Counterfactual Proximity Signal Calibration¶

Calibrate how much an almost-happened better or worse outcome should teach, motivate, warn, or matter.

Overview¶

Counterfactual Proximity Signal Calibration is the pattern for using almost-events responsibly. It preserves the factual outcome while adding a separate, bounded signal for how close the case came to a better or worse plausible counterfactual. This is useful because many systems learn too late when they wait for completed outcomes only, but they also drift when imagined alternatives are treated as facts.

The key move is not simply asking “what if?” The key move is asking: what actually happened, what nearby value-changing alternative was plausible at the time, how close was it, how large was the value difference, and what bounded response should that proximity signal trigger?

When to use it¶

Use this archetype when the difference between two cases with the same outcome label matters. A failed attempt that nearly succeeded, a safe event that almost injured someone, or a decision that barely missed a better path can all contain useful information. The archetype is especially useful when actual losses or rewards are rare but almost-losses and almost-rewards are frequent enough to guide learning.

Do not use it to launder actual outcomes. A real harm remains a harm. A real failure remains a failure. A real success remains a success. The counterfactual proximity signal is a second layer beside the factual record, not a replacement for it.

Core components¶

Component	Description
Factual Outcome Anchor ↗	The factual outcome anchor records what happened before any alternate path is considered. This includes the observed result, timing, known constraints, and relevant decisions. It prevents “almost” from becoming a substitute for the actual record.
Valued Counterfactual Contrast ↗	The valued counterfactual contrast names the nearby better or worse state. It might be injury instead of no injury, reward instead of no reward, breach instead of blocked attempt, or passing threshold instead of narrowly missing it. The alternative must differ in value, not merely in description.
Proximity Metric or Ordering ↗	The proximity metric defines what “close” means. Sometimes closeness is measured in seconds, meters, points, dollars, probability mass, causal steps, or procedural barriers. Sometimes it is ordinal: distant, moderate, close, critical. The important requirement is that the dimension is named.
Value Delta Frame ↗	Proximity alone is not enough. A tiny value difference near a threshold should not necessarily dominate a larger but less close threat. The value delta frame states how much better or worse the counterfactual would have been and whose value frame is being used.
Signal Transformation Rule ↗	The signal transformation rule converts distance and value into a bounded learning, warning, motivational, or decision signal. It prevents two opposite errors: treating near misses as irrelevant because the outcome did not occur, and treating them as if the outcome fully occurred.
Eligibility and Plausibility Gate ↗	The plausibility gate is the discipline that keeps this archetype from becoming hindsight storytelling. A counterfactual should be weighted only if it was reachable under what was known, possible, and constrained at the time.
Learning Update Boundary ↗	The learning boundary states what may change because of the proximity signal. A near miss may trigger review, retraining, redesign, or retesting. It should not automatically create full credit, full blame, or a completed-outcome statistic.

Common mechanisms¶

A near-miss distance scorecard scores closeness across explicit dimensions. A counterfactual value-delta table pairs plausible alternatives with their value difference from the actual outcome. A threshold band map shows which cases are far misses, close misses, threshold crossings, and close escapes. A close-call review protocol investigates almost-events as evidence while preserving actual outcome accounting.

For decision review, a regret-weighted decision log records whether the better alternative was genuinely available at the time. For training and product experimentation, an almost-reward annotation allows targeted feedback without granting full reward. For validation, a proximity signal backtest checks whether past near-miss signals actually predicted later risk, learning, or improvement.

Parameter dimensions¶

The most important parameters are proximity dimension, value-delta magnitude, plausibility confidence, signal cap, response tier, and calibration horizon. A strong draft or implementation should say what each parameter means in context. For example, a cybersecurity near miss may use procedural distance from privilege escalation, while an education near miss may use conceptual distance from a correct proof.

Invariants to preserve¶

The factual outcome and counterfactual signal must remain separate. Every weighted counterfactual must have a value polarity, a distance basis, and a plausibility basis. Signal strength must be bounded. Near misses may inform learning and response, but they must not become completed harms or rewards. Hindsight-only alternatives should be rejected or marked as low-plausibility. Vividness should not substitute for evidence.

Neighbor distinctions¶

Counterfactual Comparison compares actual and alternative paths to isolate causal effect or decision value. This archetype is narrower: it calibrates the signal generated by closeness to a nearby alternative.

Prediction-Error Learning Calibration compares expected and received outcomes. This archetype compares the actual outcome with a plausible unrealized outcome of different value. An event can have zero reward prediction error and still be counterfactually close to harm, or it can have a large prediction error without being a near miss.

Counterfactual Subtraction estimates effect by subtracting a constructed baseline. This archetype does not estimate intervention effect; it decides how much an almost-event should teach, warn, motivate, or trigger action.

Near-Miss Normalization is a failure mode: repeated close calls become accepted as normal. This archetype can help prevent it by assigning calibrated response weight before a full failure occurs.

Examples¶

In aviation safety, a loss of separation that does not become a collision is recorded as no collision plus a high-proximity near-miss signal. That signal can trigger review of spacing, procedure, communication, or equipment without falsifying the outcome record.

In education, an answer that is one inference away from a correct proof should not receive full credit, but it may deserve a different instructional response from a blank or unrelated answer. The almost-reward signal points to the missing step.

In product experimentation, a feature that narrowly misses an adoption threshold might be refined and retested rather than discarded, but only if the proximity is meaningful and not just a convenient story.

In cybersecurity, a blocked intrusion attempt may deserve more than a “no breach” label if it came one action away from privilege escalation. The signal remains a near-miss warning, not a breach statistic.

Non-examples¶

A failed project is not “almost successful” merely because the team can imagine a better ending. A clinical harm is not a near miss if the harm actually occurred. A trader’s hindsight regret is not valid proximity if the alternative required information unavailable at the decision time.

Failure modes¶

The main failure modes are outcome substitution, hindsight counterfactual inflation, salience overweighting, near-miss normalization, threshold gaming, and ruminative counterfactual loops. Each failure comes from confusing a calibrated signal with a fact, a feeling, or a fully justified causal inference.

Reconciliation notes¶

This draft is deliberately bounded. It does not replace accepted counterfactual_comparison, prior queue prediction_error_learning_calibration, or upcoming counterfactual_subtraction. It fills the specific gap for proximity-graded counterfactual signals: the internal or institutional weight assigned to an almost-happened outcome of different value.

Common Mechanisms¶

Almost-Reward Annotation
Close-Call Review Protocol
Counterfactual Plausibility Filter
Counterfactual Value-Delta Table
Near-Miss Distance Scorecard
Near-Miss Response Tier
Proximity Signal Backtest
Regret-Weighted Decision Log
Salience Overweighting Check
Threshold Band Map

Compression statement¶

Counterfactual proximity signal calibration anchors on what actually happened, identifies a plausible nearby outcome of different value, measures how close the actual episode came to that outcome, weights the value gap by proximity, and constrains the resulting signal so near misses inform learning without being counted as wins, losses, harms, or proof.

Canonical formula: signal_weight = bounded_transform(counterfactual_value_delta × proximity_to_value_changing_counterfactual × plausibility_confidence)

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (7)

Clustering Illusion: A finite sample from a random process is misread as patterned because randomness reliably produces clumps that no null model has been compared against.
Counterfactual Proximity Weighting: An internal signal is graded not by the outcome alone but by its distance to a nearby counterfactual outcome of different value, so near misses carry near-reward signals.
Counterfactual Reasoning: Hypothetical alternatives.
Counterfactuals: Alternate hypothetical scenarios.
Precision Weighting: Signals about the same latent quantity receive influence in proportion to estimated precision, so more reliable evidence contributes more while context may revise the weights.
Regret: Disvalue from comparing an outcome against a better forgone alternative.
Reward Prediction Error: The signed difference between received and expected reward or value becomes a teaching signal for updating value estimates, policies, and reward-linked behavior.

Also references 15 related abstractions

Baseline Deviation: An observation is interpreted against a declared reference and flagged as departing from it, producing deviation as a first-class fact.
Calibration: Aligning a system's output to a trusted reference by measuring deviation, adjusting to reduce it, and monitoring for drift.
Cognitive Appraisal: The interpretive step that evaluates a situation's significance and one's coping resources, determining the emotional and behavioral response rather than the raw stimulus doing so.
Comparison: Place items in a shared frame along chosen dimensions to read off a relation between them.
Counterfactual Subtraction: An effect is estimated by subtracting a constructed baseline representing what would have obtained absent the intervention, so the inference rests on the baseline's credibility, not the arithmetic.
Emotional Reasoning: Decisions shaped by emotion.
Feedback: Outputs influence inputs.
Loss Aversion: Losses felt stronger than gains.
Near-Miss Normalization: A system reads the absence of harm in a narrowly-averted event as evidence of robustness rather than as a signal of lost margin, so it resets the margin to the smaller value and drifts its operating envelope toward the failure boundary while still counting itself safe.
Outcome-Defined Adequacy: Adequacy is specified by whether a defined outcome is produced in context, leaving the form open to any realization that achieves it.

▸ Show 5 more

Variants¶

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Near-Miss Signal Calibration · risk or failure variant · recognized

A safety or reliability variant that treats almost-harm as graded evidence without counting it as harm itself.

Distinct from parent: The parent covers any proximity-weighted counterfactual signal; this variant specializes in avoided-loss and safety-learning cases.
Use when: A process narrowly avoided failure, injury, breach, outage, or loss; Actual incident counts are too sparse to reveal danger early enough; The organization needs to learn from proximity without creating panic or blame inflation.
Typical domains: safety engineering, medicine healthcare, aviation, cybersecurity, disaster management
Common mechanisms: close call review protocol, near miss distance scorecard, near miss response tier

Almost-Reward Feedback · affective or cognitive variant · recognized

A learning variant in which coming close to a reward changes attention, motivation, or policy despite the reward not being received.

Distinct from parent: The parent includes both near-gain and near-loss signals; this variant emphasizes almost-success and learning motivation.
Use when: A learner, agent, team, or product nearly reached a valued target; Outcome-only scoring would treat the episode as equivalent to a distant failure; The system needs to preserve the informational value of almost-success without awarding full credit.
Typical domains: education, product experimentation, reinforcement learning, coaching, sales operations
Common mechanisms: almost reward annotation, counterfactual value delta table, proximity signal backtest

Regret Proximity Weighting · affective or cognitive variant · recognized

A decision-review variant that gives stronger learning weight to choices close to a better forgone outcome.

Distinct from parent: The parent covers internal signal calibration generally; this variant emphasizes decision review and regret governance.
Use when: A decision’s actual result is being compared with a plausible nearby alternative; Decision-makers may overlearn from distant fantasies or underlearn from close misses; The review needs proportional regret rather than binary success/failure judgment.
Typical domains: strategy, investment decisions, clinical review, public policy, career planning
Common mechanisms: regret weighted decision log, counterfactual plausibility filter, salience overweighting check

Near names: Counterfactual Closeness Weighting, Near-Reward Signal Calibration, Close-Call Weighting.