Proxy Retirement Decision Record¶
Decision record — instantiates Proxy–Target Divergence Detection and Recalibration
Documents, with rationale and a named owner, the decision to downgrade, recalibrate, replace, or retire a proxy — and what claims must change as a result.
Proxy Retirement Decision Record is the loop's governance artifact — the durable written act that turns "the evidence says this proxy has decoupled" into an executed, accountable decision that the organization can later inspect. Its defining move is documentation-as-commitment: it records which path was chosen (recalibrate, supplement, replace, or retire), why, which claims the proxy may no longer support, and — most importantly — who now owns the consequences. It does not detect divergence or measure anything; it consumes that evidence and closes the loop by making the response official, reversible only by another record, and impossible to quietly forget. Without it, divergence is documented in a report, discussed in a meeting, and then ignored while the dashboard rolls on.
Example¶
For a decade, a large search engine has used raw click-through rate as a headline proxy for result relevance: more clicks, better results. Accumulated evidence — audits, complaints, an incentive review — finally establishes that clicks now track curiosity-gap and clickbait pull more than genuine usefulness, especially on health and news queries. The relevance team writes a retirement decision record. It states the finding, cites the fidelity evidence, and selects the path: click-through is not retired outright but downgraded — demoted from a headline relevance metric to one weak input among several, with dwell time and satisfaction surveys added. The record then does the two things a meeting cannot: it narrows the claim scope in writing ("click-through no longer certifies relevance for YMYL queries; do not cite it as such in launch reviews"), and it names an accountable owner responsible for the new blend and for the review date twelve months out. Anyone who later reaches for click-through as proof of relevance is met with the record explaining why they can't.
How it works¶
- Anchor to the triggering evidence. Cite the specific divergence findings — audit results, drift flags, incentive analysis — that justify acting, so the decision rests on record rather than opinion.
- Choose and state the path. Select recalibrate, supplement, split-by-context, replace, or retire, and say explicitly what changes in the pipeline as a result.
- Rewrite the claim scope. Enumerate what the proxy may still be used to support and, in plain language, what it may no longer support — the demotions are the point.
- Assign the owner and review date. Name a person or role accountable for the change and its consequences, and set the date the decision itself must be revisited.
- Publish and make binding. Put the record where downstream consumers of the proxy will encounter it, so the narrowed scope travels with the metric.
Tuning parameters¶
- Action strength — the choice along downgrade → recalibrate → replace → retire. Stronger actions cut decoupling risk but destroy continuity and sunk investment; weaker ones preserve continuity but may under-respond.
- Reversibility — whether the record includes a defined path back if the proxy recovers, or is a one-way retirement. Reversible records lower the cost of acting early.
- Scope of the downgrade — global versus context-specific ("invalid for these queries only"). Surgical downgrades preserve the proxy where it still works.
- Bindingness — advisory note versus enforced gate in launch/compliance workflows. More binding records are harder to ignore but slower to issue.
- Review horizon — how far out the mandatory re-examination is set; shorter horizons guard against premature retirement of a proxy that recovers.
When it helps, and when it misleads¶
Its strength is that it defeats institutional inertia and amnesia. By naming an owner and rewriting the claim scope in a durable place, it stops the most common loop failure — divergence noticed, then quietly ignored — and it leaves an auditable trail explaining why a long-trusted number was demoted, which protects the decision when it is later questioned. The convention echoes the architecture decision record, whose whole value is preserving the reasoning behind a choice for the people who inherit it.[n1]
Its failure mode is that a record is only as good as its enforcement and its timing. A beautifully written record that no workflow actually consults changes nothing — dashboard theater with better prose. And because retirement is costly and visible, the record is prone to over-retirement: killing a proxy that was merely stressed and would have recovered, or that still worked fine outside one context. The classic misuse is writing the record to rationalize a decision already made for other reasons rather than to follow the evidence. The guarding discipline is to tie the record to a binding workflow gate, to prefer scoped downgrades over wholesale retirement, and to set a mandatory review date so both the decision and its reversal stay live.
How it implements the components¶
recalibration_or_replacement_path— the record is the executed path decision: it names whether the proxy is recalibrated, supplemented, replaced, or retired and what changes downstream.claim_scope_downgrade— it rewrites, in durable form, what the proxy may and may not be used to claim from now on, so unsupported interpretations cannot survive the decay.accountable_proxy_owner— it assigns a named owner responsible for the change, its consequences, and the next review, giving the proxy's fate a locus of accountability.
This record documents and executes a decision; it does not detect the break or fire the trigger that prompts it — the divergence_sentinel and decoupling_threshold_rule are supplied by Drift and Change-Point Detection, whose evidence this record consumes.
Related¶
- Instantiates: Proxy–Target Divergence Detection and Recalibration — it is the loop's closing act, turning divergence evidence into an accountable, recorded response.
- Consumes: Drift and Change-Point Detection and Holdout Ground-Truth Audit supply the evidence the decision rests on.
- Sibling mechanisms: Drift and Change-Point Detection · Holdout Ground-Truth Audit · Incentive Impact Review · Proxy–Target Correlation Refresh · Reference-Standard Recalibration Review · Sentinel Outcome Dashboard · Shadow Target Measurement · Triangulated Proxy Panel
Editorial Notes¶
Form Classification¶
Form family: Record, Log & Register
Rationale: Proxy Retirement Decision Record operates by preserves the triggering evidence, chosen retirement path, revised scope, owner, and effective date as provenance. That concrete deployed or enacted form is Record, Log & Register under the frozen taxonomy.
Nearest alternative: Decision, Gate & Allocation — Although Decision, Gate & Allocation can support this mechanism, the frozen evidence makes its operative form the act that preserves the triggering evidence, chosen retirement path, revised scope, owner, and effective date as provenance; the alternative is therefore secondary rather than defining.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Data Science & Analytics
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Reassessing, recalibrating, or removing a production proxy after drift is closest to model and metric governance in data science; NIST calls for regular metric reassessment and treats recalibration or removal as documented responses.
Related originating lineages:
- Accounting & Auditing — Audit documentation supplied the durable rationale and downstream-claim trace.
- Organizational & Management Science — Named ownership, rationale, and downstream claim changes derive from organizational decision-record practice.
Review resolution: The blind reviewers disagreed on primary lineage. Light authoritative research resolves the defining form in favor of data_science: Reassessing, recalibrating, or removing a production proxy after drift is closest to model and metric governance in data science; NIST calls for regular metric reassessment and treats recalibration or removal as documented responses. The other materially formative traditions are retained as alternates; current breadth of use remains separate as domain_reach=multi_domain.
Attribution caveat: The decision-record wrapper is an encyclopedia synthesis around proxy governance. The exact retirement record appears to be an encyclopedia synthesis of metric governance practices.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; medium confidence.
Sources consulted:
Notes¶
[n1] An architecture decision record (ADR) is a short document capturing a significant choice, its context, and its consequences, kept alongside the system it governs. Its enduring value is not the decision but the preserved reasoning, so future maintainers understand why — the same job a retirement record does for a demoted proxy. ↩