---
name: analyze-feedback-stability
description: Analyze feedback-loop stability with backend-independent OpenADA evidence. Use for differential, common-mode, nested, feed-forward, regulator, or switched-loop questions involving operating-point validity, loop injection and probe identity, gain and phase crossover review, phase or gain margin, closed-loop correlation, oscillation diagnosis, or choosing one focused compensation experiment without relying on simulator-specific commands.
---

# Analyze Feedback Stability

Produce an evidence-bounded diagnosis, not a universal compensation recipe.
Use `$openada:openada` to invoke semantic operations and interpret results. Keep the
topology, prepared testbench, PDK/models, and native artifacts authoritative.

## Separate the claims

Use these exact layers:

- Establish analysis evidence with
  `openada.operation/circuit.simulate/v1alpha2` and
  `openada.assertion/simulation.evidence.valid/v1alpha1`.
- Bind the exact native AC Cartesian vectors through
  `openada.operation/result.series.extract/v1alpha1` and
  `openada.assertion/series.extraction.valid/v1alpha1`.
- Derive the implemented loop metrics only through
  `openada.operation/result.transfer.measure/v1alpha2` and
  `openada.assertion/transfer.measurement.valid/v1alpha1`.
- Derive ordinary time-domain or real-series scalars only through
  `openada.operation/result.measure/v1alpha2` and
  `openada.assertion/measurement.valid/v1alpha1`.
- Judge an explicit stability limit only through
  `openada.operation/specification.evaluate/v1alpha1` and
  `openada.assertion/specification.satisfied/v1alpha1`.

Inspect capability records, then run `openada profile show
openada.operation/result.transfer.measure/v1alpha2` before constructing a
request. Do not invent a metric kind or fall back to a backend expression when
the requested `result.transfer.measure` metric kind is not supported. Mark that
metric **not evaluated**.
None of these layers alone establishes silicon, reliability, or signoff.

## Freeze one baseline

Record:

- authoritative design/testbench identity and design phase;
- supply, bias, common mode, differential stimulus, load, corner, temperature,
  and operating mode;
- enabled paths, stages, switches, feedback elements, and intended loop state;
- exact loop break or injection point, direction/sign convention, return
  probe, loop-response signals, and requested closed-loop probes;
- selected operation/feature, driver, native product/version, request/result
  IDs, and evidence destination.

Classify differential, common-mode, nested, local, global, and feed-forward
paths from authoritative connectivity evidence. Signal names and a remembered
topology are not proof. If topology, polarity, enabled state, injection, or
probe orientation remains ambiguous, ask one narrow question and stop before
quantitative margin claims.

The simulation operation consumes a prepared testbench; it does not create a
safe loop break or mutate the design. Require a reviewed testbench or a separate
authorized design/testbench-construction task. Record how the injection source
preserves the relevant DC state and loading.

## Pass the DC gate

1. Require capability
   `openada.feature/simulation.analysis.op/v1alpha1` and request nominal OP
   evidence under the frozen conditions.
2. Treat analysis-evidence pass only as proof of valid OP evidence.
3. Use `result.measure` for the topology-specific common-mode, bias, rail,
   headroom, saturation, or device-operating facts needed by the review.
4. Evaluate explicit acceptable ranges through the specification contract when
   they exist. Otherwise label the engineering acceptance as an identified
   reviewer assumption, not a specification result.

Stop quantitative stability interpretation when OP evidence fails or is
unknown, when required DC `result.measure` results are invalid/unavailable, or
when the simulated state is inconsistent with the intended loop mode. Report
that gate; do not explain railing, saturation, or a floating state as
phase-margin failure.

## Acquire complete loop evidence

Require capability `openada.feature/simulation.analysis.ac/v1alpha1` for the
declared frequency range. Bind the request to the same baseline and retain the
exact generated simulator input and native result artifacts.

After analysis evidence passes, request supported `result.transfer.measure`
metrics that preserve:

- loop-response signal definition, orientation, and sign convention;
- full magnitude and unwrapped phase over the declared frequency range;
- every falling 0 dB crossing candidate used to decide whether the result is
  unique or ambiguous, including interpolation method;
- phase and phase margin at the unique falling unity-gain crossing under the
  stated convention;
- ambiguous, missing, or multiple crossings rather than selecting the most
  favorable scalar.

The implemented transfer profile constructs `H = output / input` from four
explicit same-unit real series on one positive-Hz axis. It retains magnitude
and deterministically unwrapped phase, calls the first point
`first-simulated-frequency-not-dc`, and supports:

- `low_frequency_gain_db` at that first positive frequency;
- `bandwidth_3db` at the unique falling first-gain-minus-3.0-dB crossing;
- `unity_gain_frequency` at the unique falling 0 dB crossing; and
- `phase_margin` as 180 degrees plus unwrapped phase at that crossing, only
  when the request explicitly declares `loop-gain-negative-feedback`.

Crossings use linear response interpolation over log10 frequency. An absent
crossing is typed `not_found`; multiple falling crossings are `unknown`. Gain
margin, phase-crossing search, favorable-crossing selection, smoothing,
fitting, and a stability claim remain unavailable. Preserve the valid AC
artifact and mark those metrics not evaluated. A plot may motivate a
hypothesis, but it is not a `result.transfer.measure` result.

## Correlate the closed loop

Use a separate AC or transient request, with the matching advertised feature,
for the exact closed-loop input/output and common-mode probes. Current profiles
can derive declared settling behavior or the first-simulated-frequency AC
transfer gain from evidence bound to the same baseline. Overshoot and ringing
metrics remain unavailable; do not infer them from a plot or rename an
`result.measure` extremum result.

Do not answer a closed-loop question from an open-loop signal. Do not declare a
physical instability until the evidence distinguishes it from:

- unintended switch or operating mode;
- floating nodes, saturation, or invalid bias;
- convergence or numerical artifacts;
- model/configuration changes or simulator-version effects;
- incorrect injection, probe orientation, or `result.transfer.measure`
  convention.

## Evaluate specifications and one hypothesis

Evaluate phase margin, crossover range, settling, or other implemented limits
only when the user/project supplied explicit bounds, units, and conditions.
Keep gain margin not evaluated in v1alpha1.
Keep one result per loop and condition; one passing crossover does not cover a
multi-loop system.

When evidence supports a diagnosis, propose one change tied to one hypothesis.
Run it in a separately identified baseline, preserve the design/testbench diff,
and repeat the same OP, loop, and closed-loop evidence. Do not change topology,
compensation, or authoritative design data without authorization.

## Route status

| State | Interpretation |
|---|---|
| Analysis evidence pass | The requested native analysis is trustworthy enough for supported result operations |
| Analysis evidence fail | The solver's defined terminal failure is proven; no margin conclusion follows |
| Analysis evidence unknown | Resolve evidence, binding, configuration, or execution uncertainty |
| `result.transfer.measure` pass | The stated metric is valid for its exact input/output phasors, interpretation, and method |
| `result.transfer.measure` fail/unknown/unavailable | Margin or response is not evaluated; repair or add the semantic primitive |
| Specification pass/fail | Only the explicit limit at the frozen condition was evaluated |

## Report

Return the frozen baseline, loop classification, DC gate, injection and probes,
all simulated crossovers, closed-loop correlation, contract status at each
layer, exact driver/artifact lineage, confounders, and one smallest next
experiment. Mark statements as **simulated**, **inferred**, or **proposed**.
Finish with `signoff: not claimed`.