Distributional-Assumption Card¶
Documentation — instantiates Distributional-Assumption Governance
A one-page record that pins a distributional commitment — modeled quantity, family, support, rationale, evidence, decision use, owner, and expiry — into an inspectable, hand-off-safe contract.
A distributional choice usually lives nowhere: it is a line of fitting code, a package default, a habit. Distributional-Assumption Card is the artifact that gives it an address. It is a short, structured record that states, in one place, what uncertain quantity is modeled and at what unit, which family and support and tail behavior are being asserted, why the process could plausibly generate that shape, what evidence backs it, for which decision it will be used, who owns it, and when it expires. The card does not test the assumption and does not decide whether it may be used — it makes the assumption a written thing that a later reviewer, a downstream team, or a future auditor can inspect and contest. Its whole value is turning an invisible default into a reviewable contract that travels with the model instead of being lost in the person who wrote it.
Example¶
A biostatistics team on a phase-III oncology trial models time-to-progression to feed a downstream health-economics group that will set a reimbursement case. Before the hand-off, they fill a card. Modeled quantity: progression-free survival, per randomized patient, over the trial horizon. Family and support: a Weibull on nonnegative time, with the shape parameter estimated (not fixed) and administrative censoring at data cut noted explicitly. Rationale: a monotone hazard is clinically plausible for this mechanism, and the Weibull's hazard shape is interpretable to clinicians. Evidence: prior trials in the indication, plus the study's own diagnostic profile. Decision use: the survival curve is extrapolated to feed a cost-effectiveness threshold — a use far more demanding than in-trial description. Owner: a named biostatistician. Revision trigger: re-open if enrollment demographics shift or if follow-up maturity falls below a stated level.
The outcome is not a better curve but a better hand-off. The health-economics group inherits a contract, not a spreadsheet column: they can see that the tail is being extrapolated well past observed follow-up, that this was a flagged limitation, and that a named person is accountable for it. Nothing is silently assumed.
How it works¶
- Fixed field set. The card is a template, so the same commitments are recorded every time and none can be quietly omitted: quantity and unit, family, support and boundaries, parameters, transformations, rationale, evidence, decision use, alternatives considered, limitations, owner, approvals, and expiry.
- Separate the kinds of claim. A good card distinguishes fixed structure (asserted, not estimated), estimated quantities, conventions adopted for convenience, and unresolved uncertainty — so a reader can tell an argued commitment from a default.
- Carry an owner and an expiry. Every card names who is accountable and states the condition under which it must be revisited, so it cannot harden into inherited truth.
- Travel with the model. The card is attached to the artifact and re-surfaces wherever the distribution is reused downstream.
Tuning parameters¶
- Depth — a one-paragraph note for a reversible low-stakes description, or a full dossier for a safety or reimbursement use. More depth costs authoring time; too little makes the card decorative.
- Field granularity — how finely support, tails, and dependence claims are itemized. Finer records surface hidden commitments but take longer to write and read.
- Expiry specificity — a vague "review periodically" versus a named, checkable trigger. Sharper triggers make expiry enforceable but require anticipating what would invalidate the model.
- Audience register — a technical appendix versus a decision-facing memo. Pick the one the actual downstream reader can act on.
When it helps, and when it misleads¶
Its strength is that it converts a distribution from a private default into a public, inspectable commitment that survives a hand-off — the same move as a machine-learning model card, which reports a model's intended use[1], data, and limitations so downstream users are not left to guess. It is the artifact that makes provenance and contestability possible at all.
Its failure mode is documentation theater: a card that exists but changes nothing. When a card is written after the decision, to justify a choice already made, it launders convenience into apparent rigor; when its exception status has no operational effect, an unsupported commitment still drives the action while the card sits beside it looking responsible. The card is a record, not an enforcer — it cannot by itself stop a bad assumption from being used. The discipline that keeps it honest is to write it before commitment, to keep the owner and expiry fields live, and to route the card through a checkpoint that can actually say no; a record without a gate behind it is a caption, not a control.
How it implements the components¶
random_quantity_estimand_and_unit_definition— the card's opening fields pin exactly what is modeled, at which unit, aggregation, and horizon, so the quantity cannot drift between fitting and decision use.distribution_family_commitment_register— the body of the card is the register: named family, support, parameters, tail and transformation claims, and which of these are fixed versus estimated versus convention.provenance_communication_equity_and_accountability— owner, approvals, limitations, affected groups, and downstream uses are stated in decision-facing language so the commitment is traceable and contestable.
The card records the commitment but never rules on it: it does not set the ex-ante bars for fit_for_use_acceptance_and_exception_thresholds or specify the robust_decision_fallback_and_escalation_path for when evidence is thin — pronouncing accept / limit / reject and mandating a fallback is the Independent Assumption-Challenge Gate.
Related¶
- Instantiates: Distributional-Assumption Governance — the card is the record that makes every later governance step have something inspectable to act on.
- Sibling mechanisms: Independent Assumption-Challenge Gate · Candidate-Family Comparison Grid · Distribution-Shift Trigger Dashboard · Support, Shape, and Tail Diagnostic Suite
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: The operative form is a one-page specification of the modeled quantity, distribution, support, rationale, evidence, decision use, owner, and expiry.
Nearest alternative: Record, Log & Register — Ownership and revision metadata aid handoff, but the card externalizes one current modeling commitment rather than accumulating a history of actual events or transitions.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Statistical modeling cohered explicit distributional commitments about family, support, dependence, parameters, evidence, and decision use.
Related originating lineages:
- Data Science & Analytics — Machine-learning model cards supplied the compact, handoff-safe documentation form.
- Ethics of Technology & AI Governance — Model governance supplied ownership, review, limitation, and expiry requirements.
Review resolution: Both current reviews place distributional_assumption_card primarily in statistics_experimental_design; the reconciled classification retains only lineages that materially shaped the mechanism and keeps breadth of origin separate from reach.
Attribution caveat: The documented object is statistical, while the card format and lifecycle controls come from model-governance practice.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; medium confidence.
Notes¶
The card and the Independent Assumption-Challenge Gate are deliberately split: the card is the artifact, the gate is the decision. Keeping them separate is what lets a team improve or correct a recorded commitment without re-litigating whether the model may ship, and lets the gate refuse a model without anyone having to reconstruct what was assumed.
Draft mechanism page for the Encyclopedia of Abstractions.
References¶
[1] Mitchell, Margaret, et al. "Model Cards for Model Reporting". Proceedings of the Conference on Fairness, Accountability, and Transparency, 220–229. ACM (2019). Proposes model cards that disclose a model's intended uses, data and evaluation context, and limitations so downstream users can judge appropriate use. registry ↩