Research from locked tests.
Long-form guides for engineers searching real problems—mid-band ring, peak chase, delay failure, adaptive normals. Every figure traces to the Evidence page and Audit Workbench packs. Same philosophy as our other evidence sites: sources first, limits included, no paid miracles.
Identity — what the filter is
Start here if you need the dynamical story, not a scoreboard.
What GSRF really is (attractor stabilizer, not a tracker)
Locked identity: log-space soft-servo to x*, wrong-setpoint experiment, step settles ~55, k_return is the product.
Peak compression is not tracking
Why the step settles at ~55 after a jump to 65—and why that is the product, not a bug.
Frequency sweet spot & gain plateau ~0.37
Pure-sine characterization: EMA → 1.0 on slow waves; GSRF plateaus; osc wins at P≈50–120.
Performance — locked wins
Where the workbench said PASS under pre-declared rules.
GSRF vs EMA for mid-band oscillation
~+35% osc31 and ~−70% sweet-band spectral power—plus when osc31 still favors EMA.
Peak compression 49/52 vs MAD 0/52
Excellence hunt v2 scoreboard: the spring wins peaks and loses tracking by design.
Adaptive x* for permanent regime shifts
Frozen normal ~9.7 error; adaptive thermostat ~1.5—without throwing away the spring.
Failures — published red lights
If a vendor only shows wins, keep your wallet closed.
When GSRF lost
Delay, fixed-FA residual recall, tracking MAD—three locked losses with numbers.
Feedback delay: why balanced GSRF fails as a command path
+59% / +97% / +106% MAD-to-measured vs EMA at 3 / 6 / 10 min delays.
When not to use GSRF (the safety cage)
Green / red / note for engineers and managers in one page.
Applications — where it lives
Pre-AI sensor hygiene for industrial models
Calm mid-band ring and peaks before inference—without selling a false-alarm detector.
GSRF vs EMA vs Kalman
Job-by-job chooser from locked characterization—not a miracle ranking.
Quantum QEC / BP preprocessing path
Code-capacity d=5–13 under NDA—not industrial homepage evidence.
Latest long-tail guides
Additional claim-safe pages for search and agents (BRM-style coverage).
How to choose: GSRF vs EMA vs Kalman
Job-first decision table with locked reasons—not a single champion.
k_return: the spring that is the product
Mechanism posture: buy peaks/osc, pay MAD.
Fixed-FA residual detection loss
Recall ~0.44 vs EMA ~0.83 at matched FA.
SpO2-flat regimes (mixed, not locked)
Why we do not claim universal physiological wins.
Log-space filtering explained
Positive signals, exp(x), soft setpoint in log domain.
Industrial actuator batteries (bounded)
19 synthetic files: 7 clear / 7 partial / 5 mixed / 0 losses — not “wins actuators.”
Pre-model pipeline signal stabilizer
Where GSRF sits before industrial AI / PdM inference.
How claims get locked
Workbench packs, pre-declared rules, claim ladder, refused wording.
FAQ hub (schema-rich)
Short answers for crawlers and buyers; deep links to notes.
Questions people actually ask
Does GSRF always beat EMA?
No. Mid-band osc and peaks: often yes on locked packs. Tracking MAD, delay command, fixed-FA detection: no. See When GSRF lost.
Where do numbers come from?
Audit Workbench packs cited on Evidence. Notes do not invent new scores for SEO.
How do I try it?
/try — full Python snippet + static synthetic output table. Evaluation only.
Is there open-source production rights?
No. Public snippet is evaluation. Production needs a commercial license. Pricing.
Application mini-sites
Long-tail guides outside /notes/ (same claim ladder):