Pragmatic Utility and Value Alignment

About this pattern

This is a generated FPF pattern page projected from the published FPF source. It is canonical FPF content for this ID; it is not a FPF Reference product feature page.

How to use this pattern

Read the ID, status, type, and normativity first. Use the content for exact wording, the relations for adjacent concepts, and citations to keep active work grounded without pasting the whole specification.

Type: Part E FPF evaluation and repair pattern Status: Stable Normativity: Normative unless a section is explicitly informative

Use this pattern when a project treats a visible measure, score, proxy, benchmark, dashboard, quality value, review result, release posture, or evidence volume as if it were the practical value or objective itself.

Keywords

  • pragmatic utility
  • proxy-to-value alignment
  • Goodhart
  • Campbell
  • surrogation
  • minimally viable value slice.

Relations

E.13builds onThe Eleven Pillars
E.13coordinates withDecision Theory (Decsn-CAL)
E.13coordinates withEvidence Graph Referring (C-4)
E.13explicit referenceDecision Theory (Decsn-CAL)
E.13explicit referenceHuman-Centric Working-Model
E.13explicit referenceQuality Improvement Loop Method
E.13explicit referenceEvidence Graph Referring (C-4)

Content

Use This When

Use this pattern when a project treats a visible measure, score, proxy, benchmark, dashboard, quality value, review result, release posture, or evidence volume as if it were the practical value or objective itself.

Typical moments:

  • a metric improves, but the team cannot say what intended value improved;
  • a quality score, all-5 posture, assurance level, citation count, source count, or review pass becomes the target;
  • a proxy is used as a gate, incentive, resource-allocation signal, reputation signal, or release argument;
  • a model, method, pattern, or system is formally better while users, operators, safety, maintainability, learning, or decision quality get worse;
  • an evaluation loop adds apparatus to satisfy the evaluator instead of improving the object of concern.

First useful move. Name the intended value or objective, name the proxy or visible measure, and state how that proxy is being used now: measure, target, incentive, gate, release argument, decision driver, reputation signal, repair target, or orientation cue.

What goes wrong if missed. The team optimizes the proxy and loses the value. It can produce a better score, cleaner review proof, larger source packet, or more complete record while practical utility gets worse.

What this buys. FPF can keep measurement, evaluation, and quality loops useful without letting their visible outputs replace the value they were meant to serve.

Not this pattern when.

  • If the question is whether a measurement scale is legal, use C.16.
  • If the question is ordinary pattern quality, use E.21; use E.13 only when a visible quality value is being treated as the practical value.
  • If the question is DRR adequacy, use E.9.DA; use E.13 only when DRR marks become a surrogate for decision usefulness.
  • If the question is whole-FPF Pillar adequacy, use E.2.DA; use E.13 only when Pillar values become the target.
  • If the question is assurance, gate passage, evidence sufficiency, or decision authority, use the governing pattern for that claim before treating the visible proxy as value.

Problem Frame

Practical work often needs visible measures. Teams use scores, dashboards, quality coordinates, tests, evidence counts, source freshness rows, release checks, and worked examples because invisible value is hard to steer directly.

The danger starts when the visible measure becomes the object being optimized. A proxy can be useful as a signal and harmful as a target. A pattern can become easier to defend while harder to use. A safety dashboard can look better while unmeasured hazards increase. A review result can look more complete while the decision it was meant to support becomes less decisive.

E.13 governs the proxy-to-value repair. It asks whether the visible measure still serves the intended value in the declared use, and what became worse when the measure improved.

Problem

Without E.13:

  1. Measures replace objectives. Teams speak as if the score, metric, benchmark, assurance level, or all-5 posture is the value.
  2. Evaluation loops become reward functions. A reviewer asks for improvement; the author adds fields, guards, source rows, proof sketches, and relation catalogues until the visible evaluation looks better.
  3. Unmeasured value is damaged. Usability, safety margin, maintainability, learning, domain fit, affordability, or operator action quality gets worse while the proxy improves.
  4. Proxy use is not typed. The same metric is treated as orientation cue, target, incentive, gate, and release proof without saying which use is live.
  5. No value slice exists. The text claims practical payoff, but no minimally viable slice shows the value being realized in a case.

Forces

ForceTension
Measurement vs valueProjects need visible signals, but signals can replace the value they indicate.
Local optimization vs protected qualitiesA local score can improve while another value-bearing dimension worsens.
Evaluation pressure vs object improvementA reviewer-facing mark can be easier to raise than the object is to improve.
Proxy affordability vs value evidenceA proxy is cheap; demonstrating value can be expensive.
Release confidence vs ongoing distortionA proxy may be safe for orientation but unsafe as a gate, incentive, or release argument.

Solution

Use ProxyToValueAlignment as a short repair note, not a new bureaucracy.

ProxyToValueAlignment:
  ObjectOfConcern:
  IntendedValueOrObjective:
  ProxyOrVisibleMeasure:
  ProxyKind:
  CurrentProxyUse: <orientation | measure | target | incentive | gate | release argument | decision driver | reputation signal | repair target>
  AffectedDecisionOrWork:
  ProtectedQualities:
  WhatImproved:
  WhatGotWorse:
  MinimallyViableValueSlice:
  AdmissibleUseNow:
  BlockedOverread:
  RepairOrStop:
  ReopenCondition:

Keep the note as small as the case allows. The fields exist to restore the value relation, not to create another checklist target.

Name the Value Before the Proxy

Name the intended value, objective, or practical payoff in terms of the work it is supposed to improve. If only the proxy can be named, lower the claim: the project has a measure, not a demonstrated value relation.

Type the Proxy Use

A proxy can be harmless as an orientation cue and dangerous as a target. State the current proxy use explicitly.

Proxy useAdmissible useDanger
Orientation cueHelps decide where to look next.Mistaken for evidence of value.
MeasureReports one declared characteristic under C.16.Treated as the whole objective.
TargetWork is optimized to move the proxy.Goodhart pressure.
IncentivePeople or agents are rewarded for the proxy.Behavioral distortion and gaming.
Gate or release argumentPassage depends on the proxy.Proxy becomes authority.
Reputation or status signalPeople, teams, models, or patterns are ranked by the proxy.Surrogation and status gaming.
Repair targetThe object is changed to raise a coordinate or score.Apparatus is added instead of value.

Ask What Got Worse

Whenever a proxy improves under optimization pressure, ask what became worse or more fragile. Check at least usability, affordability, safety or harm boundary, maintainability, domain fit, source preservation, decision quality, learning, and neighboring-pattern fit when they are live in the case.

If nothing worsened, say which loci were checked. If no loci were checked, do not claim value alignment.

Require a Minimally Viable Value Slice

Do not require every project to create a lifecycle artifact named MVE. Require a minimally viable value slice: one compact case, worked slice, observation, trial, user/operator moment, or decision replay where the intended value is visible enough for the declared use.

The value slice may be small. It must show the value, not merely the proxy.

Repair by Value Movement

When the proxy has displaced the value, repair one of these:

  • change the proxy use from target/gate/incentive to orientation or bounded measure;
  • add a protected quality or counter-metric that names the value at risk;
  • change the work or design so the value slice improves, not only the proxy;
  • split the claim: one measure report, one value claim, one assurance or gate claim if needed;
  • stop the value claim until a value slice or better proxy relation exists.

Archetypal Grounding

CaseProxy pressureE.13 repair
Pattern quality loopAll-5 pattern-quality posture becomes the target.Use E.21 values as measurements; repair only substantive content movement and record what worsened when apparatus grew.
DRR reviewSource rows and selected-locus tables grow while the decision remains vague.Use E.9.DA; the DRR improves only when selected answer, source payload, or first drafting move improves.
Safety dashboardA lower incident count is used as proof of safety.Split measure, reporting behavior, unreported hazard, and safety assurance; use the safety/assurance pattern for the stronger claim.
AI reward modelA model gets higher reward or judge score by exploiting the specification.Treat the score as proxy; inspect unmeasured intended outcome and blocked value dimensions.
Manufacturing throughputThroughput rises while rework, fatigue, or latent defect risk rises.Keep throughput as a measure; add protected qualities and a value slice for delivered usable output.

Conformance Checklist

CheckRequirement
CC-E13-1The repair names the intended value or objective before the proxy.
CC-E13-2The proxy or visible measure is typed by current use: orientation, measure, target, incentive, gate, release argument, decision driver, reputation signal, or repair target.
CC-E13-3If a proxy improved, the repair asks what got worse and names checked loci or protected qualities.
CC-E13-4A minimally viable value slice shows the intended value for the declared use, or the value claim is lowered.
CC-E13-5The repair does not treat evaluation values, source counts, review praise, all-5 posture, assurance level, or release status as value by itself.
CC-E13-6Stronger claims are sent to their governing patterns: measurement to C.16, quality evaluation to E.21/E.9.DA/E.2.DA, assurance to B.3, gate passage to A.21, decision authority to C.11, and value/proxy alignment here.
CC-E13-7The repair changes value movement, proxy use, protected qualities, claim split, or stop condition; it does not close by adding proof apparatus alone.

Common Anti-Patterns

Anti-patternSymptomRepair
Score as valueA higher score is reported as practical improvement.Name intended value, proxy use, and value slice.
All-5 targetingA pattern or DRR is rewritten to make every coordinate defensible as 5.Use the evaluation as measurement; repair content movement and protected trade-offs.
Source-count proofMore citations or source rows are treated as better decision quality.Ask which decision payload changed.
Dashboard myopiaA visible dashboard metric improves while unmeasured harm rises.Add protected qualities and split measure from value.
Proxy as gate authorityA proxy becomes a release or gate argument without the governing gate or assurance pattern.Send gate or assurance claims to their governing patterns and keep proxy use bounded.
Value slice missingPractical payoff is asserted but never shown in a case.Add a minimally viable value slice or lower the payoff claim.

Rationale

FPF needs measurement, evaluation, assurance, and release checks, but those checks remain instruments. They are not the value by themselves. E.13 keeps the visible instrument attached to the intended value and asks whether the value survives optimization pressure.

The pattern is intentionally small. Goodhart-style failure is not repaired by another large audit apparatus. It is repaired by restoring the relation among value, proxy, use position, protected qualities, and a small slice where the value is visible.

SoTA-Echoing

ClaimSource lineageLocal adoption
A measure used for decision or control can corrupt the process it monitors.Goodhart and Campbell indicator-pressure lines.CurrentProxyUse distinguishes measure, target, incentive, gate, and release argument.
Proxy optimization has distinct failure modes.Manheim/Garrabrant Goodhart variants and later proxy-failure work.WhatGotWorse and protected qualities prevent a single proxy from standing for value.
Measures can replace the strategic construct in decision makers' minds.Management-accounting surrogation work by Choi, Hecht, Tayler, and later studies.The proxy is never named as the value; the intended value is named first.
Optimizing an imperfect reward or specification can satisfy the formal signal while missing the intended outcome.AI safety specification-gaming and reward-hacking work, including formal reward-hacking analyses and current reasoning-model specification-gaming evaluations.Evaluation values, judge scores, and all-5 posture are treated as proxies that require value-slice and protected-quality checks.
Useful measures should be derived from goals and questions.Goal-Question-Metric and GQM+Strategies measurement alignment.E.13 asks for intended value/objective before proxy and asks which decision or work the proxy affects.
Human values require stakeholder and use-context inquiry, not only formal metrics.Value Sensitive Design and value-oriented design lines.The minimally viable value slice may include user, operator, manager, safety, or affected-stakeholder evidence when those values are live.

Relations

  • Implements: E.2 Pillar P7 Pragmatic Utility.
  • Complements: E.12 for cognitive ergonomics and E.14 for human-facing working models.
  • Coordinates with: E.8 for authoring practical-payoff claims, E.19 for review/admission proxy-to-value checks, E.22/E.23 for improvement framing and repeated improvement loops, C.16 for measurement legality, C.25 for engineering quality-family endpoints, E.21 for pattern quality, E.9.DA for DRR adequacy, E.2.DA for whole-FPF Pillar adequacy, B.3 for assurance, A.21 for gate passage, C.11 for decisions, and A.10 for evidence.
  • Used by: improvement loops, release checks, pattern reviews, dashboards, metric-driven work, AI reward or judge-score cases, and any project where visible performance may displace intended value.

Consequences

  • FPF can use scores and metrics without making them the object of optimization.
  • Improvement loops gain a simple value-proxy stop condition.
  • Practical payoff claims need at least a small value slice.
  • Some attractive proxy improvements are rejected, split, or lowered.
  • The cost is a small proxy-to-value check whenever a visible measure becomes a target, incentive, gate, release argument, or repair target.

E.13:End


Last Updated: 2026-06-11 — this section last modified in upstream FPF commit 3f9a2dd6 (github.com/ailev/FPF)