Skip to content

Positive / Negative Example Deck

Training artifact — instantiates Prototype-Centered Category Modeling

A curated deck of clearly-labeled positive and negative examples — each carrying its rationale and the action it triggers — that installs a category's center and purpose in a new judge.

A Positive / Negative Example Deck is a teaching artifact: a curated, labeled collection of clear category members and clear nonmembers, organized to onboard a new judge into what the category is for and where its unambiguous cases lie. Its distinctive move is that it leads with the center and the purpose, not the boundary. It chooses examples people already agree on — obvious positives, obvious negatives — and attaches to each the rationale and the action it triggers, so the learner absorbs not just a label but why the category exists and what to do about it. The contested knife-edge is deliberately left to a near-miss set; the deck's job is to plant a shared, purposeful prototype.

Example

A platform is onboarding new content moderators to a "hate speech" policy. The deck opens with clear positives — unambiguous slur-based attacks, each card citing the exact policy clause and the action remove — and clear negatives — heated political argument, reclaimed in-group usage, a slur quoted inside straight news reporting — each carrying the action allow or escalate. Every card teaches the why and the do, so a trainee learns that the category is not "offensive language" but "targeted attack on a protected class," and that the point of the category is a moderation action, not a moral verdict. The deck deliberately includes examples that don't fit the naïve picture — harassment that uses no slurs at all, offensive-sounding speech that policy nonetheless permits — so the prototype the trainee forms is the policy's, not a stereotype of what hate speech "looks like."

How it works

  • Anchor on agreement, not the edge. Choose examples with near-universal consensus to establish the category's core and its purpose; hand the contested boundary to a near-miss set.
  • Every card carries label + rationale + action. The deck teaches the action context — what the category is for — not just the class name, so judges know what to do once they classify.
  • Keep it as a living reference. The labeled deck persists as the shared library judges return to and calibrate against, long after onboarding.
  • Sample deliberately. Spread cards across the recognized varieties of member and nonmember so the taught center is representative rather than a single stereotype.

Tuning parameters

  • Clarity threshold — how unambiguous a case must be to earn a place. A high bar keeps the deck a clean reference but leaves the boundary untaught (delegated to a near-miss set); a low bar smuggles in contested cases that will provoke disagreement.
  • Rationale depth — a bare label versus a full policy citation plus worked reasoning. Depth teaches transfer to new cases but bloats the deck.
  • Action coupling — whether each card names the downstream action or only the class. Coupling prevents "I can label it but don't know what to do."
  • Coverage spread — how widely cards sample the varieties of member/nonmember; wider spread resists a monoculture prototype.
  • Size and refresh — how many cards, and how often the deck is renewed from fresh decisions before it goes stale.

When it helps, and when it misleads

Its strength is that it is the fastest way to install a shared, purposeful category in new judges: the labeled set doubles as the reference everyone calibrates against, and coupling each card to an action prevents classification that goes nowhere. Its failure mode is that a deck freezes one curator's view — its blind spots quietly become everyone's, and it drifts out of date as the phenomenon changes while the deck stands still. An over-clean deck leaves judges helpless at the very boundary it never showed them. The classic misuse is stacking the deck — choosing only flattering examples to manufacture consensus for a contested policy. The discipline that guards against this is to pair the deck with a near-miss set for the edge, to refresh it from real decisions, and to track inter-judge agreement so a deck teaching the wrong prototype is caught early. Studied worked examples like these can teach a category faster than unguided practice — but only if the examples are the right ones.[1]

How it implements the components

Positive / Negative Example Deck fills the teaching-and-reference components a training artifact can hold:

  • category_purpose_and_action_context — each card ties the class to why the category exists and the action it triggers; the deck's spine is purpose, not just label.
  • calibration_case_library — the persistent labeled deck is the reference library judges calibrate against and return to.

It does not build the minimal-pair boundary cases — those are the Near-Miss Comparison Set — and it does not empirically rank the anchors or elicit the graded scale, which is the Typicality Rating Exercise. Its representation spread helps guard against a stereotyped prototype, but the formal bias_and_stereotype_guardrail lives in the Similarity Dimension Rubric.

Notes

The deck's authority is entirely curatorial, so who curates it and how it is revised matter more than any single card. It depends on a drift signal to know when to refresh: a deck is only as current as the decisions it was last rebuilt from.

References

[1] The worked-example effect from cognitive-load research: studying fully worked examples can produce faster, more durable learning than unguided problem-solving early in skill acquisition — provided the examples chosen are representative of what the learner must generalize to. ↩