GrumpyTanker · GrumpyTanker · Jun 8, 2026 · Jun 4, 2026 · Jun 4, 2026 · Jun 4, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,64 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.8.0] — 2026-06-03
+
+Active/standby load-state split for bimodal loads — classify each record by
+current, report the two states separately, correct the session energy three
+ways, headline the active-state power factor, and harden the auto reverse-CTs
+heuristic to decide on the dominant high-current state. Parity-tested Python↔JS.
+Built from the real P115RE coating-rectifier session.
+
+### Added — load-state split (`load_states`)
+- **Current-gated classifier.** Each record is **active** when its mean
+  per-phase average current `(I_a_avg+I_b_avg+I_c_avg)/3` is **≥ the threshold**
+  (default **50 A**), else **standby**. Current — not power — because the power
+  *sign* at low current is exactly the thing in question. Configurable with
+  **`--standby-threshold-a N`**.
+- **`load_states.csv` + `load_states.json`** — one row per state (active,
+  standby): records, hours, duty %, kWh, P avg/min/max (kW), I avg (A), S avg
+  (kVA), PF avg, V_LN avg, V_THD p95. JSON also carries the threshold, the three
+  energy figures, and the standby-sign caveat. Emitted automatically (cheap);
+  `--load-states` is an explicit opt-in. A compact table is embedded in
+  `summary.txt` and the HTML report.
+- **Three explicit energy figures** (never silently changing the historic
+  number): `energy_as_measured_kWh` (signed sum — current behavior),
+  `energy_active_kWh` (active records only), and
+  `energy_net_clip_standby_kWh` (standby real power clipped to ≥0). The standby
+  real-power sign is unreliable at low current, so the active/clip figures are
+  the defensible consumption.
+- **Headline PF is the active-state PF.** The narrative, summary, and HTML
+  report now headline the active-state power factor (the blended whole-session
+  PF is meaningless for a bimodal load); the raw whole-session PF is kept but
+  de-emphasized.
+
+### Changed — magnitude-weighted reverse-CTs auto-detect
+- `detect_ct_reversal` / `--auto-reverse-cts` now decide on the **dominant
+  high-current (active) state** — *is real power negative when current is
+  high?* — instead of the fragile whole-session negative-P count, which a
+  bimodal load defeats. The whole-session count fields are still reported for
+  context (`basis: "active"` vs `"whole_session"`, with `active_records`,
+  `active_frac_negative`, `active_mean_p_w`); the operator notice keys off the
+  active state. **Manual `--reverse-cts` behavior is unchanged** — only the AUTO
+  heuristic + its printed notice were improved.
+
+### Changed — shift integration
+- `shift_comparison.csv/json` rows gain **`active_records`**,
+  **`active_duty_pct`**, **`active_kWh`**, and **`active_PF_avg`** so each shift
+  shows its active load too. `summary.txt`'s shift table shows the new columns.
+
+### Parity / tests / docs
+- JS port in `web/analysis.js` (`classifyLoadStates`, `loadStateRows`,
+  `sessionEnergy`, `activeStatePf`; active-state `detectCtReversal`; active
+  columns on `shiftComparisonRows`) with a Python↔JS parity test
+  (`load_states_parity.test.js`) against a shared bimodal golden fixture.
+- New Python tests: classifier (threshold, balanced bimodal fixture), the three
+  energy figures, the magnitude-weighted reverse-CTs decision (active-positive
+  and active-negative cases + fallback), shift active columns, narrative
+  active-PF/energy, and CLI `load_states` outputs.
+- `docs/LOAD_STATES.md` (concept, the energy caveat, the standby-sign
+  explanation); README options; version bump to 0.8.0.
+
 ## [0.7.0] — 2026-06-03
 
 Generalized, named, configurable shift/period splitting — so usage can be

diff --git a/README.md b/README.md
@@ -15,6 +15,7 @@ This project gives you:
 - **A Python CLI** for scripting, batch jobs, and publication-quality gnuplot output. `pip install -e .` and you're going.
 - **Auto event detection** — outages, voltage dips, swells, high-current peaks, frequency excursions, NEMA imbalance spikes, sudden load steps.
 - **CT reversal correction** — `--reverse-cts` flag handles iFlex probes installed backwards (extremely common mistake). See [`docs/CT_REVERSAL.md`](docs/CT_REVERSAL.md).
+- **Active/standby load-state split** — for bimodal loads (e.g. a rectifier that toggles between a heavy active draw and a light standby), classify each record by current, report the two states separately, headline the active-state power factor, and correct the session energy three ways. See [`docs/LOAD_STATES.md`](docs/LOAD_STATES.md).
 
 ## Web app — no install
 
@@ -100,7 +101,9 @@ fluke-analyze path/to/ES.NNN -o output/ \
 | `--tod-bin MINS` | `1` | Time-of-day bin width in minutes. |
 | `--demand-window SECS` | `900` | Rolling peak-demand window. Reports peak demand + the window it occurred in (`demand.json` + XLSX). See [`docs/DEMAND.md`](docs/DEMAND.md). |
 | `--tz ZONE` | UTC | Render report timestamps in local + UTC for an IANA zone (e.g. `America/Chicago`). Default UTC only. |
-| `--auto-reverse-cts` | off | Auto-detect a reversed-CT install (sustained negative real power) and apply `--reverse-cts` automatically, with a loud notice. See `docs/CT_REVERSAL.md`. |
+| `--auto-reverse-cts` | off | Auto-detect a reversed-CT install and apply `--reverse-cts` automatically, with a loud notice. Decides on the dominant **high-current (active)** state — robust for bimodal loads. See `docs/CT_REVERSAL.md`. |
+| `--standby-threshold-a A` | `50` | Per-phase mean current (A) at/above which a record is **active** load (else standby). Drives the active/standby split (`load_states.csv`/`.json`), the energy correction (three figures), and the magnitude-weighted reverse-CTs decision. See [`docs/LOAD_STATES.md`](docs/LOAD_STATES.md). |
+| `--load-states` | auto | Force the active/standby load-state report (emitted by default; flag is an explicit opt-in). |
 | `--rules-file FILE` | off | JSON/TOML EventRules overrides keyed by asset name. See [`docs/RULES_FILE.md`](docs/RULES_FILE.md). |
 | `--no-stats` | | Skip whole-session statistics (`stats.json`/`stats.csv` + XLSX sheet). |
 | `--format` | `png` | `png` or `svg` |

diff --git a/ROADMAP.md b/ROADMAP.md
@@ -4,6 +4,28 @@ Shipped releases live in [CHANGELOG.md](CHANGELOG.md). This file is the
 backlog of ideas we've discussed but not yet committed to a specific
 release.
 
+## Shipped in v0.8 — active/standby load-state split
+
+Bimodal-load handling, built from the real P115RE coating-rectifier session,
+parity-tested Python↔JS:
+
+- **Current-gated classifier** (`--standby-threshold-a`, default 50 A) splitting
+  each record into active vs standby, reported separately in
+  `load_states.csv`/`.json`.
+- **Three explicit energy figures** (as-measured / active-only / net-clip
+  standby) — the standby real-power sign is unreliable at low current, so the
+  historic signed sum understates consumption; the report never silently
+  changes it.
+- **Headline PF = active-state PF** in the narrative/summary/HTML (the blended
+  whole-session PF is meaningless for a bimodal load).
+- **Magnitude-weighted reverse-CTs auto-detect** — decides on the dominant
+  high-current (active) state instead of the fragile whole-session negative-P
+  count. Manual `--reverse-cts` unchanged.
+- Shift rows gain `active_duty_pct` / `active_kWh` / `active_PF_avg`.
+- New `analysis.py` funcs (`classify_load_states`, `load_state_rows`,
+  `session_energy`, `active_state_pf`) + JS port + parity test +
+  `docs/LOAD_STATES.md`.
+
 ## Shipped in v0.7 — generalized shift splitting
 
 Operator-defined, named, midnight-wrapping shift windows so usage can be

diff --git a/docs/LOAD_STATES.md b/docs/LOAD_STATES.md
@@ -0,0 +1,129 @@
+# Load states: active vs standby (`load_states`)
+
+Many factory loads are **bimodal** — they alternate between a heavy *active*
+draw and a light *standby* state. Blending the two into one session mean buries
+the real consumption and produces a meaningless average power factor. The
+load-state report classifies every record as **active** or **standby**, reports
+the two states separately, surfaces the **active-state PF**, and corrects the
+session energy.
+
+This was built from the real P115RE coating-rectifier session (ES.004,
+`--reverse-cts`, `America/Chicago`):
+
+| state   | duty | I/phase | P_total | P1 (fundamental) | PF    |
+|---------|------|---------|---------|------------------|-------|
+| active  | ~49% | ~239 A  | +97 kW  | —                | +0.47 |
+| standby | ~47% | ~16 A   | −7.6 kW | −7.7 kW          | −0.64 |
+
+## Why classify by *current*, not power
+
+The whole question is whether the **power sign** is trustworthy. With the global
+`--reverse-cts`, the active state reads +97 kW (correct — that is the coating
+draw), but the standby state reads −7.6 kW. A rectifier in standby should draw
+small *positive* core/copper losses, not export. The standby reading is balanced
+across all three phases and the **fundamental** P1 is also negative (−7.7 kW), so
+it is not a harmonic artifact — it is simply that **no single CT polarity makes
+both states physical**, and the low-current sign is unreliable.
+
+So we gate on **current**, which is unambiguous:
+
+> A record is **active** when its mean per-phase average current
+> `(I_a_avg + I_b_avg + I_c_avg) / 3` is **≥ the threshold** (default **50 A**),
+> otherwise **standby**.
+
+A single dropped phase is ignored (the mean is taken over the finite phases), so
+one NaN does not drag a record into standby. The threshold is configurable with
+`--standby-threshold-a N`. A single threshold is used (no transition band); set
+it between the two clusters — for the P115RE the active state is ~239 A and the
+standby state is ~16 A, so the 50 A default sits comfortably between them.
+
+## The three energy figures
+
+Because the standby sign is unreliable, the as-measured signed energy is
+**understated** (the bogus −7.6 kW standby subtracts). The report surfaces all
+three figures explicitly — it never silently changes the historic number:
+
+- **`energy_as_measured_kWh`** — the signed sum of `P_total_avg_W` over all
+  records. This is the existing/historic behavior. *(P115RE: ~6,054 kWh.)*
+- **`energy_active_kWh`** — energy from the **active** records only.
+  *(P115RE: ~6,638 kWh.)*
+- **`energy_net_clip_standby_kWh`** — active records pass through unchanged;
+  standby real power is **clipped to ≥ 0** (a rectifier in standby never
+  exports). *(P115RE: ~6,684 kWh.)*
+
+All three use the tool's standard convention: per record (1 s) energy =
+`P_total_avg_W / 1000 / 3600`, summed; non-finite samples are skipped.
+
+> **Caveat (carried in every output):** standby real-power sign is unreliable at
+> low current, so the **active** and **clip** figures are the defensible
+> consumption — not the as-measured signed sum.
+
+## Headline power factor
+
+For a bimodal load the blended whole-session PF is meaningless (the P115RE
+blended PF is −0.09). The **meaningful** figure is the **active-state PF**
+(~0.47), so that is what the narrative, summary, and HTML report headline. The
+raw whole-session PF is still reported, but de-emphasized and labeled.
+
+## Outputs
+
+`load_states.csv` — one row per state (active, then standby):
+
+| column          | meaning                                  |
+|-----------------|------------------------------------------|
+| `state`         | `active` or `standby`                    |
+| `records`       | record count in the state                |
+| `hours`         | hours (records / 3600)                   |
+| `duty_pct`      | percent of all records in the state      |
+| `kWh`           | mean P × hours for the state             |
+| `P_avg_kW`      | mean real power (kW)                      |
+| `P_min_kW`      | min real power (kW)                       |
+| `P_max_kW`      | max real power (kW)                       |
+| `I_avg_A`       | mean per-phase current (A)               |
+| `S_avg_kVA`     | mean apparent power (kVA)                 |
+| `PF_avg`        | mean power factor                        |
+| `V_LN_avg_V`    | mean L-N voltage (outage zeros excluded) |
+| `V_THD_p95_pct` | 95th-percentile V_THD                    |
+
+`load_states.json` carries the same `states` rows plus `standby_threshold_a`,
+the three `energy` figures, and the caveat `note`. A compact load-state table
+(and the three energy figures) is embedded in `summary.txt` and the HTML report.
+
+The report is emitted automatically (it is cheap — a couple of streaming
+passes); `--load-states` is accepted as an explicit opt-in but is not required.
+
+## Integration with shifts
+
+When `--split-by shifts` is used, each shift row in `shift_comparison.csv/json`
+also carries its **active** load, using the same current cut:
+
+- `active_records` — active record count in the shift
+- `active_duty_pct` — percent of the shift's records that are active
+- `active_kWh` — active-only energy for the shift
+- `active_PF_avg` — active-state PF for the shift
+
+## The magnitude-weighted reverse-CTs decision
+
+The same active/standby insight hardens the **auto** reverse-CTs heuristic
+(`--auto-reverse-cts`). A naive count-based test ("is P negative for ≥ 50 % of
+records?") is fragile for bimodal loads: the P115RE reads negative more than half
+the time (standby) while the real consumption — the active state — is clearly
+positive. So the auto-detect now decides on the **dominant high-current (active)
+state**: *is real power negative when current is high?* The whole-session count
+fields are still reported for context, but `reversed` and the operator notice key
+off the active state (`basis: "active"`). If there is no active population at all
+(everything below the threshold), it falls back to the whole-session count
+(`basis: "whole_session"`). The manual `--reverse-cts` behavior is unchanged —
+only the AUTO heuristic was improved.
+
+## Tuning the threshold
+
+```bash
+# Default 50 A active/standby cut
+fluke-analyze ES.004 --reverse-cts --tz America/Chicago
+
+# Lower the cut so a small steady draw counts as active
+fluke-analyze ES.004 --reverse-cts --standby-threshold-a 25 --tz America/Chicago
+```
+
+See also [CT_REVERSAL.md](CT_REVERSAL.md) and [SHIFTS.md](SHIFTS.md).
diff --git a/python/pyproject.toml b/python/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "fluke-3540-analyzer"
-version = "0.7.0"
+version = "0.8.0"
 description = "Parser, event detector, and chart generator for Fluke 3540 FC three-phase power-quality sessions. See https://github.com/GrumpyTanker/fluke-3540-analyzer."
 requires-python = ">=3.10"
 license = {text = "MIT"}

diff --git a/python/src/fluke_3540/__init__.py b/python/src/fluke_3540/__init__.py
@@ -4,4 +4,4 @@
 The canonical field map and binary layout live in spec/field_map.json at
 the repo root and are shared with the JavaScript port. See README.md.
 """
-__version__ = "0.7.0"
+__version__ = "0.8.0"