministic model while adding smarter cache reuse around automatic-mode target and preset search.

2) Input robustness: TXT and WAV/IR

Modern workflows mix REW text exports and WAV/IR measurements.

Stability improvements include:

  • deterministic WAV parsing policy aligned to TXT baseline behavior
  • path sanitation and local WAV existence validation
  • higher minimum FFT robustness for WAV IR -> FR conversion
  • alignment guard between delay estimate and impulse-peak alignment

This reduces source-format-dependent surprises.

3) Bounded correction prevents inverse-filter runaway

CamillaFIR keeps each stage bounded:

Magnitude bounds:

  • max_boost_db, max_cut_db
  • low-bass safety controls
  • smoothing and regularization (filter_smooth, reg_strength)

Shape bounds:

  • slope limits (max_slope_db_per_oct)
  • optional asymmetric slope limits for boost vs cut

Decay bounds:

  • TDC strength and max reduction limits
  • optional TDC slope limit

When limits are active, behavior remains controlled instead of extreme.

4) Confidence-guided correction improves repeatability

Room data confidence is frequency-dependent. CamillaFIR uses confidence-aware behavior to avoid overreacting to artifacts:

  • Confidence Pull logic
  • A-FDW support
  • smoothing that favors robust trends over narrow features

Practical effect: small mic-position changes usually alter details, not the whole correction profile.

5) Phase-domain safeguards

Phase processing is constrained by design:

  • phase_limit bounds correction bandwidth
  • Mixed-phase excess correction is faded from LF to HF
  • Mixed-only limits cap excess delay and pre-ringing risk
  • adaptive excess-phase clamping and conditional GD spike guards improve robustness

These controls prevent unstable phase behavior and reduce ringing risk.

6) Headroom reproducibility and stereo consistency

Level handling is deterministic and tied to realized correction:

  • auto-headroom model computes attenuation from realized max boost + margin
  • shared auto-gain behavior is available for stereo-linked runs
  • plot-only compensation keeps analysis visuals readable while preserving output safety

This keeps output gain decisions consistent across repeated runs.

7) Comparable evaluation across fs/taps

Changing fs/taps can change analysis grids even when audible behavior is similar. CamillaFIR addresses this with:

  • comparison mode (fixed reference grid for scoring/match)
  • multi-rate export with time-length-aware tap mapping

So you can compare settings more fairly without grid-induced score drift.

8) Observability and audit trail

Reproducibility needs traceability. CamillaFIR provides:

  • Summary version stamp and key effective parameters
  • run timing breakdown (Read, DSP, ZIP/PNG, Render, Total)
  • expanded System Health warnings for source completeness and risky settings

If two runs differ, diagnostics help explain why.

Reproducible-run checklist

  1. Use the same app version (v3.5.5 or a pinned newer release).
  2. Use identical measurement sources and source type (TXT vs WAV) per comparison.
  3. Keep base fs/taps fixed, or enable comparison mode.
  4. Keep correction guardrails fixed (boost/cut/slope/regularization/phase limit).
  5. Keep phase strategy fixed.
  6. Keep headroom margin and stereo-link policy fixed.
  7. Verify System Health is clean before exporting.
  8. Compare Summary outputs and timing data when validating repeated runs.

Bottom line: CamillaFIR is reproducible because correction strength is bounded, phase is safety-limited, and the full run is observable end-to-end.

Disclaimer

AI was used to translate this document from Finnish to English.