Distortion Model Card¶
Model card — instantiates Source Distortion Modeling
Documents the assumed distortion pattern, supporting evidence, scope, counterevidence, and expiry conditions.
The Distortion Model Card is a short, standardized document that states one assumed distortion pattern for a source and pins it in place: the transformation believed to act on the account, the evidence for that belief, the scope over which it is assumed to hold, the counterevidence that would break it, and — its signature field — the expiry conditions that retire it. Its defining move is that it gives a distortion hypothesis a shelf-life. The card is not a verdict and not a plausibility test; it is a dated, falsifiable assumption written down so that everyone downstream applies the same correction and knows exactly when to stop trusting it.
Example¶
A hydrology team relies on a satellite soil-moisture product. They notice it reads systematically too dry for roughly two days after heavy rain — the retrieval algorithm mishandles standing surface water. Instead of trusting the feed or discarding it, they write a Distortion Model Card. Distortion: a negative bias in the first 48 hours after precipitation, larger in densely vegetated cells. Evidence: comparison against ground stations over the last two wet seasons (figures illustrative). Scope: vegetated basins only, not arid ones. Counterevidence that would void it: the bias disappearing in fresh ground comparisons, or a vendor firmware change. Expiry: re-validate each wet season, or immediately if the vendor ships a new retrieval version. Downstream models apply the documented correction only inside the card's scope, and a calendar trigger forces re-validation. A year later the vendor updates the algorithm; the expiry condition fires and the card is retired rather than silently mis-applied to a sensor that no longer behaves that way.
How it works¶
- One card, one distortion. Each card documents a single hypothesized transformation, not a general reputation for a source.
- Record scope explicitly — where the pattern is assumed to hold and where it is not — so the correction is not over-applied.
- Name the breakers. List the counterevidence that would falsify the card, up front and in writing.
- Set an expiry. Every card carries revision or escalation conditions — a date, an event, a drift threshold — that force re-examination; a card with no expiry is a defect.
Tuning parameters¶
- Scope tightness — narrow (one window, one regime) versus broad. Narrow cards are safe to apply but proliferate; broad cards are convenient but over-reach.
- Expiry aggressiveness — how soon or easily the card retires. Short expiries stay honest but churn; long ones risk applying a stale correction.
- Evidence bar — how much support a card needs before it is trusted enough to drive corrections rather than mere caution flags.
- Correction form — whether the card prescribes a numeric adjustment, a downgrade, or only a "treat with caution" marker.
When it helps, and when it misleads¶
Its strength is that it turns tacit "everyone knows that sensor runs hot" folklore into an explicit, scoped, expiring artifact that downstream users apply consistently and can audit later. The format borrows directly from model cards in machine learning — the practice of shipping a model with its intended use, evaluation conditions, and known limitations rather than as a black box.[n1]
Its failure mode is calcification: the assumed distortion becomes a permanent "fact" applied long after the source changed, especially when nobody honors the expiry. A confidently written card can also over-correct, manufacturing a clean signal that is really the correction's own artifact. The classic misuse is writing a card to justify discarding inconvenient data — "that feed is known-biased" — without the evidence field ever earning it. The guarding discipline is no card without an evidence basis and an expiry, and treating a fired expiry as a mandate to re-derive rather than renew by habit.
How it implements the components¶
distortion_hypothesis_set— the card is the durable home of one selected distortion hypothesis: the transformation, its evidence, and its scope, written down and citable.revision_and_escalation_trigger— the expiry field is a revision trigger: dated or event-based conditions that retire or re-open the card.
It records an assumed distortion but does not test whether that distortion is plausible from the source's incentives (motive_constraint_and_stakes_profile) — that adjudication is the Motive-Opportunity-Bias Analysis — and it does not rank the independent evidence (corroboration_and_countertrace_set) it cites — that is the Corroboration Ladder. The card documents and dates; its siblings test and rank.
Related¶
- Instantiates: Source Distortion Modeling — the archetype's durable record of one hypothesized transformation.
- Consumes: Motive-Opportunity-Bias Analysis — the plausibility verdict the card records in its evidence field.
- Sibling mechanisms: Narrator Reliability Matrix · Account/Event Reconstruction Table · Vantage-Bias Interview Protocol · Corroboration Ladder · Contradiction Timeline · Motive-Opportunity-Bias Analysis · Claim Release Gate
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Distortion Model Card operates as a non-executable information artifact that externalizes static or prospective structure because it documents the assumed distortion pattern, supporting evidence, scope, counterevidence, and expiry conditions.
Independent corroboration: The frozen evidence defines Distortion Model Card as 'Documents the assumed distortion pattern, supporting evidence, scope, counterevidence, and expiry conditions', so its operative form is Representation, Specification & Plan.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Ethics of Technology & AI Governance
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Responsible-ML practice established model cards documenting intended use, evaluation conditions, limitations, and operating envelopes for downstream users.
Related originating lineages:
- Data Science & Analytics — Machine-learning practice supplied the model artifact and its performance evidence.
Review resolution: The original Model Cards publication frames the artifact as transparent and responsible reporting, making AI governance primary; this entry adapts it to source distortion.
Attribution caveat: The artifact arose inside machine learning, but its distinctive institutional purpose is responsible-technology disclosure and accountability.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
Notes¶
[n1] In machine learning, a model card (proposed by Mitchell and colleagues) ships a model together with its intended use, evaluation conditions, and known limitations, so downstream users apply it within a documented envelope rather than blindly; the distortion model card adapts the same discipline to a distorted source. ↩