[M2 TRACKER] Reproducible Research Loop v0.1

[andreazedda]

Milestone objective

Demonstrate one end-to-end, lineage-bound research loop on the pseudonymized longitudinal case without overclaiming clinical or causal validity.

Primary work item

• [P1][RESEARCH] Close the lineage-bound real-patient Research Loop v0.1 #18 Close the lineage-bound real-patient Research Loop v0.1

Required prerequisites

• [P0][CANONICALIZATION] Publish the validated Twin-lineage candidate and migrate the research database safely #11 Twin-lineage candidate canonical and DB migrated
• [P0][GOVERNANCE] Establish canonical intended use and remove prescriptive legacy outputs #12 intended-use contract active
• [P0][REPRODUCIBILITY] Lock dependencies and standardize Python quality tooling #15 deterministic dependency environment
• [P1][REPRODUCIBILITY] Implement immutable run manifests, model registry, and typed scientific contracts #19 run manifest/model registry available at the level required by the loop

Exit definition

BASELINE_INPUT_FROZEN = true
BASELINE_TWIN_LINEAGE = valid
PRE_CALIBRATION_RESIDUALS = complete
CALIBRATION_ACCEPTED_OR_EXPLICITLY_REJECTED = true
POST_CALIBRATION_LINEAGE = valid
EXPOSURE_AWARE_WHAT_IF = complete
TEMPORAL_BACKTEST = complete
RUN_MANIFEST = complete
PHI_IN_GIT = 0
CLINICAL_SUPERIORITY_CLAIM = false

Evidence required

• computational-input manifest and hashes;
• parent/child Twin lineage evidence;
• pre/post residual report;
• calibration acceptance/rejection analysis;
• schedule-resolved what-if results;
• leakage-free temporal backtest;
• uncertainty/sensitivity findings;
• final research report with epistemic labels;
• reproducibility instructions.

Failure is an allowed milestone result

This milestone can close with a scientifically useful negative result such as:

INSUFFICIENT_DATA
CALIBRATION_FAILED
MODEL_NOT_IDENTIFIABLE
BACKTEST_UNACCEPTABLE

provided the failure is reproducible, explained and not converted into a positive therapeutic claim.