Exemplar Feedback Registry¶
Feedback ledger — instantiates Nearest-Exemplar Response Reuse
Logs what happened every time an exemplar was reused and uses those outcomes to broaden, narrow, or retire each stored case's authority — so the case memory sharpens instead of fossilizing.
An Exemplar Feedback Registry is the memory-maintenance backend of the archetype. It does no matching and answers no query; it sits behind every reuse and records what the reuse produced. Each time a stored case guides a new decision, the registry logs the outcome, and it rolls those outcomes up into a running authority for each exemplar: a case whose reuses keep succeeding earns a broader mandate, one whose reuses keep failing has its authority narrowed or is retired, and its quality label moves with the evidence. Layered on top is a guardrail deciding which cases may be retained and allowed to gain weight at all, so that a harmful or tainted exemplar cannot accumulate authority simply by being reused. Its distinctive contribution is the write-back: turning the archetype's loop into something that closes, so history is a curated asset rather than an accreting pile.
Example¶
A content-moderation platform reuses past decisions as exemplars — one canonical case, say, established that a particular reworked meme counts as protected satire and is left up. Every time that exemplar guides a new moderation call, the registry logs the pairing and, later, the outcome: was the decision upheld on appeal, reversed, or did it lead to a documented harm?
Over a quarter the signals accumulate. The satire exemplar is upheld again and again across contexts, so the registry broadens its authority and raises its quality label — moderators can lean on it more widely. A second exemplar about "borderline harassment" starts getting reversed on appeal at a rising rate; the registry narrows its authority to the single language and region where it still holds and downgrades its quality label, flagging it for review. A third exemplar turns out to have rested on a proxy for a protected class; the bias guardrail purges it from the store and blocks it from ever regaining weight, no matter how "similar" future cases look. None of this touched the matcher — but the pool of cases the matcher draws from is measurably sharper than it was three months earlier.
How it works¶
- Log every reuse. On each reuse, record the exemplar used, the query it guided, the response, any adaptation, and — when it returns — the outcome.
- Roll outcomes into authority. Aggregate per-exemplar success, reversal, and harm signals into an authority level and a quality label; broaden, narrow, or retire the exemplar accordingly.
- Guard retention and weighting. Screen which cases may stay in the store and gain authority, purging tainted or privacy-violating exemplars regardless of their apparent usefulness.
- Surface the decaying. Flag exemplars whose outcomes are drifting or whose freshness has lapsed for review, so staleness is caught before the next reuse.
Tuning parameters¶
- Update aggressiveness — how fast authority moves per outcome. Fast updates track real change but overreact to noise and small samples; slow updates are stable but let a decaying exemplar keep binding too long.
- Harm-vs-success weighting — how heavily a harmful reuse counts against an exemplar relative to a successful one. Asymmetric weighting protects against bad-case amplification but can retire useful cases on a single bad break.
- Retirement threshold — the authority floor at which an exemplar is pulled from reuse.
- Decay half-life — how quickly an un-refreshed outcome loses weight, so old evidence does not prop up a stale case forever.
- Guardrail strictness — how aggressively suspect or privacy-sensitive cases are purged versus retained with restricted authority.
When it helps, and when it misleads¶
Its strength is that it is the component that keeps a case memory calibrated: it catches drift and staleness, and it is the specific defense against bad-case amplification — the unverified or harmful response that gets reused again and again precisely because it is stored. Without a registry, every other mechanism in the archetype is retrieving from a pool that only ever grows and never learns.
Its own failure mode is the feedback loop: reused decisions generate the very outcomes that appear to confirm them, so an exemplar can look successful only because its reuse foreclosed the alternative that would have exposed it.[n1] Outcome signals are often delayed, sparse, or absent entirely, and updating authority on thin or gamed evidence can retire good cases and entrench lucky ones. The classic misuse is treating short-term proxy metrics as ground-truth outcomes and letting them drive authority. The guarding discipline is to separate outcome capture from the reuse decision, to require a real outcome trace before moving authority (holding judgment where none exists rather than inventing one), and to watch explicitly for self-reinforcing loops where an exemplar's success is an artifact of its own reuse.
How it implements the components¶
outcome_feedback_update_record— its core artifact: the per-reuse ledger of what happened, and the authority updates it drives.case_quality_label— recomputes each exemplar's quality and authority from accumulated outcomes, moving it up or down as evidence lands.privacy_and_bias_guardrail— screens which cases may be retained and gain weight, purging tainted or privacy-violating exemplars from the store.
It does not encode a query_case_profile (Case-Based Reasoning System, Expert Case Recall Checklist), compute a similarity_metric (Similarity Search over Case Embeddings), or run a nearest_exemplar_selection_rule (K-Nearest-Neighbor Case Matcher, Incident Playbook Lookup) — it maintains the store after the fact rather than matching against it. Where Precedent Matching Workflow also runs a bias guardrail, it applies it in the moment of decision and distinguishes cases with a countercase_and_exception_probe; this registry instead updates exemplar authority from accumulated outcomes over time.
Related¶
- Instantiates: Nearest-Exemplar Response Reuse — this is the archetype's write-back-and-govern loop, the component that keeps the memory from fossilizing.
- Sibling mechanisms: Case-Based Reasoning System · K-Nearest-Neighbor Case Matcher · Similarity Search over Case Embeddings · Precedent Matching Workflow · Incident Playbook Lookup · Expert Case Recall Checklist · Case Similarity Rubric
Editorial Notes¶
Form Classification¶
Form family: Record, Log & Register
Rationale: Exemplar Feedback Registry operates as a durable record, ledger, register, or trace whose value depends on preserving actual state or history because it logs what happened every time an exemplar was reused and uses those outcomes to broaden, narrow, or retire each stored case's authority — so the case memory sharpens instead of fossilizing.
Independent corroboration: The frozen evidence defines Exemplar Feedback Registry as 'Logs what happened every time an exemplar was reused and uses those outcomes to broaden, narrow, or retire each stored case's authority — so the case memory sharpens instead of fossilizing', so its operative form is Record, Log & Register.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Computer Science & Software Engineering
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Case-based reasoning formalized retaining prior cases, reusing them, testing outcomes, revising failed adaptations, and maintaining the case library from feedback.
Related originating lineages:
- Education & Pedagogy — Exemplar-based learning materially supplies the authority and calibration role of stored cases.
- Law & Governance — Precedent maintenance, distinguishing, and overruling materially shape broadening, narrowing, and retiring case authority.
- Organizational & Management Science — Maintaining and revising a reusable case repository from outcome feedback follows organizational knowledge-management practice.
- Ethics of Technology & AI Governance — Bias, privacy, and harmful-example retention safeguards materially shape governance of the exemplar store.
Review resolution: AAAI identifies retrieval, reuse, revision, and retention as the core CBR cycle. Organizational learning, precedent, and exemplar governance materially shape the registry's authority and retirement controls.
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¶
The registry is consumed by the other mechanisms rather than consuming them: matchers and reasoning systems retrieve from the store it curates, and it earns its keep only if outcomes actually return to it — an archetype whose loop never closes degrades into an ever-growing, never-corrected memory.
[n1] A feedback loop here means that acting on a decision changes the world in ways that then look like evidence for the decision — reusing an exemplar can suppress the counter-evidence that would have revealed it as wrong. It is why an exemplar's apparent success rate can be inflated by the very reuse the registry is trying to evaluate. ↩