RDS preserves exact types, item/evidence order, hashes, covariance, numerical modes and prediction data, including E3-MCMC draws. Reading never refits or migrates a posterior. Only open trusted RDS files. Unknown external Phase A provenance remains missing; computed payload hashes are always present.
Arguments
- state
A
start_link_session()result.- path
File path for an RDS session.
- overwrite
Whether to replace an existing file.
- input
Optional expected prepared input (or list of inputs) on load. It must match the saved input exactly. Use
resume_link_session()to append new evidence after this identity check.
Details
Hashes are verified using the current rlang engine. A failed check
with recorded link-v1-rlang provenance from before rlang 1.3.0 raises
pairwiseLLM_link_hash_incompatibility (also a linking-contract error), with
stored_hash_scheme, stored_hash_engine_version, and
current_hash_engine_version condition fields. Other mismatches remain
integrity/contract errors; a different version alone is not proof of
incompatibility. Missing engine provenance cannot establish the cause.
Keep the original artifact and hashes. Explicitly prepare new inputs from
the original saved judgments and compatible Phase A points/draws using
prepare_link_input(). This preparation needs no new LLM judgments or
sampling. A new standalone reference can reuse a saved completed fit through
prepare_linked_rubric_reference() only if its fit-time evidence identity
verifies. Otherwise investigate in its producing environment before deciding
separately whether a refit is needed. Creating a new E3 result itself refits
the saved outcomes; MCMC remains an explicit, potentially expensive choice.
Loading/validation never rebuilds, migrates hashes, overwrites artifacts,
contacts providers, or starts sampling.
See also
prepare_link_input(), fit_link(), start_link_session(), save_link_session()
Other linking:
fit_link(),
predict_link(),
prepare_link_input(),
start_link_session(),
summary.pairwiseLLM_link_result()