# Research Signals

## Purpose

Coulomb Research needs observable signals that help participants and agents understand what is attracting attention, shaping downstream work, surviving criticism, or generating useful activity.

Signals are measurements of the research system.

They are not truth.

## 1. Foundational Separation

Coulomb should keep at least four concepts distinct:

```text
ATTENTION
What are people looking at?

INFLUENCE
What does other work depend on?

EPISTEMIC SUPPORT
What survives available evidence and criticism?

UTILITY
What is being reused?
```

A fifth signal family is especially important:

```text
GENERATIVITY
What produces valuable new questions, methods, experiments, or explanations?
```

No single signal should silently stand in for the others.

## 2. Signal Families

### 2.1 Attention

Possible observations:

- views,
- reads,
- downloads,
- bookmarks,
- external mentions,
- discussion activity,
- watch/follow events.

Attention may be positive, negative, or merely fashionable.

It must not be interpreted as quality.

### 2.2 Influence

Possible observations:

- conventional citations,
- typed citations,
- propositions depending on the artifact,
- methods reused downstream,
- later work extending the artifact,
- inclusion in syntheses,
- downstream research programs.

Typed relations are preferable to raw citation count.

Example:

```text
depends-on
extends
supports
challenges
uses-method
uses-data
uses-software
```

### 2.3 Epistemic Support

Possible observations:

- independent evidence,
- successful replication,
- prediction success,
- unresolved criticism,
- contradictory evidence,
- retractions or supersession,
- methodological robustness.

Support should usually be represented as a multi-dimensional state rather than one score.

### 2.4 Utility and Reuse

Possible observations:

- software imports,
- dataset reuse,
- benchmark reuse,
- method adoption,
- protocol adoption,
- implementation reuse,
- educational reuse,
- operational deployment.

Reuse may reveal value that citation systems miss.

### 2.5 Criticism

Possible observations:

- substantive critiques,
- counterexamples,
- failed replications,
- methodological objections,
- conceptual contradictions,
- corrections.

A high criticism signal is not inherently negative.

It may indicate that an artifact is important enough to test aggressively.

### 2.6 Generativity

Possible observations:

- new questions spawned,
- new propositions created,
- new experiments triggered,
- research programs formed,
- software or datasets produced,
- cross-domain extensions,
- independent follow-up work.

Generativity measures whether research creates productive new problem spaces.

### 2.7 Velocity

Signals should often be time-normalized.

Examples:

```text
citations / month
reuses / month
replications / year
new dependent propositions / week
```

Velocity helps surface emerging work before accumulated totals dominate.

## 3. Scope and Normalization

Signals should support explicit comparison scopes.

Possible dimensions:

- field,
- subfield,
- year,
- publication age,
- artifact type,
- cohort,
- geography,
- research program.

Comparing raw totals across biology, mathematics, AI, and philosophy is usually misleading.

Percentiles are often preferable to universal thresholds.

## 4. Signal Representation

A research artifact should expose a signal vector rather than a universal score.

Example:

```text
Attention       98th percentile
Influence       91st percentile
Reuse           84th percentile
Replication     62nd percentile
Criticism       77th percentile
Generativity    95th percentile
Support         contested
```

The exact metrics and normalization method must remain inspectable.

## 5. No Canonical Impact Score

Coulomb should avoid a single global "impact score".

Reasons:

- it collapses unlike dimensions,
- it invites gaming,
- it hides value judgments,
- it creates prestige hierarchies,
- it becomes a Goodhart target,
- it mistakes observability for truth.

If composite scores are offered, the weighting function must be explicit and contestable.

## 6. Signals as Control Inputs

Signals can drive research-system behavior.

Examples:

```text
high attention
   → allocate additional criticism

rapid reuse
   → prioritize replication

foundational dependency
   → audit provenance and robustness

high controversy
   → synthesize competing explanations

high generativity
   → inspect what made the work productive
```

Signals therefore participate in the cybernetic feedback loop.

## 7. Actor Signals

Actor-level summaries should be derived from contribution events, not maintained as universal status.

Examples:

- replication history,
- prediction history,
- criticism history,
- artifact reuse,
- successful resource allocation,
- correction acceptance.

Different communities should be able to derive different reputation functions.

## 8. Anti-Gaming Principles

Signals should:

- rely on attributable events,
- discount obvious self-interaction where appropriate,
- distinguish automated from human activity,
- expose suspicious activity,
- use time windows,
- use field normalization,
- preserve raw events for independent recomputation.

Whenever possible, Coulomb should publish enough event data for alternative ranking functions to be constructed independently.

## 9. Research Question

The signal system itself belongs to RP-000.

Key question:

> Which observable signals best predict durable explanatory progress without creating destructive incentives?