Attributing Contribution in Collaborative Production

The problem

A group forms, produces something together, and the output is tokenised as a single asset. The asset has to be divided:

\sigma = (\sigma_1, \dots, \sigma_n), \qquad \sum_{i=1}^{n} \sigma_i = 1

Every scheme for choosing $sigma$ is a statement about what the system values, and every scheme becomes a target the moment it is legible. Three properties make this harder than it looks.

Contributions are incommensurable. One person invents a character, another writes dialogue, another runs the community, another brings a commercial partner. There is no natural exchange rate between these, and no observable quantity to count.
Nobody outside the group can see the work. The information about who actually contributed exists only inside the group. Any mechanism that relies on an external assessor is guessing.
The participants are the beneficiaries. The only people who know the answer are precisely the people with a direct financial interest in distorting it.

That third point rules out the obvious approach. Ask collaborators to rate each other, average the ratings, and split accordingly:

\hat{c}_i = \frac{1}{n-1} \sum_{j \neq i} e_{ji}

This is a straightforward invitation to collude, or simply to inflate. Under naive averaging, reporting that everyone else contributed nothing weakly dominates reporting honestly. The mechanism has to do more work.

Mechanism

Paying for honesty, not just for opinion

The move is to stop treating peer evaluation as data collection and start treating it as a game whose equilibrium you get to choose.

Peer evaluation as an elicitation problem

After a joint project completes, each participant evaluates the others. This produces an evaluation matrix:

E = (e_{ij})_{i \neq j}, \qquad e_{ij} \in [0,1]

The mechanism takes $E$ as input and returns the ownership split. The design question is not how to aggregate — that part is easy — but how to make truthful reporting the strategy a self-interested participant would choose anyway.

This is a well-studied problem outside our setting. Eliciting subjective judgements that nobody can verify — because there is no ground truth to check them against — is exactly what Bayesian Truth Serum was built for (Prelec 2004). The insight it turns on is that a respondent's prediction of what others will say carries information about their own private signal, and comparing the two exposes dishonesty.

The two-part score

Each participant supplies two things: their assessment of their collaborators, and their prediction of the distribution of everyone else's assessments. Their score decomposes accordingly:

s_i = \underbrace{\log \frac{\bar{x}_i}{\bar{y}_i}}_{\text{information}} \; + \; \alpha \underbrace{\sum_k x_{ik} \log \frac{\bar{x}_k}{y_{ik}}}_{\text{prediction}}

The information score rewards answers that are surprisingly common — more common in the actual population than the population itself predicted. The prediction score rewards accurately anticipating what others would say. Neither can be gamed in isolation: inflating your own contribution costs you on the first term, and misrepresenting your beliefs about others costs you on the second.

The property that makes this usable:

\mathbb{E}[s_i \mid \text{truthful}] \; \geq \; \mathbb{E}[s_i \mid \text{any other report}]

Truthful reporting maximises expected score, and it does so without the mechanism ever needing to know the right answer. That is the entire point — the mechanism is Bayesian Nash incentive compatible against a population whose private information it cannot observe.

Why this fits creative collaboration specifically. Truthfulness holds when participants lack informative prior knowledge of each other's capabilities. A newly-formed group of strangers assembling around a project is close to the ideal case. A group that has worked together for years is not — they can coordinate, and the guarantee weakens.

Splitting the score

In implementation the reputation credited to a participant separates the two channels explicitly:

\hat{\rho}_i = w_p \, \pi_i + w_t \, \tau_i

where $pi_i$ is what peers said about you and $au_i$ is how honestly you assessed them. Keeping these separate rather than collapsing them into one number matters for two reasons. It lets the system weight them independently as a governance parameter, and it means a participant who contributed modestly but evaluated honestly still accrues standing — which is what you want, because the mechanism depends on those people showing up.

Application

From scores to ownership

The mechanism produces relative standing. Turning that into a claim on future revenue requires deciding how much of the future is on the table.

The contribution budget

A project sets aside a portion of its future revenue as a reward budget — the share it is willing to distribute to people who help beyond the core production process:

B \in [0, 1] \quad \text{of future revenue, set per project}

This is optional and set per project. It exists to solve a cold-start problem: a new project has no revenue and no audience, but it does have a credible claim on its own future, and it can pay with that. The mechanism then allocates:

\sigma_i = \frac{\hat{\rho}_i}{\sum_{j} \hat{\rho}_j} \cdot B

Requirements we held the design to

Property	What it rules out
Scalability	Schemes whose behaviour changes as the asset becomes popular. The split must mean the same thing at a hundred users and a hundred thousand.
Fairness across kinds of work	Any implicit privileging of creative contribution over promotional contribution, or vice versa. Both can be what made the asset succeed.
Proportionality over time	Early contributions being mechanically worth more than later ones purely because they were early — while still recognising that early risk is real.
Continuity	One-shot allocations that leave nothing for contributors who arrive in year three of a ten-year asset.

Proportionality and continuity pull against each other, and the tension is not resolvable in general. A single perpetual allocation gives early contributors unbounded upside and leaves later ones with nothing to earn; repeated per-period allocations fix that but make each period's reward depend on revenue timing rather than contribution quality, and push support toward assets that are already winning. We treated the choice as a per-project parameter rather than a protocol-level answer, because the right point differs between a project that needs a cold start and one that needs sustained contribution.

What did not survive contact

Several candidate mechanisms were evaluated and set aside, and the reasons are more useful than the shortlist.

Trust-graph propagation (EigenTrust and relatives). Elegant for reputation over a network of repeated pairwise interactions. Wrong shape here: contribution to a single joint output is not a transitive trust relation, and the graph is too sparse at the scale of one project team to say anything.

Bayesian online ranking. Strong when there is a repeated observable outcome to update against — wins and losses. There is no such signal in creative production. The outcome arrives years later, once, and is confounded by everything else.

Market valuation of contributions. Attractive because it sidesteps elicitation entirely, and the direction I would still most want to explore. It needs liquidity and a tradeable claim, and at the point where attribution has to happen there is neither.

Assessment

Where this is fragile

The guarantee is real but conditional, and the conditions are worth stating plainly.

Known limits

Small groups. The truthfulness guarantee is asymptotic. Creative teams are small — often under ten — and the small-population variants of the mechanism are meaningfully weaker than the limiting case.
Repeated interaction erodes it. Collaborators who expect to work together again have a shared future to trade against, which is precisely the coordination the mechanism assumes away.
Non-contributors need separate handling. Somebody who added nothing should be excluded rather than assigned a small share, and that is a screening question the scoring mechanism does not answer on its own.
It is a survey. Every participant must complete a two-part elicitation after every project. That is real friction on people who joined to make things, and no amount of mechanism design removes it.

None of these are reasons not to use it. They are reasons the mechanism is one input to an ownership decision rather than the whole of it, and reasons the surrounding system needs a way to override an outcome that is obviously wrong.

References

Prelec, D. (2004). A Bayesian Truth Serum for Subjective Data. Science 306(5695).
Witkowski, J. & Parkes, D. (2012). A Robust Bayesian Truth Serum for Small Populations.
Carvalho, A. & Larson, K. (2012). Sharing Rewards Among Strangers Based on Peer Evaluations. Decision Analysis.
Kamvar, S., Schlosser, M. & Garcia-Molina, H. (2003). The EigenTrust Algorithm for Reputation Management in P2P Networks.