Model cards and evidence levels
Distinguish a complete specification from fitted parameters, independent model evaluation and a deployed subscriber route.
A card identifies the decision object
A useful model card describes intended inputs, availability, target, transformations and action before it displays a result. It identifies a deterministic rule or an actual fitted estimator, not merely an algorithm family. The source contract should say what happens when a required field is late or absent. The target should say when a label resolves and which rows own that outcome. The action should name its threshold or rule and permitted exceptions. Version identity ties these declarations together. A filled card can make the design inspectable without implying that anyone trained, evaluated or deployed the design.
Separate levels instead of using one badge
Specification completeness answers whether the process is described. Fitting evidence answers which parameters were actually learned under which scope. Independent economic evaluation answers whether the frozen financial system met its declared requirements. Operational evidence identifies what artifact actually serves the subscriber. These layers can disagree legitimately. A fitted candidate can fail evaluation; a completed research master can reject all candidates; a documented research success can still lack production promotion. Each state needs its own artifact rather than one green status. A model nickname or a file labelled LIVE does not answer all four questions.
Missing fields should retain their meaning
Use absent or unresolved for records that do not exist. Do not create plausible source hashes, reviewer names or evaluation dates to make a template look finished. The educational cards deliberately include null actual parameters, master, deployed source identity and financial outcomes. Their anonymous observations are authored rather than pulled from market history. This is a complete document about an incomplete real-evidence situation. If genuine records arrive later, retain the teaching classification and create a new identified research artifact. Silently turning the fiction into an alleged historic experiment would defeat the purpose of the example.
A comparison requires compatible scopes
Two cards can be compared for documentation coverage without ranking their financial quality. Ask which field is supported by a primary source and what remains unverified. A rule with no fitted weights can still have a large selection history. A classifier with many parameters can still use leaked features. Complexity does not order the evidence levels. The release workbench helps retain references to each question, but manually selecting documented does not authenticate the reference. Real instrument research still requires its full authoritative factory and independent master rather than a card-only completion shortcut.
Fictional model evidence case
The fictional prior-reference card has a fully written availability and action contract, but no genuine instrument source, master or deployed route. The classifier card adds a proposed label and training boundary while leaving actual learned parameters absent. The successor card changes an input contract and cannot inherit a validation badge from either. None claims a live or profitable system. The comparison is about which document would answer each question, not an evaluation of market outcomes. Its teaching IDs are not source-control hashes.
| Object | Record | Bounded finding |
|---|---|---|
| Specification | Written teaching card | Document completeness only |
| Training | Actual parameters absent | No fitted candidate claim |
| Master | No independent artifact | No financial completion |
| Release | No subscriber route | No live-model inference |

Primary references and scope
Source review: 14 October 2026. Official mechanism descriptions are not this desk's paid-account test, independently captured fills or vendor-return audit.