Migrating to causalrl 2.0¶

causalrl 2.0 is the first release with a breaking change. It is small, mechanical, and was announced with a FutureWarning since 1.5. This page tells you exactly what changed and how to update.

The one breaking change: certificates by default¶

Three routines that had both a legacy return type and a certificate-returning variant now return a Certificate by default:

Routine	1.x default	2.0 default
`identify_effect`	`Estimand`	`Certificate` (`IDENTIFIED`)
`ipw_sensitivity_bounds`	`Interval`	`Certificate` (`BOUNDED`)
`msm_policy_value_bounds`	`Interval`	`Certificate` (`BOUNDED`)

Everything else in the public API is unchanged: no other function changed its return type, and no name was removed.

If you want the certificate (recommended)¶

Do nothing — the default now gives you a unified, serializable Certificate that carries the claim, the epistemic kind, the numeric value, the assumptions consumed, and provenance:

from causalrl import ipw_sensitivity_bounds

cert = ipw_sensitivity_bounds(outcomes, propensities, gamma=1.5)  # -> BOUNDED Certificate
print(cert.kind, cert.value)          # Kind.BOUNDED  Interval(lower=..., upper=...)
print(cert.to_json())                 # serialize the whole claim

If you want the legacy value¶

Pass return_certificate=False. This is the exact 1.x behaviour and it is not deprecated — it is a supported escape hatch that returns the raw Estimand / Interval:

interval = ipw_sensitivity_bounds(outcomes, propensities, gamma=1.5, return_certificate=False)
lo, hi = interval                     # unchanged 1.x Interval

A mechanical migration is: find bare calls whose result you index, unpack, or compare as an Interval / Estimand, and add return_certificate=False. The certificate wraps exactly the same numerics (a byte-stability test pins this), so nothing else needs to change.

Removed¶

The pre-2.0 FutureWarning and the private causalrl._deprecation helper are gone (the flip they announced has happened).

Error behaviour is unchanged: invalid inputs still raise the same exceptions on the default path.

What else is new in 2.0¶

2.0 is mostly additive — the breaking change above is the only removal. Highlights across the 1.3 → 2.0 line:

Unified certificates (Certificate, Kind ∈ IDENTIFIED / BOUNDED / EMPIRICAL) that the shipped DecisionCertificate / PivotalityCertificate / TransportRegretCertificate conform to via as_certificate.
Continuous causal core: doubly-robust certify_effect, sharp estimated-propensity and heavy-tailed bounds, conformal wrappers, neural / flow mechanisms with exact abduction, sequential DR, and transport estimation.
Multi-agent core (causalrl.magames): typed populations, certify_equilibrium, per-agent CausalEnv views, and a duck-typed PettingZoo adapter.
Scale & data plane: streaming certificate kernels (stream_policy_value, stream_msm_bounds, stream_quantile_certificate) over a columnar TrajectoryLog — in memory or streamed from Parquet without materialising it — plus an optional JAX backend.
Interop & scale: an SBI/NumPyro posterior → Regime bridge (regimes_from_posterior, across_regimes), a generic columnar-simulator adapter (ColumnarSimulator), and a deepened d3rlpy path (TrajectoryLog ↔ MDPDataset both ways, certify_fqe, policy_actions).

See the task guides for runnable, end-to-end walkthroughs, and the assumption glossary for what each certificate's assumptions mean.