World Cup 2026 Project — Glossary

Terms used throughout the pipeline. Football-specific, project-specific, and pipeline-specific jargon.

Pipeline & Architecture

MD1 / MD2 / MD3

Matchday 1/2/3 — the three rounds of group-stage fixtures. Each team plays 3 matches (one per matchday). MD3 matches within a group kick off simultaneously to prevent collusion.

Dead Rubber

A match where neither team has anything competitive at stake — both already eliminated, both already qualified, or the result can't change tournament position. Affects predictions: motivation is uncertain, lineups rotate, intensity drops, Finding 12 (conservation) behaves differently.

RAPID Analysis

A no-gathering ensemble — writers fire from MD1 form only, no live gathering (scout/injury/referee), no T-60 confirmed lineups, no match-day weather. Used when T-8 or less to kickoff (full pipeline can't fit). All XIs flagged PROJECTED-UNCONFIRMED. Validated: reliable for clear favorites (>65%) + generational scorers; unreliable for coin-flips.

T-60 Workflow

The standard pre-match timing: gather early (T-24h+), then pull confirmed XIs ~60 minutes before kickoff, then fire the ensemble. Ensures lineups are confirmed, not projected.

TASK-99 (Independent Resolution)

Pipeline change (Jun 23, 2026): each match-writer resolves the 6-step decision graph (doc-5) independently, instead of the lead pre-resolving it and passing a synthesis block. Produces honest spreads (3-7pt) instead of artificial 1pt convergence. Pre-TASK-99 per-bot stats were noise (all 3 anchored on lead's synthesis); post-TASK-99 they're genuine signal.

TASK-97 (Post-Event Writer + Rubric)

Pipeline change (Jun 23, 2026): a post-event-writer agent + 100-point calibration rubric (doc-1) replaces the lead's vibe-based ~X% scoring. Components: Winner (0-25), Scoreline (0-25), Who Scores (0-20), Findings (0-20), Model-breaking (0-10), with −5 deductions. Bands: A (90+), B (75-89), C (60-74), D (45-59), F (<45).

Agents (7 total)

Gathering Layer (run in PARALLEL)

Synthesis Layer (run in PARALLEL, 3 models for ensemble)

Post-Event Layer

The 22 Calibration Findings (doc-1)

The empirical rules validated across the tournament. Key ones referenced frequently:

Key Concepts

Modal Scoreline

The single most-likely exact score (highest probability in the distribution). E.g., "POR 2-0 at 17%." Hitting the exact modal is rare (~20% of matches) and scores 25/25 on the rubric.

Who Scores #1

The top-ranked scorer pick for each team, with a probability %. The pipeline's scorer predictions are one of its strongest signals — especially for generational milestone-chasing players (Messi, Mbappé, Haaland, Ronaldo all braced when #1 unanimous).

Finding Resolution

Each writer walks doc-5's 6-step decision graph and explicitly states whether each Finding fires (applies), doesn't fire, or fires partially — with reasoning. This is the analytical engine, not just number-crunching.

Spread (Win% Spread)

The range of favorite win% across the 3 models. Pre-TASK-99: 1-2pt (artificial anchoring). Post-TASK-99: 3-7pt (honest uncertainty). The widest-spreading (most defensive) agent tends to be closest on upset/draw matches.

Category 0 Gate

The no-fabrication hard rule: the final score MUST be confirmed from ≥2 independent sources (BBC/Guardian/FotMob/ESPN) before it enters any review or summary. A single stale live-blog is NOT sufficient (CZE-RSA was mis-reported as 1-0; actual was 1-1).

Project Infrastructure

Canonical Profiles (sources/reference/)

Stable, shared intelligence updated per matchday — NOT contamination:

• teams/{code}.md — squad, manager, tactical signature, GK, form
• referees/{surname}.md — career profile + assignments log
• venues/{city-slug}.md — roof, capacity, altitude, climate-control

Per-Match Directory (sources/matches/{code}/)

Thin delta — match-specific only:

• scout/, team-news-{t1}/, team-news-{t2}/, referee/, weather/, lineups/, post-event/, archive-superseded/

RAG (Retrieval-Augmented Generation)

Local semantic search database (rag/). 800+ chunks embedded via ollama nomic-embed-text (768d). Search: node .pi/skills/mundial-rag/search.mjs "query". Complements, not replaces, direct profile reading.

TOML Metrics Block

Machine-parseable block at the end of every analysis file ([match], [win_probability], [scorelines], [who_scores], [calibration], [predicted_stats]). Auto-extracted by rag/toml-extract.mjs and agent-accuracy.mjs.

Calibration Rubric (doc-1)

100-point scoring system for post-event reviews:

Bands: A (90+, exceptional), B (75-89, good), C (60-74, mixed), D (45-59, weak), F (<45, miss).

File Naming Conventions

Last updated: June 25, 2026. Living document — updated as the pipeline evolves and new findings are validated.