The action taken
Switch to SSI-DATA. Notch the 24.9 Hz harmonic. Raise the model order. Re-window band two. Every move is named in plain language.
action ⋅ notch_harmonic
Deep Estimate — Automatic Mode Estimation
The world's first closed-loop Automatic Mode Estimation engine.†
It diagnoses its own results, switches methods, re-runs the identification, and reverts any step that degrades the evidence — logging every decision.
† Based on published capabilities of commercial OMA/EMA software and peer-reviewed literature as of July 2026. Substantiation dossier available on request.
01 — The loop
Automatic Mode Estimation that iterates
Automatic Mode Estimation that iterates
like an engineer: diagnose, adapt, re-run, defend.
One-shot automation clusters poles and hopes. Deep Estimate runs the analysis the way a senior analyst would: it reads the record, finds what is wrong, changes strategy, and runs the identification again — as many times as the evidence demands. It re-plans across the FDD family and the SSI sub-variants, calls FSDD and NExT-ERA in as bounded cross-validators, and targets remediation at the specific frequency band that misbehaves.
STEP 1PerceiveCharacterize the record: stationarity, SNR, spectral content, harmonic risk.
STEP 2DiagnoseDetect rotating-machinery harmonics, close modes, weak excitation, aliasing exposure.
STEP 3PlanChoose the remediation: switch SSI variant, adjust model order, re-window, notch.
STEP 4Re-identifyExecute the plan and re-run identification with the adapted strategy.
STEP 5Cross-validateConfirm across independent methods; commit only the modes that survive the gates.
The field's latest state-of-the-art survey (Machines, 2025 — 120 papers) lists automated cross-method arbitration as an open research goal. Modal Master ships it.
02 — What "closed-loop" means here
A loop that can reject
A loop that can reject
its own bad ideas.
Iteration alone is not intelligence. What closes the loop is the regression guard: after every corrective action, Deep Estimate re-scores the result across a multi-axis evidence vector — and if the action made the evidence strictly worse, it reverts the action and records why. The revert itself becomes part of the log.
THE REGRESSION GUARD
Every action must earn its keep.
Harmonic notching, channel exclusion, projection reduction, model-order changes, method switches — each is applied as a guarded experiment on an isolated scratch session. An action that degrades the committed evidence across the guard's eleven evaluation axes is rolled back before it can contaminate the result. Your working session is untouched until the final mode set commits.
- GuardedIn-loop actions are scored against the evidence vector; strictly-worse outcomes revert.
- IsolatedThe whole loop runs on a scratch session — the hard boundary that protects your data.
- RecordedReverted actions stay in the decision log with their revert reason, visible at review.
CROSS-METHOD ARBITRATION
Two families, one verdict.
The primary route to a confident commit demands agreement between two independent method classes — frequency-domain and subspace — correlated by cross-method MAC. Where the data can only support a single family, tightly guarded single-family routes exist by design, and the evidence trail says so explicitly: no mode pretends to more corroboration than it has.
- Per bandRemediation carries band bounds — the fix lands on the octave that needs it.
- Data branchesHarmonic-notched, channel-zeroed, and projection-reduced branches run as first-class guarded actions.
- ArbitratedCandidates from every engine converge through one literature-cited commit gate.
03 — The decision log
Every iteration answers:
Every iteration answers:
what, why, how — and what happened.
The decision log is a product surface, not a debug file. Each iteration records what the engine did, why it chose that action, how it executed, and what the evidence said afterwards — including the actions it tried and took back. Each committed mode carries its own evidence bundle with provenance hashes.
The diagnostic that fired
Actions trace back to a diagnostic finding — harmonic contamination, weak excitation, close-mode interference — not to a heuristic whim.
finding ⋅ spectral kurtosisThe evidence verdict
After every action, the guard's re-score: kept because the evidence improved — or reverted, with the reason printed in the log.
reverted ⋅ reason loggedAn audit for each commit
Which gates the mode passed, the measured value against each threshold, and the literature citation that defines the bar.
MAC 0.981 ≥ 0.90Hashes on the bundle
Per-mode evidence bundles carry provenance hashes, so a reviewer can verify that the evidence on screen is the evidence that was computed.
bundle ⋅ hashedBuilt to be interrogated
The log reads top to bottom as an engineering narrative — the same story you would want from the analyst at the desk next to yours.
timeline ⋅ legibleSame record, same settings — same modes, same decisions, same log. Deep Estimate is deterministic by construction, because a result you cannot reproduce is a result you cannot defend.
— Reproducibility is a requirement, not a feature
04 — The certification discipline
Certified against a benchmark
Certified against a benchmark
built to make it fail.
Most automation is demonstrated on friendly data. Deep Estimate is certified against a 45-dataset benchmark designed to be hostile: close-mode pairs, harmonic contamination, SNR ladders down to the noise floor, nonstationary excitation, interior spectral gaps — each dataset generated with exact ground truth, so there is nowhere to hide.
Known truthEvery benchmark record is synthesized with exact modal parameters by the Modal Master Signal Generator — the engine is graded against truth, not opinion.
Hard gatesA pass demands the expected modes committed within tight frequency tolerance, no undeclared extra modes, and full candidate and run provenance.
Byte-stableCertification requires two independent runs to reproduce the committed mode set byte-for-byte; the committed signature is SHA-256-pinned.
Hostile by designThe suite grew by adding datasets that broke the engine — every defect class found became a permanent regression tripwire.
No lucky runsA result that cannot be reproduced identically is treated as a failure. Nondeterministic passes are worthless and are scored as such.
Re-certifiedAny change to the engine re-runs the full battery. The baseline signature moves only when the evidence says the change is sound.
05 — Beyond synthetic data
The same discipline,
The same discipline,
on real structures.
Synthetic certification proves control; real records prove relevance. The platform's evidence tooling faces recorded ambient data from instrumented structures — a cable-stayed footbridge, a centuries-old mosque, a five-story reinforced-concrete apartment block — under the same rules: archived protocols, hashed inputs, and byte-identical reproduction across independent runs.
THE REPLICATION PROTOCOL
Re-run, hash, compare. Twice.
A real-data claim here means an archived protocol re-run in full fidelity: every input record hashed, every configuration field reproduced, and the result required to be byte-identical across isolated fresh-process runs. The discipline is not reserved for the flagship: the research-badged TOMA evidence engine shipped only after clearing its own three-structure gate under this protocol, in July 2026.
A result that cannot be replicated under the archived protocol does not ship. No lucky runs, no it-worked-once screenshots.
THE GHOST THAT DIDN'T SURVIVE
The discipline catches what matters.
On the apartment-block record, an earlier engine build had committed two spectral ghosts — plausible-looking modes that were artifacts, not structure. The evidence discipline flagged them, the root cause was traced in the identification chain, and on the current engine the committed set contains exactly the three true modes. The re-run proves it: same record, same protocol, ghost-free — byte-for-byte, twice.
This is what defensibility means in practice: the system is built so that a wrong mode
cannot hide behind a confident interface.
06 — Damping, cross-examined
Four independent estimators.
Four independent estimators.
The baseline never votes on itself.
Damping is the most contested number in structural dynamics. Deep Estimate cross-validates every committed damping ratio with up to four automatic estimators — half-power bandwidth, CFDD curve fit, Hilbert envelope decay, and continuous wavelet transform — spanning two bias families. The estimator that produced the committed value is deliberately excluded from its own cross-check: agreement means something precisely because the judges are independent. Disagreement is not averaged away; it is shown, and it lowers the quality tier. The Hilbert and wavelet judges fit a band-passed free-decay proxy of the record — not raw ambient ridges — and a fit that fails its acceptance gates is reported as not identified, with the reason, never as a silent zero.
Half-powerBandwidth of the spectral bell around the committed frequency.
CFDD fitParametric SDOF curve fit in the frequency domain.
HilbertEnvelope decay of the band-isolated correlation function.
CWTContinuous wavelet ridge decay — a separate bias family entirely.
07 — Scope, stated plainly
What Deep Estimate does not do.
A defensibility product owes you its boundaries. These are Deep Estimate's, in writing.
Evidence tools
TOMA (PSDTM-SVD) and the other research-grade evidence tools inform review — they never commit modes. Committed results come from the FDD and SSI identification families through the commit gate.
Spectral settings
The loop adapts identification strategy — method, model order, banding, data branches. It does not silently re-tune your spectral-estimation parameters or re-segment your record.
Confidence routes
The primary confident route requires two independent method classes. Guarded single-family routes exist for data that supports only one family — and the evidence trail labels them as such.
Guard boundary
The regression guard governs in-loop actions on the isolated scratch session. It is a guard on evidence quality during the loop — not a claim that every conceivable outcome is superior to every alternative.
See it on your data
Watch it diagnose, adapt,
and defend the result.
Bring an ambient record you know well. Deep Estimate will identify it, justify every commit, and hand you the decision log — the same trail your reviewer will read.
Enterprise evaluations · Technical dossiers under NDA · Volume licensing