Roadmap

Each stage is one method; a stage ships when it has a benchmark row and a test. This page names what shipped at each stage and what it still owes.

Stage	Theme	Status
1	Store + data-physics layer	shipped (3W real, TEP real)
2	Features + splits + first published number	shipped (synthetic backbone)
3	Provisional baselines (MAD / moving-range screen)	shipped (synthetic; 3W real with caveats)
4	Refined per-tag regime baselines	shipped (synthetic; 3W real with caveats)
5	SPC	shipped (synthetic; 3W real with caveats)
6	MSPC (PCA T2/SPE)	shipped (synthetic; 3W real with caveats)
7	ML detectors + static evidence viewer	shipped (synthetic; 3W real with caveats)
8	LLM narration over the evidence ledger	shipped (offline-gated)

Stages 1 to 6 are reachable from the CLI (profile, segment, screen, spc, mspc, compare, switchback). Stage 7 is library-only, because it needs feature rows over many windows and the ml extra. Every stage ships benchmark rows in BENCHMARKS.md, computed from SYNTHETIC data, either the validation backbone or seeded noise where the dataset column says so, and compared byte for byte in CI. The REAL sections follow the synthetic table, one per study in examples/studies/.

Real-data numbers are published under group holdout by asset; each names its fold split. The reading rules are in each study's REPORT.md in examples/studies/.

Status

Stages 1 to 8 ship in 0.3.0, with the evaluation protocol as tsdive.eval.

Parked, each with the measured reason:

• Conformal martingale alarm. It lives in tsdive.eval and is not wired into screen or spc. On 3W normal records it fires on 76.3% at delta 0.05 against a 5% bound, and the permutation check gives 0.04%, so the bound needs an exchangeability the records do not have. Report: examples/studies/3w_conformal/REPORT.md.
• Compression detection. A Thornhill-style compression-factor estimate saturates on data rounded to a lattice: the estimate stays near 2 whether the true factor is 1 or 40. It waits for an archive with documented compression settings.
• Control-loop performance (Harris index, stiction). It waits for a loop archive with a loop schema. A candidate archive exists (Bauer 2019, IECR).
• Moving baselines in the shipped tools. A baseline refitted over the three windows before each scored window takes the false-alarm rate on 3W records with no fault window from 71.2% to 29.3% per window, and on the SKAB anomaly-free record from 95.7% to 47.2%, against a 25.0% floor. It also keeps a 3W fault above its threshold for a median of 2 consecutive fault windows where the static baseline keeps it for 10, and it cannot be defined on the six-window onset-aligned design. screen and spc keep the fixed history. Report: examples/studies/baseline_drift/REPORT.md.
• Intervals on the compare level shift and spread ratio. On synthetic AR(1) series with 480 samples per period the 95% hac and ewc intervals cover 0.872 and 0.878 at phi 0.9, and no method holds its coverage in every cell up to phi 0.9. A placebo date on the 538 3W instances with no fault window gets an interval that excludes 0 on 65.6% (hac) and 63.8% (ewc) of (instance, tag) pairs against a 10% bar, and on 24.4% and 41.5% of the Turbine Upgrade placebo weeks. Regressing on the other tags leaves the 3W interval at 0.927 of the raw width. On the TEP fault-free runs ewc covers the level at 0.924 to 0.942 over three variable groups, and at 0.939 to 0.955 once a persistence refusal chosen after the 3W result drops the rows with fewer than 10 effective samples. compare keeps the shift without an interval. Report: examples/studies/shift_intervals/REPORT.md.
• Zero-shot forecasters as shipped detectors. The Chronos-Bolt study gives the only false-alarm rate under the floor (18.2%) and a post1 detection rate of 27.3%. It stays a benchmark subject. No forecaster ships as a tool.

Next beds: SKAB is the bed where records return to normal after a fault, and its numbers are in examples/studies/skab/REPORT.md. What the frozen-sensor and rounding checks change in the 3W detector numbers is measured in examples/studies/3w_audit_effect/REPORT.md: the zero-scale check drops 690 of 3,870 (instance, variable) pairs, and without it the MAD screen scores AUC 0.818 instead of 0.867 and detects 90.6% of the faults it now catches at 100%.

Stage 0, sources

Shipped: parquet archives, CSV and parquet ingest, and the 3W and TEP converters, each answering a window read under a stated contract through tsdive.store.source.Source. Owed: an OPC UA history client and a message-bus capture. A source ships with a converter or a Source implementation, a benchmark row and a test (docs/SOURCES.md).

Stage 1, store and data-physics layer

Shipped: the store, tsdive profile, and every check it prints. The check-by-check table, with the error each one raises, is the What it checks section of the README. A tag is (source_id, point_id) and the display name is metadata. loop_id and role in {PV, SP, OP, MODE} are stage-1 schema fields, not a later retrofit. A read declares its sampling contract as a positional argument and gets a stable digest back. Provenance fingerprints produce the replay verdicts MATCH, CODE_DRIFT, DATA_CHANGED and BOTH. Two refusal reference cases ship with committed outputs.

Stage 2, features, splits and the first published number

Shipped: contract-aware feature extraction (features/), the synthetic validation backbone with injected fault archetypes and ground truth (data/backbone.py), and provenance-stratified splits whose proportions are exact within each stratum (data/splits.py), all labelled SYNTHETIC in BENCHMARKS.md. stratified_split cuts rows inside a stratum, so it cannot hold a whole group out; group_holdout closes that. Whole groups go to one fold, taken largest first, placed where the per-stratum counts end up least spread. A seeded SHA-256 rank breaks size ties. GroupSplit.leakage_check() raises GroupLeakage if a group reaches two folds, and it runs before the split is returned. Folds come out uneven by construction, and fold_sizes reports that. Shipped: tsdive.eval, which carries the evaluation protocol as an API: ranking_metrics, clock_control, worst_baseline_threshold, fires, far_floor and the group holdout re-exported beside them.

Stage 3, provisional baselines

Shipped: MAD and moving-range screens, each object carrying its provisional caveat, with sparse or censored history raising InsufficientQuality rather than producing a wide interval. The benchmark row states why they are provisional: the regime-blind MAD misses a level shift that stage 4 catches.

Stage 4, refined baselines

Shipped: MODE-keyed per-regime median and MAD baselines. A regime with too few GOOD samples raises RegimeTooSparse. Benchmark: 100% regime screen detection against stage 3's 0% on the same shift.

Shipped: segment (changepoints/pelt.py), which supplies regimes when no MODE tag keys them, which is most tags. It runs PELT with the L2 cost over MAD-scaled values, so one penalty means the same thing on a flow and on a temperature. The penalty defaults to 3*log(n).

Shipped: cross-instance population baselines (baselines/population.py). population_baseline summarises every other training-fold window of the same (asset, variable). screen_population fires on a frozen window only when that pair holds more than one value in at least 95% of its training windows. A pair with fewer than MIN_BASELINE_WINDOWS training windows raises PopulationTooSparse.

Shipped: compare (compare.py), which reads two periods of one unit and reports what moved between them. Table 1 gives each tag the after period's quality read against the before period's, the level shift in before-period MAD-sigmas, the spread ratio, the share of after samples a before-period MAD screen flags, and the first changepoint at or after the after period starts. Table 2 gives each pair the Pearson correlation of first differences in each period and a moving block bootstrap interval on the change, and prints only the pairs whose interval excludes 0. Table 3 fits one PCA on the before period and reads it on the after one. A tag changed and a pair decoupled. Neither word names a cause.

Shipped: a plan that sets before and after runs compare over all its archives at once, as the last step.

Shipped: switchback plan and switchback analyze (src/tsdive/switchback/). plan draws a balanced random schedule of settings A and B over a window, writes it with a SHA-256 digest, and reads its power off a history window. analyze verifies the digest, the block times, the balance and the seed, raising ScheduleMismatch on any difference, and reports the B minus A difference in means with its randomization p-value and inverted 95% interval, plus an estimate adjusted on declared covariates. A design with too few blocks raises DesignTooSmall. Benchmark: the claim rate at a zero shift and the detection of a 0.5 sigma shift on seeded AR(1) records. On fresh assignments of 3W, TEP, Turbine Upgrade and SKAB records where nothing was changed, the claim rate at a zero shift is 3.9% to 5.3% per bed, where the first-half against second-half split of the same 3W records claims 65.2%. Design guidance: 0.8 detection at a 0.25 sigma shift was reached by schedules of 212 and 414 one-day blocks on the turbine pairs (0.922), and 16 blocks of 15 min on a 240-minute 3W record reach 0.113. Report: examples/studies/switchback/REPORT.md. Use: docs/SWITCHBACK.md.

Shipped: segment --mode-out FILE writes the segments as a MODE archive, one label (S1, S2, ...) per sample the window read, and screen --mode FILE builds one baseline per label from it.

Owed: mutual information between two tags, and per-tag attributions from a fitted model. Both wait for a benchmark row showing them find something the Spearman column and the SPE contributions do not.

Owed: a population baseline that generalises across assets. Under a held-out-asset split the current one has no training windows for the pair it is asked about and refuses (PopulationTooSparse).

Stage 5, SPC

Shipped: individuals limits from baseline statistics, xbar and R charts from the Shewhart constant tables, and a rules engine (BEYOND_3SIGMA, RUN_9_SAMESIDE, TREND_6) reporting each rule's hits separately. Limits come from baseline windows, never from the monitored data.

Stage 6, MSPC

Shipped: PCA-based Hotelling T2 and SPE over aligned multi-tag matrices. Alignment raises MspcAlignmentError on insufficient common-grid coverage instead of interpolating across tags. Control limits are empirical training percentiles, so they carry no distributional claim. SVD signs are fixed so models replay identically. The benchmark shows variance faults surfacing in SPE while T2 stays quiet. The two statistics are reported separately and never summed.

Stage 7, ML detectors and the viewer

Shipped: IsolationForest over physics-gated feature rows. Incomplete rows are excluded and counted rather than dropped, and random_state is pinned. Shipped: tsdive report-html, a self-contained static HTML evidence viewer with no server and no JavaScript dependencies. Studied, not shipped: a zero-shot time-series foundation model (Chronos-Bolt) as a benchmark subject on the onset-aligned 3W windows, under the same protocol as the detectors (examples/studies/3w_chronos, group tsfm).

Stage 8, LLM narration

Shipped as the evidence-ledger boundary. Narration consumes only the serializable EvidenceLedger (profiles, benchmark rows, refusals). The offline scripted narrator is what tests and CI use. A remote narrator raises NarratorUnavailable unless the operator sets TSDIVE_LLM_ENDPOINT. No endpoint, key or network call exists in this repository's default code paths. Narration produces advisory prose over cited evidence. It cannot write to a historian, because the store is read-only, and it cannot override a typed refusal. tsdive run fills the ledger by walking one plan over several archives. A step that raises for one archive writes a ledger row rather than propagating.

Deferred, pending a benchmark row and a test

• Granger causality, PCMCI and transfer entropy: no replayable causal benchmark exists.
• Batch machinery (EVT, batch SPC): every current tool is invalid on batch data, so batch needs its own physics layer first.
• Stiction detection and the Harris index: these need a control-loop schema at scale and a benchmark archive with known stiction cases.

Cut

Anything that needs write access to a historian. Anything with no refusal path.