The Fathom Protocol

Architecture Overview

This paragraph will introduce the three main components of a fathom-assessment: 1. a concept, representing an assessable quality 2. the concept-ontology, which relates all concepts to each other. 3. assessments which draw individuals from the concepts to be assessors and, upon a positive outcome, add new individuals to them.

Concept

Concept is an umbrella term to capture any kind of skill, quality, piece of knowledge or fact that can be established about an individual. Therefore, each concept \(C\) has the following properties:

  • Parent Concepts: This is the set of concepts \(P\) that \(C\) is a subset of . For example, the concept ‘Math’ could be a parent of ‘Linear Algebra’. If there is no suitable parent, a concept can be located beneath the ‘mew’-concept - a specially designated concept with no skills associated to it and with no parents.
  • Connection-strength(s): For each parent \(p\) \(\in P\), a concept \(C\) denotes a connection strength \(c_p\) from 0 to 1, specifying the degree of similarity or difference. All connection strengths collectively add up to 1, so that when an assessment needs to draw assessors from its parents, the connections strenghts can used to determine how many assessors should stem from each parent.
  • Expiration time: Concepts can specify expiration times \(e_c\) to reflect that some skills become outdated, need to be maintained, or change over time. An example would be concepts related to taxation-laws, which are changed on a relatively frequent basis and where false or outdated information can lead to significant losses. Members who have been assessed in a concept longer ago than specified by the expiration time do not lose their certificates, but can no longer take part in the process of assessing others.
  • Members: a set of individuals who have passed an assessment in the concept in question or in one of its children
  • Weights: For each member the concept stores a set of weights, a positive integer and date, corresponding to the latest assessment in the concept. Weights are used to probabilistically call a member to act as assessors in assessments [1].
  • Owner: An ethereum address that controls what data is saved on the concept and which can move the concept around by changing its parents. [3].

The Concept-Ontology

As all concepts have at least one parent, the entirety of all concepts forms a directed graph. The only concept without parents is the initial concept, the \(mew\)-concept. It does not present any particular subject or skill but serves as the root-node of the graph and as parent to all new concepts that are unrelated to the already existing ones. Thus, moving father away from the ‘mew’, concepts become more specific.

This network of relationship among knowledge-communities is valuable when sampling members to create a pool of potential assessors. If a concept has not enough members to create a pool of assessors of sufficient size, additional assessors will be drawn from the parent concepts.

As described in Governance & Upgrades the set of concepts and its definitions will be changing over time, with concept owners expected to update a concept’s description and relation to parent concepts (in coordination with its members) and with members cloning and migrating their weights in case disagreements can not be resolved.

The Assessment Game

An assessment is the process by which a jury of qualified individuals (assessors) decides whether or not some candidate (assessee) fulfills the necessary conditions to become a member of a concept. When initiating an assessment in a concept, the assessee decides how many assessors they want and how much they are willing to pay to each one of them. That offer is forwarded to potential assessors (see setup for drawing specifics) who must stake the offered amount in order to accept. Thus, a market forms around assessments, allowing the system to scale from easy to assess, and hence cheap, concepts, to more involved, complicated, and hence expensive ones.

Upon completion of the assessment, assessors are paid the price offered by the assessee and a proportion of their stake if they come to consensus around the applicants skill. The proportion of the stake being paid back is proportional to the assessor’s proximity to the average score of the biggest cluster of scores and will also be added a portion of the stake of dissenting assessors, should there be any (see payout for details).

Also, the mechanism by which assessors log in their scores is designed such that colluding assessors can double cross each other, thereby creating a coordination problem in an adverserial environment, where the only point of coordination (schelling point) left is a truthful assessment. In case the majority vote of the assessors is positive, the assessed candidate will get i) a score in the assessed concept, similar to a grade in university or school, ii) a weight in the concept and iii) a weight in all parent-concepts, proportionally reduced by their respective connections strengths.

Assessment Process

A fathom assessment goes through five phases: A setup phase, where the assessors are called from the concept tree, the assessment, where the assessors determine the assessee’s skill, commit- and reveal-phases, where the assessors log in their score and, at last, the calculation of the result.

For each phase this section is gonna depict the choices of the involved participants, their interactions and what happens if they deviate from the protocol.

Setup

Creating the assessment:

Wanting to be certified in a concept \(C\), the assessee needs to specify the following parameters:

  1. A time period during which they would like the assessment to start and end (latest start and end time).
  2. The number of assessors \(N_a\) to be assessed by. While there is a minimum number of five assessors to guarantee a fair voting, the assessee might want to be assessed by a bigger number in order to receive a higher weight and higher chances to become assessors themselves (given that they end the assessment with a passing score).
  3. The price \(cost_a\) that each assessor will be paid.

Calling assessors from the concept-tree:

A pool of potential assessors is created by probabilistically drawing members from the concept and its parents. The selection of potential assessors happens according to a tournament-selection of size 2, starting at the assessed concept:

  • Two members are picked at random and their weights are compared.
  • The member with the higher weight is being added to the pool. In case of a draw, the member that was drawn first wins.

Thus each member has a chance of being called as assessor, whilst giving a higher chance to those with higher weights.

To make it hard to predict who will be in the pool, no more than half of the members of each concept can be called as assessors. Therefore, the maximum number of tournaments per concept \(c\) is limited by the number of members in the concept. Specifically \(Y\) is:

\[Y = \min (N_{req}, \frac{m_c}{2})\]

with \(N_{req}\) being the number of required assessors and \(m_c\) the number of members in the concept \(c\).

After \(Y\) attempts, this selection process is repeated for each of the \(n_{p_c}\)’s parent concepts of \(c\), using the connections strengths to the different parents to determine how many members are drawn from each parent.

The minimal size and the ideal size for the pool of assessor are subject to parameters and will grow with the amount of members in the network.

Assessors confirm by staking:

Each assessor that is being called, can decide to participate in it by staking the offered price. Once the desired number of assessors has confirmed, the assessment moves to the next stage. Assessors from the pool self-select whether they think would be competent judges on the concept in question. If so, they signal their intent to participate by staking the offered price. More considerations why assessors would or wouldn’t want to confirm are elaborated in the incentive section. If not enough assessors can be found before the desired start-time of the assessment, the assessment is cancelled and everybody who deposited collateral is refunded.

Assessment of the candidate

In a fathom-assessment there is no notion or form what constitutes a test and the form or procedure of how candidates are evaluated is left to each individual assessor. Ultimately, assessors express their verdict of assessee’s skill as a number on a scale (e.g. between 0 and 100) - with everything above half being considered a passing score.

Yet, what exactly defines a failing, passing or barely passing assessment can be different for each concept as well and should be agreed upon by the community. Moreover, the assessment could also be the place to put up some sybil protection mechanism in the form of extra requirements that make it hard to repeat an assessment (see sybil-attack for more details on how this could work).

Committing a Score

Sending in a score follows the commit-reveal procedure common in blockchain applications. Assessors signal that they have decided on a score by concatenating it with a secret element, also referred to as ‘salt’ and submitting its hashed value (hash=sha3(score+salt)).

If any assessors fails to commit a score before the assessment ends their stake is being burned. If, as a consequence, less assessors than would be required for the minimum size of a viable assessment have committed, the assessment is cancelled and everyone is being refunded. Otherwise, the assessment progresses to the next stage.

Steal and Reveal

To end the assessment, the assessors reveal their verdict by submitting their score and salt separately. Any assessor (or external person) who knows about another assessor’s score and salt, can do so as well, thereby stealing half of the assessor’s stake, burning the rest and eliminating him/her from the assessment game. This prevents the assessors from credibly guaranteeing each other their committment to logging in a specific score, thus making it harder to collude.

While stealing is possible at all times after an assessor has committed (even if others have not yet), revealing will only be possible after all assessors have committed and a buffer period of 12 hours has elapsed. The buffer ensures that there is time to challenge someone’s commit, even if they waited until moments before the end of the assessment period to send it in. Should any assessor fail to reveal, their stake is burned and they are eliminated from the assessment game [2]. If the number of assessors decreases below the necessary minimum, the rest of the participants is being refunded and the assessment ends without a score.

Determining the Outcome

In order for an assessment to result in a final score, one score must be in consensus with enough other scores to form a 51% majority. Two scores are considered to be in consensus if their difference is less than the consensus-distance \(\phi\). If such a score \(s_{origin}\) exists, the final score \(s_{final}\) is computed as the average of all scores that are in consensus with \(s_{origin}\).

Should there be two scores with majorities of equal sizes, the one that will result in a lower final score wins. If there is no point of consensus, the assessment is considered invalid and all stakes are burned. Otherwise the assessors’ payments will be computed as described in the following section.

Also, in case the result \(s_{final}\) is a passing score, the assessee is registered as new member of the concept with a weight \(w_i = s_{final} * N_{in}\).

Payout of Assessors

Payments to assessors consist of two parts, their returned stake and the assessee’s reward. Both are attributed differently, depending on whether or not the assessor is inside out outside the majority cluster of winning assessors.

Therefore, an assessor \(i\)’s distance \(dist_i\) from the final score \(s_{final}\) is measured against the consensus range \(\phi\):

\[dist_i = \frac{\lvert s_i - s_{final} \rvert}{\phi}\]

Outside assessors (\(1<dist_i<2\)) only get back a part of their stake, reduced linearly in relation to their distance from the final cluster. Thus, an outside assessor \(j\)’s payout is computed as:

\[payout_{out_j} = stake_j * \frac{(2 - dist_j)}{2}\]

Inside assessors (\(dist_i <= 1\)) get back their entire stake, any stake that is not returned to outside assessors (distributed equally among them) and a share of the assessee’s reward, proportional to their proximity to the final score:

\[payout_{in_i} = stake_i + \frac{1}{N_{in}} * \sum_j^{N_{out}} (stake_j-payout_{out_j}) + r_a * (1 - dist_i)\]

,with \(N_{out}\) and \(N_{in}\) denoting the number of assessors outside and inside the winning cluster and r_a being the reward of the assessee.

Thus, the best case scenario for an assessor is to be inside the winning cluster, close to the final score, with a large minority outside of it.

Any payout that is not returned to the assessors will be burned (if it were redistributed, assessors could collude to cover a range of scores and redistribute amongst them without any loss).

This figure summarizes the payout mechanism in a single graph:

Figure 1: The payout mechanism: Three Assessors are inside the winning cluster, two are outside of it.

[1]Although it’s possible to repeat an assessment, only the result of the most recent assessment will be taken into account for the weight.
[2]This should be unlikely, as at that point assessors have nothing to lose but rewards to gain.
[3]While this is not specified by the protocol, we intend the data-field (a bytes array) to be the ultimate source of truth of what a concept is about. As the owner is just any ethereum address, this allows for a variety of governance schemes to be implemented. The owner can also transfer ownership to another address or set it to zero, in which case the data-field becomes immutable.