diff --git a/README.md b/README.md
index 3566ca2d..c805f32e 100644
--- a/README.md
+++ b/README.md
@@ -632,6 +632,8 @@ Verify the plugin structure: `bash plugins/ruview/scripts/smoke.sh`. Full detail
 |----------|-------------|
 | [User Guide](docs/user-guide.md) | Step-by-step guide: installation, first run, API usage, hardware setup, training |
 | [Build Guide](docs/build-guide.md) | Building from source (Rust and Python) |
+| [Calibration & Room Training Guide](docs/calibration-guide.md) | What `calibrate`/`enroll`/`train-room` actually enforce: minimum frame counts, per-anchor quality gates, the pet/small-motion presence-detection caveat, and empty-room baseline conditions — grounded in the real code, not just ADR-135/151 |
+| [Trust State & Engine Errors](docs/trust-and-engine-errors.md) | What `engine_error_count` and `demoted` mean on `/api/v1/status`, exact trigger conditions, the current diagnostic gap (no per-cause breakdown), the `WDP_GUARD_INTERVAL_US` recovery path, and why a converted Hugging Face model isn't shown to be the cause in code |
 | [**Home Assistant + Matter Integration**](docs/integrations/home-assistant.md) | **Works with Home Assistant** via MQTT auto-discovery + **Works with Matter** (Apple Home / Google Home / Alexa / SmartThings) — full entity catalog, 3 starter blueprints, Lovelace dashboards, privacy mode, threshold tuning ([ADR-115](docs/adr/ADR-115-home-assistant-integration.md)). |
 | [**BFLD — Beamforming Feedback Layer for Detection**](v2/crates/wifi-densepose-bfld/README.md) | New privacy-gated WiFi sensing layer that measures + structurally prevents identity leakage from 802.11ac/ax Beamforming Feedback Information. Three type-enforced invariants (raw BFI never exits node, identity embedding is in-RAM-only, cross-site correlation cryptographically impossible via per-site BLAKE3 keyed hash + daily rotation). Ships full operator surface (`BfldPipeline`, `BfldPipelineHandle`, the Soul Signature §3.6 per-channel matcher `EnrolledMatcher`/`SoulMatchOracle` — experimental; named identity is data-gated, **measured** as not-separable on WiFi-only channels alone), MQTT topic router + HA-DISCO + availability + LWT, 3 operator HA blueprints, two runnable examples, eclipse-mosquitto:2 CI service container. 327+ tests. [ADR-118](docs/adr/ADR-118-bfld-beamforming-feedback-layer-for-detection.md) umbrella + sub-ADRs [119](docs/adr/ADR-119-bfld-frame-format-and-wire-protocol.md)/[120](docs/adr/ADR-120-bfld-privacy-class-and-hash-rotation.md)/[121](docs/adr/ADR-121-bfld-identity-risk-scoring.md)/[122](docs/adr/ADR-122-bfld-ruview-ha-matter-exposure.md)/[123](docs/adr/ADR-123-bfld-capture-path-nexmon-and-esp32.md). Research dossier: [`docs/research/BFLD/`](docs/research/BFLD/) (11 files, 13,544 words). |
 | [**SENSE-BRIDGE — rvagent MCP server**](tools/ruview-mcp/README.md) | Dual-transport MCP server (`@ruvnet/rvagent`) bridging the RuView sensing stack to AI agents (Claude Code, Cursor, ruflo swarms). 6 tools wired: `ruview.presence.now`, `ruview.vitals.get_{breathing,heart_rate,all}`, `ruview.bfld.last_scan`, `ruview.bfld.subscribe`. stdio + Streamable HTTP (`POST /mcp`, Origin-validated, bearer-token auth, `127.0.0.1` bind). Full 20-tool Zod schema barrel + 5 RUVIEW-POLICY governance tools. 93 tests. [ADR-124](docs/adr/ADR-124-rvagent-mcp-ruvector-npm-integration.md). Try: `npx @ruvnet/rvagent stdio`. |
diff --git a/docs/calibration-guide.md b/docs/calibration-guide.md
new file mode 100644
index 00000000..dd174cc8
--- /dev/null
+++ b/docs/calibration-guide.md
@@ -0,0 +1,300 @@
+# Calibration & Room Training Guide
+
+This guide explains what actually happens — and what is actually *enforced* —
+when you run `wifi-densepose calibrate`, `enroll`, and `train-room`. It is
+written for the person setting up a room, not for developers.
+
+Everything below was checked against the real Rust implementation in
+`v2/crates/wifi-densepose-calibration/`, `v2/crates/wifi-densepose-signal/src/ruvsense/calibration.rs`,
+and `v2/crates/wifi-densepose-cli/`, not just the design ADRs. Where the design
+documents (ADR-135, ADR-151) describe something that isn't actually built yet,
+this guide says so explicitly.
+
+## The three-step pipeline
+
+```
+wifi-densepose calibrate --port <PORT>      # Stage 1: empty-room baseline (no people)
+wifi-densepose enroll --room <NAME>         # Stage 2+3: 8 guided anchors (~4 minutes)
+wifi-densepose train-room --room <NAME>     # Stage 4: fit the specialist bank
+wifi-densepose room-status --room <NAME>    # check what trained / what's stale
+wifi-densepose room-watch --room <NAME>     # live inference
+```
+
+`calibrate` must run first — `enroll` refuses to start without a baseline file
+(`--baseline ./baseline.bin` by default), and `train-room` refuses to start
+without an enrollment file. Each step writes a file the next step reads; there
+is no way to skip a step.
+
+---
+
+## 1. Is there a minimum amount of data required?
+
+**Yes, and for the empty-room baseline it is a hard, enforced minimum — not a
+recommendation.**
+
+`wifi-densepose calibrate` will not produce a baseline file with fewer than
+**600 recorded frames** (the default for every PHY tier: HT20, HT40, HE20,
+HE40). This is `DEFAULT_MIN_FRAMES = 600` in
+`v2/crates/wifi-densepose-signal/src/ruvsense/calibration.rs:48`, and it is
+checked in `CalibrationRecorder::finalize()`
+(`ruvsense/calibration.rs:532-538`): if fewer than `config.min_frames` frames
+were recorded, `finalize()` returns
+`CalibrationError::InsufficientFrames { got, need }` and calibration fails
+outright — there is no partial/degraded baseline. This is pinned by a unit
+test (`finalize_requires_min_frames`, same file) so it isn't accidental
+behavior.
+
+**Important subtlety:** the CLI's `--duration-s` flag (default 30 seconds)
+and the 600-frame minimum are checked *independently*. The capture loop in
+`v2/crates/wifi-densepose-cli/src/calibrate.rs:135-183` stops as soon as
+**either** the duration timer expires **or** 600 frames have been recorded,
+whichever comes first. If your node streams CSI slower than the assumed 20 Hz
+(e.g. congested WiFi, a busier ESP32), 30 seconds may not be enough to reach
+600 frames, and `calibrate` will fail with an explicit
+`"insufficient frames: have X, need 600"` error rather than silently
+producing a short baseline. If you hit this, raise `--duration-s` rather than
+overriding `--min-frames`.
+
+You *can* override the 600-frame floor with `--min-frames <N>` (0 = use the
+tier default). The code prints an explicit warning when you do:
+
+> `[calibrate] WARN: --min-frames=N overrides ADR-135 tier default (600 for
+> ht20). This relaxes the phase-concentration guarantee; do not use in
+> production.`
+
+(`v2/crates/wifi-densepose-cli/src/calibrate.rs:112-119`). Treat this as a
+debugging escape hatch, not a supported way to shorten setup.
+
+The CLI also independently rejects `--duration-s` below 10 seconds
+(`"Fewer frames produce unreliable phase-concentration estimates"`,
+`calibrate.rs:341-348`) and prints (but does not block on) a warning above 300
+seconds.
+
+**Guided enrollment (`enroll`) has a much lower, per-anchor floor.** Each of
+the 8 guided anchors (`empty`, `stand_still`, `sit`, `lie_down`,
+`breathe_slow`, `breathe_normal`, `small_move`, `sleep_posture`) is captured
+for a fixed duration baked into the code — 20 seconds for the static/motion
+anchors, 30 seconds for the two breathing anchors and `sleep_posture`
+(`AnchorLabel::duration_s()`, `v2/crates/wifi-densepose-calibration/src/anchor.rs:98-104`).
+This is **not** a CLI flag — you cannot currently shorten or lengthen an
+individual anchor capture from the command line.
+
+Underneath that fixed duration, the anchor is only *accepted* if it clears a
+quality gate (`AnchorQualityGate`, `v2/crates/wifi-densepose-calibration/src/enrollment.rs:43-53`):
+
+| Threshold | Default | What it checks |
+|---|---|---|
+| `min_frames` | **60 frames** | Anchor is rejected if fewer than 60 frames were captured — mainly catches "the ESP32 stopped streaming" mid-capture, not a real duration requirement (60 frames is a fraction of a second of streaming at typical rates) |
+| `min_presence_z` | 1.5 | For anchors that expect a person, the mean amplitude z-score must exceed this or the anchor is rejected as "no person detected" |
+| `empty_max_z` | 1.0 | For the `empty` anchor, the z-score must stay under this or it's rejected as "room not empty" |
+| `max_still_motion` | 0.6 (60%) | For still anchors, motion-flagged frame fraction above this is rejected as "too much motion" |
+| `min_move_motion` | 0.3 (30%) | For `small_move`, motion-flagged fraction below this is rejected as "not enough motion" |
+
+A rejected anchor is re-prompted, up to `--attempts` times (default **2**).
+If an anchor is still rejected after all attempts, `enroll` moves on without
+it and logs `"moving on without '<label>'"` — enrollment does **not** abort;
+you end up with a partial anchor set.
+
+**`train-room` itself enforces almost nothing.** It only bails if the
+enrollment file has *zero* accepted anchors at all
+(`v2/crates/wifi-densepose-cli/src/room.rs:246-248`, `"no accepted anchors …
+re-run enroll"`). There is no minimum anchor count beyond that. What actually
+happens with a partial anchor set is that individual specialists silently
+fail to train and are simply absent from the resulting bank — for example
+(from `v2/crates/wifi-densepose-calibration/src/specialist.rs`):
+
+- **presence** needs the `empty` anchor plus at least one anchor where a
+  person was expected present — missing either, `PresenceSpecialist::train()`
+  returns `None` and presence detection is unavailable in that bank.
+- **anomaly** needs at least 2 anchors total, of any kind.
+- **restlessness** needs `sleep_posture` (or `lie_down` as a fallback) *and*
+  `small_move`.
+- **posture** needs at least one anchor that establishes a posture
+  (`stand_still`, `sit`, `lie_down`, or `sleep_posture`).
+
+So a "successful" `train-room` run can still produce a bank missing one or
+more specialists if enrollment didn't collect the anchors those specialists
+need. `room-status` (`v2/crates/wifi-densepose-cli/src/room.rs`) is the way
+to check what actually trained.
+
+### What we could not verify
+
+The ADR-151 design document (§2.2) claims total guided enrollment is
+"~4 minutes of wall-clock" — that arithmetic checks out against the coded
+per-anchor durations (5 × 20s + 3 × 30s = 190s ≈ 3.2 min, plus a 3-second
+countdown before each anchor ≈ +24s, so ~3.5–4 minutes is consistent with the
+code). But we found **no integration test or measurement showing that this
+duration is sufficient for reliable specialist accuracy** — the ADR's own
+status section says the full `baseline → enroll → train-room → infer` loop is
+proven only against **deterministic synthetic CSI** (`tests/full_loop.rs`),
+not yet run start-to-finish on real hardware in an empty room. Treat the
+default durations as reasonable code defaults, not as a validated minimum for
+real-world accuracy.
+
+---
+
+## 2. Recommended duration if there's no hard minimum
+
+Where a hard minimum *does* exist (the 600-frame baseline, the 60-frame
+per-anchor floor), it's documented above. Beyond that:
+
+- **Baseline capture**: the CLI default (`--duration-s 30`) is the number to
+  use; it's what the 600-frame minimum is designed around at the assumed
+  20 Hz sensing rate. ADR-135 §2.3 argues 30 s is the shortest duration that
+  keeps the phase-concentration estimate's standard deviation under
+  0.02 rad², citing published circular-statistics error bounds — but this is
+  a paper-derived justification for the *default value*, not a code-enforced
+  floor beyond the 600-frame check itself.
+- **Enrollment anchors**: use the built-in per-anchor durations (20s/30s) —
+  there's currently no way to change them from the CLI anyway.
+
+---
+
+## 3. Will a pet get classified as "occupied"?
+
+**Honest answer: the code has no way to distinguish a pet (or any small/animal-scale
+motion) from a person.** This is a real limitation, not a solved problem —
+flagging it here rather than guessing.
+
+Presence detection (`PresenceSpecialist`,
+`v2/crates/wifi-densepose-calibration/src/specialist.rs:100-198`) is trained
+purely from two scalar channels measured during enrollment:
+
+- **variance** of the CSI amplitude series, thresholded at the midpoint
+  between the `empty` anchor's variance and the mean variance of the
+  person-present anchors;
+- **mean shift** — `|mean − empty_mean|`, thresholded at half the
+  empty→occupied mean distance.
+
+Presence fires if **either** channel crosses its threshold. Both thresholds
+are learned entirely from the amplitude statistics of your enrollment
+anchors — there is no body-size, RCS (radar cross-section), Doppler-signature,
+or any other physical feature in this code that separates "a full-grown
+adult moved" from "a cat walked past" or "a dog jumped on the couch." If a
+pet's motion perturbs the CSI amplitude by roughly the same amount as the
+`small_move` anchor did during your enrollment, `PresenceSpecialist` will read
+it as occupied, because that's mechanically what the threshold measures.
+
+The closest thing to a safeguard is `AnomalySpecialist`
+(`specialist.rs:386-448`), a generic novelty detector that flags a live
+window as "anomalous" when it's far (in embedding distance) from every
+enrolled anchor prototype. It is **not** a validated pet filter — it will
+flag *any* statistically unusual signal as anomalous or normal depending on
+how close it happens to land to your anchors, with no guarantee it
+distinguishes species or motion source. A pet whose motion pattern happens
+to resemble the `small_move` anchor would not be flagged as anomalous at all.
+
+**Practical takeaway for a homeowner with pets:** expect presence/posture
+readings to occasionally trigger on pet motion, especially larger animals or
+motion near the sensor. There is currently no configuration option or code
+path to suppress this.
+
+---
+
+## 4. Does the empty-room baseline need "typical" conditions (HVAC running) or true silence?
+
+The short answer, grounded in how the baseline is actually computed: **a
+stationary, continuously-running interferer (a fan, HVAC blower, humidifier)
+that is present for the *entire* capture window becomes part of what "empty"
+means, and gets subtracted out naturally** — that's a direct consequence of
+how the statistics are computed, not a documented feature you have to
+configure.
+
+`CalibrationRecorder` uses Welford's online algorithm to accumulate a running
+mean and variance per subcarrier over however many frames you feed it
+(`ruvsense/calibration.rs`). If a fan is running steadily the whole time you
+capture the baseline, its contribution is baked into `amp_mean`/`amp_variance`
+for every frame equally, so the resulting baseline already represents "empty
+room with the fan on" — and at runtime, `BaselineCalibration::subtract()`
+removes exactly that reference, so a room in the same steady state reads as
+quiet. The design intent documented in ADR-135 §1.1 is explicit about this:
+the whole point of baseline subtraction is to remove "hardware-induced gain
+bias and environment-fixed multipath" so downstream motion detectors aren't
+tripped by things that are always there.
+
+**What actually matters is consistency, not silence**: capture the baseline
+under whatever background conditions the room will normally be in during
+real use (HVAC/fans running as usual), and try to keep the room in that same
+steady state for the entire capture window. What the code cannot correct
+for is a background condition that **changes partway through** the capture
+(e.g. HVAC cycles on 15 seconds into a 30-second capture) — that would bias
+the Welford mean/variance toward an in-between state that matches neither
+"HVAC off" nor "HVAC on" well.
+
+There is a **real-time guard during capture** that can catch gross problems:
+`--abort-z-threshold` (default `2.0`) aborts the capture if the per-frame
+amplitude z-score median stays above that threshold for 20 consecutive
+banner intervals (`v2/crates/wifi-densepose-cli/src/calibrate.rs:82-83,
+163-178`). This is designed to catch someone walking through mid-capture, not
+necessarily short-duration mechanical noise — we found no test exercising it
+against an HVAC-cycling scenario specifically, so how it behaves for
+"appliance turns on mid-capture" is unverified.
+
+### What we could not verify — and a design gap worth knowing about
+
+ADR-135 §2.5 describes a much more sophisticated staleness-detection system:
+a `drift_score` computed from ongoing z-scores, a `BaselineDrift` event fired
+after sustained drift, and a `baseline_stale` flag published over the
+sensing WebSocket. **We searched the actual `calibration.rs` implementation
+and none of that exists in code** — there is no `drift_score` field, no
+`BaselineDrift` event, and no `baseline_stale` flag anywhere in
+`v2/crates/wifi-densepose-signal/src/ruvsense/calibration.rs`. That part of
+ADR-135 is aspirational design, not shipped behavior.
+
+What *is* implemented, at a different layer, is a much simpler check on the
+**trained specialist bank** (not the raw baseline): `SpecialistBank` stores
+the `baseline_id` it was trained against, and `SpecialistBank::is_stale()`
+(`v2/crates/wifi-densepose-calibration/src/bank.rs:102-104`) returns `true`
+whenever the *current* baseline's id doesn't match the id the bank was
+trained on. Re-running `calibrate` always produces a new baseline id, so
+**any** recalibration — whether because of furniture moving, a genuinely
+stale reference, or just re-running the command — immediately marks every
+previously trained specialist bank stale, and you'll need to re-run `enroll`
+and `train-room` afterward. There is no partial/graded staleness signal
+(no "how stale"), only this all-or-nothing id comparison.
+
+**Practical guidance:**
+
+1. Calibrate with the room in its normal, steady background state (HVAC,
+   fans, fridge compressor, etc. running as they normally would) and keep
+   that state constant for the whole `--duration-s` window.
+2. If you significantly change background conditions later (move furniture,
+   add a permanent appliance, change HVAC routine) or notice the sensing
+   quality degrade, re-run `calibrate` — this is an explicit, operator-driven
+   step; there is no code path that recalibrates for you.
+3. Re-running `calibrate` invalidates every specialist bank trained against
+   the old baseline (via the `baseline_id` mismatch above) — plan to re-run
+   `enroll` and `train-room` right after.
+
+---
+
+## Quick reference: commands and defaults actually in the code
+
+```bash
+# Stage 1 — empty-room baseline. Room must be empty for the whole window.
+wifi-densepose calibrate \
+  --udp-port 5005 --duration-s 30 --tier ht20 --output ./baseline.bin
+# Hard requirement: >= 600 recorded frames, or calibration fails.
+
+# Stage 2+3 — guided enrollment (8 fixed anchors, ~4 minutes total)
+wifi-densepose enroll --baseline ./baseline.bin --room living-room \
+  --output ./enrollment.json --attempts 2
+
+# Stage 4 — train the specialist bank from whatever anchors were accepted
+wifi-densepose train-room --enrollment ./enrollment.json \
+  --output ./room-bank.json
+
+# Check what actually trained (and whether the bank is stale)
+wifi-densepose room-status --room living-room
+```
+
+Source references for everything above:
+- `v2/crates/wifi-densepose-cli/src/calibrate.rs`
+- `v2/crates/wifi-densepose-cli/src/room.rs`
+- `v2/crates/wifi-densepose-signal/src/ruvsense/calibration.rs`
+- `v2/crates/wifi-densepose-calibration/src/enrollment.rs`
+- `v2/crates/wifi-densepose-calibration/src/anchor.rs`
+- `v2/crates/wifi-densepose-calibration/src/specialist.rs`
+- `v2/crates/wifi-densepose-calibration/src/bank.rs`
+- `docs/adr/ADR-135-empty-room-baseline-calibration.md`
+- `docs/adr/ADR-151-room-calibration-specialist-training.md`
diff --git a/docs/trust-and-engine-errors.md b/docs/trust-and-engine-errors.md
new file mode 100644
index 00000000..1ba6ca18
--- /dev/null
+++ b/docs/trust-and-engine-errors.md
@@ -0,0 +1,250 @@
+# Trust State & Engine Errors
+
+If you've seen the sensing-server log a growing `engine_error_count`, or
+noticed your deployment reads as `"demoted": true` on the status endpoint,
+this page explains — from the actual code, not the design docs — what those
+two things mean, what triggers them, where you can see them, and what your
+real options are.
+
+Everything below is grounded in
+`v2/crates/wifi-densepose-sensing-server/src/engine_bridge.rs`,
+`v2/crates/wifi-densepose-sensing-server/src/main.rs`,
+`v2/crates/wifi-densepose-engine/src/lib.rs`, and
+`v2/crates/wifi-densepose-signal/src/ruvsense/multistatic.rs`.
+
+## Two different things, easy to conflate
+
+The sensing-server runs a "governed trust cycle" every sensing tick
+(`StreamingEngine::process_cycle`, driven by `EngineBridge::observe_cycle` in
+`engine_bridge.rs:193-223`). Each cycle produces **one of two outcomes**, and
+they are tracked completely separately:
+
+1. **The cycle fails outright** (`Result::Err(EngineError)`) — nothing is
+   published for that tick. This increments `engine_error_count`, a
+   monotonically increasing counter.
+2. **The cycle succeeds but under a demoted privacy class**
+   (`Result::Ok(TrustedOutput { demoted: true, .. })`) — a belief *is*
+   published, just at a more restricted privacy class than normal. This sets
+   the `demoted` flag, which is recomputed fresh on every successful cycle.
+
+A deployment can have a high `engine_error_count` with `demoted: false` (lots
+of failed cycles, but the ones that succeed are clean), or `demoted: true`
+with `engine_error_count: 0` (every cycle succeeds, but under a downgraded
+privacy class), or both at once — which is what the issue reporter saw.
+
+## 1. Exact conditions for each
+
+### Engine errors (`engine_error_count`)
+
+`engine_error_count` increments only when
+`StreamingEngine::process_cycle` returns `Err(EngineError::Fusion(..))`
+(`engine_bridge.rs:198-222`). `EngineError` wraps
+`wifi_densepose_signal::ruvsense::multistatic::MultistaticError`
+(`wifi-densepose-engine/src/lib.rs:54-69`), which has exactly four variants
+(`multistatic.rs:36-56`):
+
+| Variant | Condition |
+|---|---|
+| `NoFrames` | No node frames were passed to fusion. In practice not reachable through the bridge: `process_cycle_from_states` returns `None` (not an error, not counted) before calling the engine at all if there are no frames (`engine_bridge.rs:168-171`). |
+| `InsufficientNodes(n)` | Fewer than 2 nodes contributing in multistatic mode. |
+| `TimestampMismatch { spread_us, guard_us }` | The spread between contributing nodes' frame timestamps exceeds the **hard guard interval**, default **60,000 µs (60 ms)** (`MultistaticConfig::default()`, `multistatic.rs:133`). |
+| `DimensionMismatch { node_idx, expected, got }` | A node's subcarrier count doesn't match the others. As of #1170 the live bridge canonicalizes every node onto a common 56-tone grid before fusion, so this is now rare on real hardware — see the comment on `observe_cycle_counts_engine_errors` in `engine_bridge.rs`. |
+
+Regardless of cause, an error is **rate-limited in the log** to one
+`tracing::warn!` line per 10 seconds (`ENGINE_ERROR_WARN_INTERVAL`,
+`engine_bridge.rs:50, 206-219`) — errors are still counted every cycle, only
+the *log line* is throttled, so a 20 Hz loop failing continuously won't flood
+your log with 20 lines/second.
+
+### Trust demotion (`demoted`)
+
+This is a **separate mechanism** from engine errors: it happens on cycles
+that *succeed*, and it downgrades the privacy class the output is emitted
+under, one step, rather than failing the cycle. From
+`wifi-densepose-engine/src/lib.rs:514-515`:
+
+```rust
+let demoted = quality.forces_privacy_demotion() || array_contradiction || mesh_at_risk;
+let effective_class = if demoted { demote_one(base_class) } else { base_class };
+```
+
+Three independent conditions can trigger it:
+
+- **`quality.forces_privacy_demotion()`** — true whenever the fusion
+  quality record carries any non-empty `contradiction_flags`
+  (`fusion_quality.rs:111-118`). These are *tolerated* disagreements, distinct
+  from a hard fusion failure:
+  - `TimestampMismatch` — spread within the **hard** guard but beyond the
+    **soft** guard (default **20,000 µs / 20 ms**, `soft_guard_us`,
+    `multistatic.rs:104-113`) — i.e. loose-but-tolerable timing alignment.
+  - `CalibrationIdMismatch` — contributing frames disagree on which
+    calibration epoch (baseline) they were captured under.
+  - `PhaseAlignmentFailed`, `DriftProfileConflict`, `CoherenceDrop`,
+    `GeometryInsufficient` — raised upstream by the array coordinator /
+    baseline drift checks.
+- **`array_contradiction`** — a separate array-level directional-fusion
+  contradiction check.
+- **`mesh_at_risk`** — the mesh is close to partitioning (`mesh_guard.rs`).
+
+`demote_one()` (`lib.rs:688-690`) steps the privacy class exactly one notch
+toward `Restricted` (it never jumps more than one step, and never relaxes a
+class in the same cycle — proven by the `forced_contradiction_never_relaxes_class`
+test). At `PrivacyClass::Restricted`, `EngineBridge::suppress_raw_outputs()`
+becomes true and `main.rs` strips per-node raw amplitude vectors from the
+published `SensingUpdate` (`engine_bridge.rs:251-260`).
+
+**Crucially, `demoted` is not sticky.** It is overwritten on every
+successful cycle to reflect *that cycle's* outcome
+(`self.demoted = trust.demoted;`, `engine_bridge.rs:203`). If the
+contradiction that caused demotion was transient, the very next clean cycle
+reports `demoted: false` again with no action from you. If you see
+`demoted: true` *persistently*, that means the underlying condition (usually
+clock drift beyond the guard, or a geometry/calibration disagreement) is
+itself persistent, not that something got "stuck."
+
+## 2. Where this is exposed
+
+Both `GET /health/ready` and `GET /api/v1/status` are wired to the same
+handler (`health_ready`, `main.rs:8128,8133`) and return a `trust` block
+(`main.rs:4589-4606`):
+
+```json
+{
+  "status": "ready",
+  "trust": {
+    "last_witness": "…64 hex chars or null…",
+    "effective_class": "Anonymous | Restricted | …",
+    "demoted": false,
+    "recalibration_recommended": false,
+    "engine_error_count": 0,
+    "raw_outputs_suppressed": false
+  }
+}
+```
+
+**This is a real, currently-shipped diagnostic surface — but it is honestly
+limited.** It tells you *that* errors are occurring and *that* the current
+class is demoted, and the total count, but not *why* for your specific run:
+
+- `engine_error_count` is a single running total. There is **no breakdown by
+  error type** anywhere in the API or in `EngineBridge`'s state — you cannot
+  tell from `/health/ready` whether your 20,000 errors are 20,000
+  `TimestampMismatch`es or 20,000 `DimensionMismatch`es.
+- `demoted` is a boolean with no accompanying list of which
+  `ContradictionFlag`s actually fired. The underlying `contradiction_flags`
+  vector exists in `QualityScore` (`fusion_quality.rs:106`) but is not
+  surfaced over the wire anywhere we found.
+- There's no error history/timeline, and no per-node breakdown (which node
+  is the one whose clock is drifting, for instance).
+
+**The closest thing to a real diagnostic today is the rate-limited log
+line itself.** Unlike the API, the log message includes the `Display` text
+of the actual `EngineError`, which for `TimestampMismatch` and
+`DimensionMismatch` includes the concrete numbers (e.g. `"Timestamp spread
+87000 us exceeds guard interval 60000 us"`, `"Dimension mismatch: node 2 has
+114 subcarriers, expected 56"`). If you're trying to diagnose a specific
+demotion/error episode today, grepping the sensing-server log for
+`"governed trust cycle failed"` is the most concrete answer available — the
+status endpoint alone will not tell you the underlying cause. Treat this as
+the honest state of the diagnostics, not a missing feature we're pretending
+exists.
+
+## 3. Is a demoted / errored state permanent? Is there a reset?
+
+**`demoted` never needs resetting** — as described above, it's recomputed
+every successful cycle from that cycle's own contradiction/mesh state. There
+is no persistence, no counter, no cooldown timer for it in the code.
+
+**`engine_error_count` has no reset mechanism at all.** It is a plain `u64`
+field on `EngineBridge`, initialized to `0` in `EngineBridge::new`
+(`engine_bridge.rs:111`) and only ever incremented
+(`self.engine_error_count += 1;`, line 207) — there is no method, admin
+endpoint, or timer anywhere in the crate that decrements or clears it. The
+only way to bring it back to zero is to **restart the sensing-server
+process**, which constructs a brand-new `EngineBridge`. If your count is
+growing and you want to confirm whether a fix actually worked, restart the
+server and watch whether the count starts climbing again — there is
+currently no lighter-weight way to "clear the counter" without a restart.
+
+**If demotion (or errors) are persistent rather than one-off**, the
+documented, real fix for the most common cause — clock drift between nodes
+exceeding the fixed 60 ms hard guard — is an environment-variable override,
+not a restart or a wait:
+
+- `WDP_GUARD_INTERVAL_US` — directly overrides the hard guard (e.g.
+  `WDP_GUARD_INTERVAL_US=200000` for a 200 ms guard). This is the escape
+  hatch a real deployment (issue #1049) needed: WiFi/ESP-NOW-synced ESP32
+  nodes were measured drifting 10–150 ms, which the published 60 ms default
+  could not absorb, causing **every** cycle to demote with "no escape hatch"
+  (see the comment at `main.rs:8336-8339`).
+- `WDP_SOFT_GUARD_US` — optionally overrides the soft (tolerated-contradiction)
+  guard, always clamped below the hard guard.
+- `WDP_TDM_SLOTS` + `WDP_TDM_SLOT_US` — derive the guard from your actual TDM
+  schedule instead of setting it directly.
+
+See `multistatic_guard_config_from_env` / `multistatic_guard_config_from`
+(`main.rs:6791-6856`) for the exact precedence rules (a direct
+`WDP_GUARD_INTERVAL_US` always wins over the TDM-derived value).
+
+## 4. Does a converted Hugging Face model explain this?
+
+**We could not find a code path connecting `--convert-model` to engine
+errors or trust demotion — they appear to be entirely separate subsystems.**
+Saying this plainly rather than speculating:
+
+- `--convert-model` (`main.rs:6976-7028`, `run_convert_model` /
+  `load_or_convert_model` at `main.rs:6925-6974`) converts a **pose-model
+  weights file** — Hugging Face `safetensors` or a `jsonl` manifest — into
+  this project's own RVF binary container format, so it can be loaded via
+  `--model`. This is entirely about which neural-network weights the pose
+  estimator uses.
+- `engine_error_count` and `demoted` come from `StreamingEngine::process_cycle`
+  in `wifi-densepose-engine`, which performs **multistatic CSI sensor
+  fusion** — checking node count, per-node timestamp spread, and per-node
+  subcarrier dimensions across your ESP32 nodes. This code path has no
+  dependency on which pose model is loaded, and `load_or_convert_model` /
+  `run_convert_model` never call into `engine_bridge` or
+  `StreamingEngine` at all.
+
+Because the code shows no coupling between the two, we are not going to
+invent one. Two possibilities that the code doesn't rule out, but also
+doesn't confirm, if you hit both symptoms together:
+
+- **Coincidence** — the deployment that had trouble loading/using a
+  converted model separately had a fusion-timing or node-count problem
+  (e.g. the #1049-style clock-drift issue, or fewer than 2 active nodes),
+  unrelated to the model conversion itself.
+- **A configuration change made alongside the model swap** — e.g. changing
+  node count, geometry, or guard settings at the same time as switching
+  models — could produce both symptoms together without the model itself
+  being the cause.
+
+If you're hitting this, the actionable step from the code is to check
+`engine_error_count` and the log line's error text (per §2 above)
+**independently** of whatever model you have loaded — if the errors are
+`TimestampMismatch`/`DimensionMismatch`/`InsufficientNodes`, the fix is on
+the sensor-fusion side (§3), not the model side, regardless of which model
+produced the report.
+
+## Quick reference
+
+```bash
+# Check current trust state
+curl -s http://localhost:3000/api/v1/status | jq .trust
+
+# Watch for the rate-limited error log line (most specific diagnostic today)
+# — look for "governed trust cycle failed" in the sensing-server's stderr/log.
+
+# If demotion/errors are persistent due to node clock drift, raise the guard:
+WDP_GUARD_INTERVAL_US=200000 wifi-densepose-sensing-server ...
+
+# The only way to reset engine_error_count is a process restart.
+```
+
+Source references for everything above:
+- `v2/crates/wifi-densepose-sensing-server/src/engine_bridge.rs`
+- `v2/crates/wifi-densepose-sensing-server/src/main.rs` (search `trust`, `health_ready`, `multistatic_guard_config_from`, `convert_model`)
+- `v2/crates/wifi-densepose-engine/src/lib.rs`
+- `v2/crates/wifi-densepose-engine/src/mesh_guard.rs`
+- `v2/crates/wifi-densepose-signal/src/ruvsense/multistatic.rs`
+- `v2/crates/wifi-densepose-signal/src/ruvsense/fusion_quality.rs`