---
type: research
title: "Elo as an Experimental Protocol"
description: "Cross-project comparison of Elo-like ratings as relative-evidence protocols across Pokémon TCG, Antigravity and Co-Scientist research systems."
tags: [research, methodology, elo, evaluation, pokemon-tcg, antigravity, co-scientist]
timestamp: "2026-08-15T17:20:00-03:00"
---

# Elo as an Experimental Protocol

## Question

Several Wikifita projects use Elo-like updates to summarize pairwise
comparisons. The shared idea is useful: a result changes a relative estimate,
and the estimate is meaningful only with its comparison population, update
rule and evidence source. The shared name does not make the resulting numbers
interoperable.

This page registers the methodological relationship. It does not create a
common leaderboard or claim that a Pokémon deck, a mathematical hypothesis and
a research model occupy one latent scale.

## Comparison map

| Project | Rated object | Initial or anchor convention | Evidence boundary |
|---|---|---|---|
| Pokémon TCG AI Battle | Agents, decks and cards; submission/model lineages remain partly planned | Current database bootstrap is 600; local and remote evidence are separated | Live SQLite tables, tournament cohorts and `ResultsDB`; invariant-deck calibration is experimental and source/cohort dependent |
| Antigravity 2.0 | Mathematical hypotheses compared with published or human reference results | Research page records a 1200 reset and a hypothesis-oriented scale | Handoff/research ledger; Elo expresses relative evaluation, not proof of a theorem |
| Co-Scientist | Candidate research systems or hypotheses in head-to-head evaluation | Its page defines its own initial scale, K-factor choices, match IDs and pair selection | Gold-set and comparison protocol defined by the Co-Scientist project |
| Red Team Arena | Safety-system evaluation pools | Separate arena conventions and rating pools | Attack, defense and judge outcomes; not a substitute for the other project ledgers |

## What transfers between projects

The following abstractions transfer safely:

- every rating needs an explicit object, population and update rule;
- a rating is relational evidence, not an intrinsic essence of the rated object;
- resets, source partitions and sampling change comparability;
- deterministic match identifiers and provenance make repeated comparisons
  auditable;
- a rating table should be accompanied by uncertainty, cohort and evidence
  status;
- a strong score does not prove the mechanism or theory that produced it.

These principles connect the Pokémon TCG database work to the broader
evaluation methodology in [[co-scientist-elo-tournament]],
[[antigravity-2.0]] and [[red-team-arena-elo]].

## What does not transfer

The following operations would be invalid without a new calibration study:

- averaging scores across projects;
- treating the 600 bootstrap in Pokémon TCG as equivalent to a 1200 reset in
  Antigravity or Co-Scientist;
- comparing K-factors or draw handling without matching the update equations;
- interpreting a deck's invariant rating as a model or submission rating;
- using a PageRank score as an Elo score merely because both are rank vectors;
- using a tournament result as proof of the proposed RoPEND, MoE or
  PageRank-Abelian architecture when the corresponding runtime is absent.

## Pokémon TCG as the detailed case

The Pokémon TCG work shows why this distinction matters. The repository has
ordinary source-separated daily ratings, an invariant-deck calculation that
combines a clipped win-rate inversion, MD10 shrinkage and a local-to-remote
translation, and a research proposal connecting translation invariance to a
sampled graph. These are three related layers, not one finished metric.

The current implementation also exposes a draw-handling difference: the
invariant-deck path uses `wins / games_played`, while the ordinary online update
uses a 0.5 score for a draw. A future comparison must state which path was used
before interpreting a number. See [[pokemon_tcg_submissions_and_elo]] and
[[pokemon_tcg_elo_pagerank_invariance]].

## Research use

For future experiments, record the following tuple with every Elo-like result:

```text
project / rated object / source / cohort or reset / update rule /
match identifier scheme / sample size / uncertainty / claim class
```

The claim class should distinguish an observed result, a code-validated
calculation, a source-reported handoff result, an inference and a proposal.
This is especially important when an orchestration transcript or a manuscript
uses completion language that the live source tree does not confirm.

## Primary sources

- [[pokemon_tcg_elo_pagerank_invariance]] — Pokémon TCG rating and graph bridge.
- [[pokemon_tcg_submissions_and_elo]] — source-separated rating lineages and current implementation boundary.
- [[co-scientist-elo-tournament]] — Co-Scientist comparison protocol.
- [[antigravity-2.0]] — Antigravity hypothesis-ranking protocol.
- [[red-team-arena-elo]] — safety-evaluation rating pools.

## Cross-references

- [[pokemon_tcg_orchestration_artifact_ledger]] — handoff artifacts and acceptance gates.
- [[pokemon_tcg_cross_harness_provenance]] — transcript and cross-harness provenance.
- [[multi-agent-methodology]] — broader agent-evaluation and review methodology.