Reputation as a Measurement Layer

Why measurement is its own layer

Most platform token designs collapse measurement into money — you did a thing, you receive tokens. That coupling is expensive. It means every change to what the platform values is a change to monetary policy, every recalibration is a supply event, and the unit you use to reason about contribution is the same unit people trade.

Separating them buys three things. Reputation can be non-transferable, so it cannot be bought and accumulating it requires doing the thing. It can be uncapped, so the system never has to refuse to acknowledge genuine work. And it can live off-chain, so its values can be revised as the platform learns which behaviours actually produce value — without a migration, a fork, or a governance fight.

Everything below is about making that revisability safe.

Structure

A vector, not a number

The first design decision, and the one most systems get wrong by default.

Reputation areas

Reputation is encoded as a set of metrics, each giving quantified insight into one dimension of past contribution:

\rho_i = (\rho_i^1, \dots, \rho_i^k)

In the MVP this was deliberately small — content reputation (quality of creative contribution), an engagement factor (activity in the network), and a set of badges for discrete achievements:

\rho_i = (\rho_i^c, \rho_i^g), \qquad \rho_i^b \in \mathcal{P}(\{1,\dots,B\})

These measure genuinely different things. Someone who writes one exceptional scene and someone who shows up every day for a year are both valuable, and averaging them into a single number destroys the distinction that makes each of them useful to reason about.

The score, and when not to use it

Aggregation to a scalar is available when you need it:

\hat{\rho}_i = f(\rho_i^1, \dots, \rho_i^k), \qquad f : X^k \mapsto [a,b]

But aggregation implicitly commits to a weighting, and that commitment is frequently wrong for the consumer of the number. A room lead assembling a team on creative strength does not want the same weighting as a mechanism deciding governance influence.

The rule we settled on: where a consumer cares about one dimension, expose the vector. Aggregate only where a single ordering is genuinely required — and treat the weights used to produce it as a parameter of that specific use, not a property of the participant.

Aggregation

Choosing a shape, not just weights

The functional form encodes a philosophy about what kind of contributor the platform wants. Better to make that explicit than to inherit it from a default.

Beyond the weighted sum

The obvious aggregator is linear:

f(\rho_i) = \omega_1 \rho_i^1 + \dots + \omega_k \rho_i^k, \qquad \sum_{k} \omega_k = 1

It has one property that is often undesirable: perfect substitutability. Being twice as good at engagement exactly compensates for being half as good at content. If the platform believes a contributor needs some of each, the linear form cannot say so.

The generalisation is a parameterized mean over weighted sub-indicators with a curvature parameter:

G(\rho_i) := \Big( \sum_{k=1}^{K} \omega_k \, (\rho_i^k)^{\gamma} \Big)^{1/\gamma}, \qquad \gamma \in [-\infty, 1)

The parameter $gamma$ spans a meaningful range of platform philosophies. At one extreme the score is dominated by your weakest dimension:

\gamma \to -\infty \; \Longrightarrow \; G(\rho_i) = \min\{\rho_i^1, \dots, \rho_i^K\}

and at the other it is a weighted geometric blend:

\gamma \to 0 \; \Longrightarrow \; G(\rho_i) = \prod_{k=1}^{K} (\rho_i^k)^{\omega_k}

Between them sits everything from "we reward specialists" to "we require well-roundedness". This is a product decision expressed as a number, which is exactly where product decisions should live — visible, adjustable, and arguable.

Weights as the governance surface

Once the shape is fixed, the weights become the natural governance lever. Adjusting $omega_k$ changes what the system rewards without touching the catalogue, the contracts, or anyone's accumulated history.

This also opens the door to local weights rather than one global set — the benefit-of-the-doubt approach from the composite indicator literature, where each participant is evaluated under the weighting most favourable to them, subject to shared constraints. It removes the argument about whose contribution type the platform secretly prefers, at the cost of losing a single common ordering.

Evolution

A catalogue that can change without lying

The part that bites in production. A scoring system that cannot evolve its own inputs is not one you can run for ten years.

The action catalogue

Points are assigned against an explicit catalogue mapping each qualifying action to a point value at a given time:

C_t : a \mapsto p_{a,t}

Making this an artefact rather than an implementation detail is most of the value. It is inspectable, it is arguable, and it is the thing governance actually votes on.

Adding and retiring actions

Catalogues change. New behaviours become valuable; old ones stop being. Both operations have to renormalise the weights in a defined way, or every change silently reprices everyone's history.

Adding a new action with intended weight $omega_{K+1}$ :

\omega_{\text{add}} = \big( \omega_1 (1 - \omega_{K+1}), \dots, \omega_K (1 - \omega_{K+1}), \; \omega_{K+1} \big)

Removing an existing action (labelled first without loss of generality):

\omega_{\text{remove}} = \Big( \frac{\omega_2}{\sum_{k=2}^{K} \omega_k}, \dots, \frac{\omega_K}{\sum_{k=2}^{K} \omega_k} \Big)

Both preserve the relative weighting among surviving actions and keep the weights summing to one. It is a small piece of arithmetic and it is the difference between a catalogue that can be maintained and one that accumulates distortion every time it is touched.

Recency and decay

Point values are dynamic. The same action can be worth more later, which is how the system expresses that recent contribution matters more than historic:

p_{a,t} \leq p_{a,t+1}

Rather than decaying accumulated balances directly — which is visible, unpopular, and feels like confiscation — the equivalent effect is achieved by inflating new distributions against a discount factor per reputation area:

p_{a,t+1} = f_{\text{calc}}(p_{a,t}, \rho, \delta_x), \qquad x \in \{c, g\}

Why inflate rather than decay. Both produce the same relative ordering. One takes something away from people who earned it; the other gives more to people earning now. The mechanism is indistinguishable and the reception is not.

References

Cherchye, L., Moesen, W., Rogge, N. & Van Puyenbroeck, T. (2007). An Introduction to 'Benefit of the Doubt' Composite Indicators. Social Indicators Research.
OECD / JRC (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide.