Interpretive truth: Verifiably constrained AI judges for arbitration at scale

manifest.json
{
  "name": "football-player-value-v1",
  "version": "2.1.0",
  "applicableTo": ["player_value", "player_ranking"],
  "description": "Tiered evidence dossier: primary on/off splits +
                  team-share metrics, secondary independent scout
                  analysis, tertiary decorative quotes.",
  "model": {
    "id": "zai-org/GLM-4.7-FP8",
    "sampling": {
      "temperature": 0,
      "topP": 1.0,
      "seed": 0,
      "maxTokens": 2000
    }
  },
  "evidenceSchema": "schemas/dossierV1.json",
  "outputSchema": {
    "outcome":    { "enum": [0, 1, 2] },
    "confidence": { "type": "number", "min": 0, "max": 1 },
    "rationale":  { "type": "string", "maxLength": 2400 },
    "citations":  { "type": "array", "items": "string" }
  },
  "promptTemplate": {
    "system":       "framework.md",
    "userTemplate": "Question: {{question}}\n\nDossier:\n{{evidence_json}}"
  }
}

framework.md
# football-player-value-v1  (v2.1.0)

"Value to club" is not the same as "transfer market value." This is the
interpretive heart of the question: reasonable analysts disagree on the
weighting. The framework declares its stance up front.

Three dimensions, in priority order — they compound, they do not add:

  1. Productive output
       goals, assists, chances created, defensive contribution,
       save percentage. Position-adjusted, per-90.

  2. Irreplaceability
       on/off team output. Structural dependence ("the system runs
       through them"). Absence of like-for-like alternatives.

  3. Ceiling and durability
       age, contract length, injury history. Realistic horizon of
       contribution, not just one match.

## Evidence hierarchy   (new in v1.3.0)

Tier 1 (primary, weight heavily)
  on/off splits, substitution patterns, team-share metrics,
  big-game start rate, trophies/standings differential.
  Things the player cannot author through PR.

Tier 2 (secondary, weight moderately)
  independent scout notes, peer-club bids actually received,
  salary % of total wage bill (what the wallet says, not the mouth).

Tier 3 (decorative, low weight)
  manager quotes, fan sentiment, media narratives.
  Included for colour, never load-bearing.

When Tier 1 and Tier 3 disagree, Tier 1 wins.

How to weigh signals:
  - In-form play > season aggregates
  - Transfer-market value is a sanity check, not a primary signal
  - Long-form context_notes are load-bearing — read them

Edge cases — when signals conflict:
  - Stats vs. narrative → prefer the side that explains more of the dossier
  - Output vs. system player → lean system player when on/off splits agree
  - Established vs. emerging → weight present output over projection

Hard rules:
  - Do not import facts outside the dossier
  - If a referenced player isn't described, return outcome: 2
  - If confidence falls below 0.55, prefer outcome: 2
  - If reasoning leans on Tier 3, cap confidence at 0.65

Return a single JSON object:

  {
    "outcome":    1,
    "confidence": 0.72,
    "rationale":  "2–4 sentence explanation, declaring which tier
                   drove the call.",
    "citations":  ["Pedri.team_share", "Pedri.on_off_splits", …]
  }

Citation conventions:
  - Cite per subject. Include at least one stats path per subject.
  - Order: team_share / on_off_splits → stats → match_reports
    → scout_notes → manager_quotes → context_notes
  - Aim for 8–16 citations — confidence signals breadth, citations
    signal weight. More citations when multiple Tier 1 fields
    contribute orthogonally.

Keep the rationale terse — it is part of the on-chain verdict and
will be replayed for verification. Extra prose changes the hash.

verdict.json
{
  "outcome":    1,
  "confidence": 0.74,
  "rationale":  "Tier 1 evidence is decisive. Haaland accounts for 41%
                 of City's PL goals vs Mbappé's 33% at Madrid, and City's
                 scoring rate drops ~40% in Haaland's off-minutes vs
                 Madrid's ~10% in Mbappé's. Tier 2 scout analysis
                 (McGuire, Lowe) agrees. Manager quotes from both sides
                 are consistent but not load-bearing under the framework — Pep
                 and Arbeloa have direct conflicts of interest.",
  "citations": [
    "Erling Haaland.team_share",
    "Kylian Mbappe.team_share",
    "Erling Haaland.on_off_splits",
    "Kylian Mbappe.on_off_splits",
    "Erling Haaland.substitution_patterns",
    "Kylian Mbappe.substitution_patterns",
    "Erling Haaland.contract_signals",
    "Kylian Mbappe.contract_signals",
    "Erling Haaland.stats.season.notes",
    "Kylian Mbappe.stats.season.notes",
    "Erling Haaland.scout_notes[0]",
    "Kylian Mbappe.scout_notes[0]",
    "Erling Haaland.match_reports[0]",
    "Kylian Mbappe.match_reports[1]",
    "Kylian Mbappe.scout_notes[1]",
    "Erling Haaland.manager_quotes[1]"
  ]
}

pinned inference

{
  "model":    "zai-org/GLM-4.7-FP8",
  "sampling": {
    "temperature": 0,
    "topP":        1.0,
    "seed":        0,
    "maxTokens":   2000
  }
}
// pinned model + sampling, executed by the
// Ritual L1 LLM precompile inside a TEE —
// the attested executor key signs the verdict.