Roadmap

The goal is a library whose results are correct by default and easy to understand. Two rounds of gap screening fixed about 90 correctness bugs, and most of them fell into six recurring classes: counter resets, interval closing, logging gaps, sample-count means, time zones and per-signal grouping. Because the same classes keep recurring, the fix is in the architecture rather than in individual detectors.

W1: signal semantics core (started)

SignalSpec, infer_specs, profile() and the spec-driven primitives in ts_shape.signals. Detectors call the primitives instead of re-deriving signal semantics. See Signal semantics and profiling.

Consistency audit (done, October 2026)

An audit of wording, names, outputs and behavior. The rules are in Naming and output conventions, and tests enforce them:

• every documented call binds against the real signature;
• detector outputs follow their declared shape, and empty results keep every column;
• one name per concept, checked by a lint over every signature, with deprecated aliases for old names;
• every public parameter is documented;
• inputs are checked the same way in every pack.

Also done in October 2026: full precision, value columns checked, best_and_worst_shifts returns a DataFrame, batch_transition_matrix returns a long table.

Open, all additive (new accepted forms, nothing removed):

• Run state in one form. DowntimeEvents reads string states ("Running") or booleans, MachineStateEvents and OEEEvents read booleans. Accept either everywhere through SignalSpec.
• One place for the signal uuid. Some detectors take it in the constructor, others per method.
• Durations in one form. OEEEvents takes ideal_cycle_time in float seconds, takt_time takes an offset string. Accept both everywhere.

In a future major release the deprecated aliases (warning since 0.0.63) are removed.

W2: 1.0, correct by default, one API, one output

• Done (October 2026): the correctness flags (handle_resets, interval_end, time_weighted, group_by_uuid, slope_per, ...) default to the correct behavior. max_gap stays None (no cap), because no single value fits every signal; set it per signal or through a SignalSpec.
• Done: no silent data loss. Skipped inputs warn with DataQualityWarning; loaders take errors="warn" | "raise".
• Move the detectors pack by pack onto the primitives, with defaults taken from the spec.
• One enforced output contract (point, interval, summary) and one keyword style for every detector.
• KPIs carry their basis (run seconds, planned seconds, excluded gaps, resets), so every number can be audited.

W3: proven correctness (harness started)

• A registry-wide invariant harness (tests/invariants) runs every detector method on empty input, on tz-aware input, with an unrelated extra signal, and with nullable, pyarrow and text-number dtypes (same result required). It is pinned to a baseline: 278 methods, 0 failures.
• Property-based tests (hypothesis) cover the primitives, counters, intervals and window bounds.
• Reference tests, done: AIAG MSA Gauge R&R (average-and-range method, K1/K2/K3 and the EV²/(n·r) correction of AV); ISO 50001 / 50006 EnPI; ISO 22400 OEE and MTBF / MTTR; Western Electric / Nelson SPC rules 1 to 8; Cp, Cpk, Pp, Ppk, Cpm (Montgomery, Introduction to SQC); GUM measurement uncertainty (JCGM 100 §4.4.3); Weibull MLE (Lieblein–Zelen ball bearings, Lawless Ex. 5.2.2).
• The docs are tested: every ts_shape name in a code block must import, and CI builds the docs with mkdocs build --strict.

W4: scale and lean install

• Done: the core install needs only pandas, numpy and scipy; cloud and SQL loaders are extras.
• Done: slow row loops vectorized (shift assignment, maintenance trend fits, routing, warm-up, stockout; 10 to 300 times faster on 200k rows). PELT change points are still quadratic on data without change points (about 6 s on 200k rows). A scale test guards the fixed methods.
• After 1.0: backend-agnostic primitives (Polars, DuckDB), TimescaleDB push-down, published benchmarks.

W5: understanding

• profile() as the entry point.
• A question-first cookbook ("Why did OEE drop?", "Which station is the bottleneck?", "Is my gauge capable?") on a bundled, realistic plant dataset.
• A generated detector reference: assumptions, input signal kinds and output shape.

W6: ecosystem

• Grafana and TimescaleDB recipes.
• A command line (ts-shape profile data.parquet).
• Done: deprecation policy. Renamed names keep working with a warning and are removed only in a major release.

Process analytics

Added: change-point detection (ChangePointEvents: exact PELT and binary segmentation), multivariate SPC (MultivariateSPCEvents: Hotelling T² with Phase I and II limits and contributions), reliability (ReliabilityEvents: MTBF and MTTR per ISO 22400, censored Weibull life model).

Candidates:

Feature	Value	Effort
Vibration: FFT spectral band energy	high	M
Tool wear vs. measurement drift; first-article and setup verification; part disposition	high	M each
MQTT or Kafka streaming loader	medium-high	M
Detector evaluation (precision and recall against labels)	medium	M
Scaling transforms	medium	S
Energy and production plot helpers in an optional [viz] extra	medium	S

Unmerged branches

Two large branches from earlier work were never merged. Both can be ported in their own pull requests after reconciling them with the core:

• claude/reporting-python-library-2fhu1a: ts_shape.reporting (8.7k lines, 211 tests): facility calendar with IANA time zones, weighted KPI roll-up, ISA-95 hierarchy, benchmarking, scorecard and variance bridge, FX normalization, reporting coverage. To reconcile: build the calendar on signals.windows / shift_windows (already DST-correct) instead of a second implementation; drop its docs CI job (the strict docs build is in ci.yml); keep SECURITY.md and the issue templates.
• claude/global-production-reporting-w5f10n: FleetProductionReport and FleetSpec (11k lines): machine-status adapters (PackML, stack light, code maps), counter wrap vs. reset, part-done pulses, cycle-time unit inference, production context with late binding. To reconcile: map the status adapters onto SignalSpec(kind="state") and OperatingStateEvents; add counter wrap as a reset_policy ("wrap") in SignalSpec and utils.counters instead of a parallel counter model; fold cycle-time unit inference into infer_specs.