Truth Resolution Agent: A Multi-Source Judicial Framework for Sports Disputes (Senna 1989 Case Study)#

Fabricio Ceolin

The Edge Agent Project

Abstract#

Sports disputes often involve conflicting narratives, incomplete evidence, and subjective interpretations that make fair adjudication challenging. This article presents a Truth Resolution Agent built with The Edge Agent (TEA) framework, designed to systematically evaluate evidence against established regulations and produce reasoned verdicts. We demonstrate the agent through a detailed analysis of the 1989 Japanese Grand Prix incident between Ayrton Senna and Alain Prost—one of the most controversial decisions in motorsport history.

We present two implementations:

LLM-only approach - Pure language model reasoning through structured nodes
Neurosymbolic approach - Combining LLM for fact extraction with Prolog for deterministic judicial reasoning

Both agents reach the same verdict: NOT GUILTY, with the original disqualification deemed unjust based on regulatory ambiguity, inconsistent enforcement, and lack of competitive advantage. The neurosymbolic version provides an additional benefit: deterministic, auditable reasoning where the same facts always produce the same verdict with a complete trace of which rules fired.

Keywords: Truth Resolution, Sports Arbitration, TEA, State Graph, Formula 1, Evidence-Based Judgment, AI Adjudication, Neurosymbolic AI, Prolog

1. Introduction#

On October 22, 1989, a two-second collision at the Casio Triangle chicane in Suzuka sparked 35 years of debate. When Ayrton Senna and Alain Prost’s McLarens interlocked wheels on Lap 47, it set in motion one of the most controversial judicial decisions in sports history.

Senna was disqualified for “cutting the chicane” after marshals pushed his stalled car and he rejoined via the escape road. The decision handed the World Championship to Prost and remains disputed to this day—particularly after FISA President Jean-Marie Balestre’s 1996 admission that he acted to favor his fellow countryman Prost.

This case presents an ideal test for automated judicial reasoning:

Complex regulations with ambiguous provisions
Disputed facts requiring careful evidence weighing
Precedent analysis showing inconsistent enforcement
Political interference allegations requiring impartial reassessment

Can an AI agent, programmed with principles of judicial fairness, reach a just verdict when human institutions failed?

2. The Truth Resolution Problem#

2.1 Why Sports Arbitration is Hard#

Sports disputes share characteristics with legal proceedings but operate under unique constraints:

Challenge	Description
Time pressure	Decisions often needed within hours of an incident
Incomplete evidence	No discovery process; rely on available footage and testimony
Regulatory gaps	Rules written for common cases; edge cases expose ambiguity
Political pressure	Governing bodies may have conflicts of interest
Precedent inconsistency	Similar cases treated differently across events

2.2 The Case for Structured AI Adjudication#

An AI judicial agent can provide:

Consistency - Apply the same standards to every case
Transparency - Document every step of reasoning
Impartiality - No nationality bias, political pressure, or personal relationships
Precedent analysis - Systematically compare with historical cases
Proportionality assessment - Evaluate if penalties fit the offense

3. Agent Architecture#

The Truth Resolution Agent uses a structured graph approach where each phase of the judicial process is a separate node. This architecture mirrors real judicial proceedings:

        flowchart TD
    START[__start__] --> AR[analyze_regulations]
    START --> AE[analyze_evidence]
    START --> AP[analyze_precedents]

    AR --> AV[analyze_violations]
    AE --> AV

    AP --> AC[assess_consistency]
    AV --> AC

    AC --> D[deliberate]
    D --> V[verdict]
    V --> ES[extract_summary]
    ES --> END[__end__]

3.1 Node Descriptions#

Node	Purpose	Output
`analyze_regulations`	Parse and structure applicable rules	Identified rules, triggers, penalties, ambiguities
`analyze_evidence`	Categorize evidence by reliability	Undisputed facts, disputed facts, documentary evidence
`analyze_precedents`	Structure historical cases	Similar cases, outcomes, enforcement patterns
`analyze_violations`	Map evidence to rule violations	Actus reus, mens rea, defenses, mitigating factors
`assess_consistency`	Check enforcement consistency	Consistency assessment, selective enforcement flags
`deliberate`	Core judicial reasoning	Detailed analysis, preliminary verdict
`verdict`	Formal verdict document	Findings, analysis, verdict, penalty, dissent
`extract_summary`	Structured verdict extraction	Final verdict and penalty

3.2 Judicial Principles Encoded#

The agent’s system prompts encode fundamental principles of fair adjudication:

# Core principles from the deliberate node
1. STRICT LEGALITY
   - Apply ONLY regulations in force at the time
   - Ambiguity interpreted in favor of the accused

2. EVIDENCE-BASED
   - Burden of proof on the ACCUSER
   - If evidence insufficient, rule NOT GUILTY

3. ABSOLUTE IMPARTIALITY
   - Ignore nationality, politics, championship implications

4. PROPORTIONALITY
   - Penalty must fit the offense
   - Consider intent vs accident, advantage gained

5. CONSISTENCY WITH PRECEDENT
   - Similar cases MUST be treated similarly
   - Inconsistent enforcement is a STRONG defense

4. The 1989 Japanese Grand Prix Case#

4.1 Regulatory Framework#

The agent was provided the 1989 FIA regulations:

Article 56.2 (Race Distance)

“The winner of the Grand Prix shall be the driver who completes the scheduled race distance in the shortest time.”

Appendix J, Article 2.3 (Marshal Assistance)

“Marshals may assist ONLY to move the car to a position of safety.”

Interpretation Note (acknowledged by FIA)

“It is ambiguous whether a car moved for safety reasons may rejoin the race if the engine restarts during the movement.”

4.2 Evidence Summary#

Undisputed Facts:

Lap 47 of 53, Casio Triangle chicane
Collision caused both cars to stall
Marshals pushed Senna’s car for safety
Engine bump-started during push
Senna used escape road, bypassed chicane
Senna lost ~30 seconds (pit stop for new wing)
Senna finished first on track
Senna disqualified; Nannini declared winner

Disputed Facts:

Who caused the collision (Prost turned in early vs. Senna’s move was optimistic)
Whether engine was fully stalled before push

4.3 Precedent Cases#

The most significant precedent: Imola 1989

Factor	Imola 1989	Suzuka 1989
Nature	Voluntary chicane cutting	Emergency (forced by collision)
Advantage	Yes (avoided slowdown)	None (30+ seconds lost)
Penalty	None	Disqualification
Drivers affected	Multiple including Prost	Only Senna

Enforcement Pattern (1979-1989):

26 total disqualifications in 10 years
0 for chicane cutting
Senna’s disqualification was unprecedented

5. Agent Implementation#

5.1 YAML Agent Definition#

The agent is defined in examples/truth-resolution-agent.yaml:

name: truth-resolution-agent
version: "1.0.0"
description: |
  An impartial judicial agent for resolving sports disputes.

state_schema:
  # Case identification
  case_id: str
  case_title: str
  date_of_incident: str
  jurisdiction: str

  # Raw inputs
  raw_regulations: str
  raw_evidence: str
  raw_precedents: str
  party_a_name: str
  party_a_arguments: str
  party_b_name: str
  party_b_arguments: str

  # Processed outputs
  regulations_analysis: str
  evidence_analysis: str
  precedents_analysis: str
  violation_analysis: str
  consistency_assessment: str
  deliberation_notes: str
  verdict: str
  verdict_reasoning: str

nodes:
  - name: analyze_regulations
    uses: llm.call
    with:
      provider: shell
      shell_provider: claude
      messages:
        - role: system
          content: |
            You are a legal document analyst...
        - role: user
          content: |
            JURISDICTION: {{ state.jurisdiction }}
            REGULATIONS TEXT: {{ state.raw_regulations }}
    output:
      regulations_analysis: "{{ result.content }}"

  # ... additional nodes for evidence, precedents, violations,
  # consistency, deliberation, and verdict

5.2 Case Input Structure#

Cases are provided as YAML files with structured fields:

case_id: "FIA-1989-JPN-001"
case_title: "Senna vs Race Stewards - 1989 Japanese Grand Prix"
date_of_incident: "1989-10-22"
jurisdiction: "FIA"

party_a_name: "Ayrton Senna (McLaren-Honda)"
party_a_arguments: |
  1. FORCE MAJEURE - Collision was shared fault
  2. NO ADVANTAGE GAINED - Lost 30+ seconds
  3. INCONSISTENT ENFORCEMENT - Imola precedent
  ...

party_b_name: "FIA Race Stewards"
party_b_arguments: |
  1. ARTICLE 56 VIOLATION - Failed to complete distance
  2. ILLEGAL PUSH START - Outside assistance
  ...

raw_regulations: |
  ARTICLE 56.2: The winner shall complete the scheduled
  race distance in the shortest time...

raw_evidence: |
  UNDISPUTED FACTS:
  1. Lap 47 of 53, Casio Triangle chicane
  ...

raw_precedents: |
  CASE 1: IMOLA 1989 - Multiple drivers cut chicane,
  no penalties including Prost...

5.3 Running the Agent#

# Convert case YAML to JSON
python3 -c "import yaml, json; print(json.dumps(yaml.safe_load(open('case.yaml'))))" > case.json

# Execute agent
tea run examples/truth-resolution-agent.yaml --input @case.json -v

6. Neurosymbolic Architecture#

The neurosymbolic version adds a critical capability: deterministic, auditable reasoning using Prolog for judicial logic.

6.1 Why Prolog for Judicial Reasoning?#

Benefit	Description
Determinism	Same facts always produce the same verdict
Auditability	Complete trace of which rules fired and why
Consistency	Automatically detects precedent inconsistencies
Transparency	Every conclusion traceable to a specific rule

6.2 Neurosymbolic Flow#

        flowchart TD
    START[__start__] --> EF[extract_facts<br/>LLM: NL → Prolog facts]
    EF --> PF[parse_facts<br/>Clean Prolog syntax]
    PF --> PR[prolog_reasoning<br/>Deterministic inference]
    PR --> GV[generate_verdict<br/>LLM: Prolog → Human text]
    GV --> ES[extract_summary]
    ES --> END[__end__]

    style PR fill:#f9f,stroke:#333,stroke-width:2px

6.3 Prolog Judicial Knowledge Base#

The neurosymbolic agent encodes judicial principles as Prolog rules:

% Defense 1: Rule Ambiguity (principle of strict legality)
valid_defense(RuleId, ambiguity) :-
    rule_ambiguity(RuleId, _).

% Defense 2: Force Majeure
valid_defense(RuleId, force_majeure) :-
    actus_reus(RuleId, Party, _),
    force_majeure(Party, _).

% Defense 3: No Advantage Gained (proportionality)
valid_defense(RuleId, no_advantage) :-
    actus_reus(RuleId, Party, _),
    no_advantage_gained(Party, _).

% Defense 4: Third Party Causation
valid_defense(RuleId, third_party) :-
    actus_reus(RuleId, Party, _),
    third_party_action(Party, _, _).

% Defense 5: Inconsistent Enforcement (strongest defense)
valid_defense(RuleId, inconsistent_enforcement) :-
    inconsistent_enforcement(RuleId).

% Verdict is NOT_GUILTY if any valid defense exists
verdict_for_rule(RuleId, not_guilty, Defense) :-
    actus_reus(RuleId, _, _),
    valid_defense(RuleId, Defense), !.

% Verdict is GUILTY only if no defense applies
verdict_for_rule(RuleId, guilty, no_defense) :-
    actus_reus(RuleId, _, _),
    \+ valid_defense(RuleId, _).

6.4 LLM-Extracted Facts (Prolog Format)#

The LLM extracts structured facts from natural language evidence:

% Rules extracted from 1989 FIA regulations
rule(article_56, 'Race distance completion', chicane_cut, disqualification).
rule(appendix_j_2_1, 'Prohibition of outside assistance', outside_assistance, disqualification).
rule(appendix_j_2_3, 'Marshal assistance for safety only', outside_assistance, disqualification).

% Ambiguity in regulations (critical for defense)
rule_ambiguity(appendix_j_2_3, 'Unclear if car may rejoin if engine restarts during safety move').

% Evidence of alleged violations
actus_reus(article_56, senna, 'Used escape road to rejoin after collision').
actus_reus(appendix_j_2_3, senna, 'Engine restarted during marshal push').

% Defense evidence
force_majeure(senna, 'Collision with Prost was racing incident').
no_advantage_gained(senna, 'Lost 30+ seconds for pit stop and wing replacement').
third_party_action(senna, marshals, 'Initiated push to move car to safety').
time_lost(senna, 30).

% Precedents showing inconsistent enforcement
precedent(imola_1989, chicane_cut, not_penalized, 1989).
precedent(monaco_1988, chicane_cut, not_penalized, 1988).

7. The Verdict#

After processing all evidence, regulations, and precedents, both agents delivered the same verdict through different reasoning approaches.

7.1 Comparison: LLM-Only vs Neurosymbolic#

Aspect	LLM-Only	Neurosymbolic
Verdict	NOT GUILTY	NOT GUILTY
Reasoning approach	Natural language analysis	Prolog inference + NL explanation
Determinism	May vary between runs	Always same result for same facts
Audit trail	Textual reasoning	Prolog trace + textual
Defenses identified	Ambiguity, inconsistent enforcement, no advantage	ambiguity, force_majeure, no_advantage, third_party, inconsistent_enforcement
Best for	Nuanced cases requiring human-like judgment	Cases requiring strict consistency and auditability

7.2 Summary#

Element	Finding
Verdict	NOT GUILTY
Penalty	No penalty (disqualification overturned)
Key reasoning	Regulatory ambiguity, inconsistent enforcement, no advantage gained

7.3 Analysis of Alleged Violations (LLM-Only Agent)#

Article 56.2 - Race Distance:

“The regulation speaks to distance, not trajectory. The escape road may add distance rather than reduce it… The escape road is a homologated feature of the FIA-approved circuit, designed for emergency use.”

Finding: Prosecution failed to establish that emergency use of an escape road constitutes failure to complete race distance.

Articles 2.1/2.3 - Outside Assistance:

“The marshals initiated the contact, not Senna. They were performing their explicit regulatory duty: ‘move the car to a position of safety.’… The prosecution’s own evidence includes the Interpretation Note acknowledging ambiguity.”

Finding: Safety-initiated marshal assistance that incidentally restarts an engine does not constitute prohibited outside assistance.

7.4 Consistency Assessment#

The agent identified INCONSISTENT enforcement:

“The Imola 1989 precedent demonstrates that similar or clearer violations went unpunished in the same season… Voluntary chicane-cutting warranted no penalty while emergency escape road use warranted disqualification. This disparity constitutes selective enforcement.”

7.5 Proportionality#

“Disqualification is the maximum penalty. It should be reserved for clear, intentional violations conferring competitive advantage. This case presents ambiguous rule application, emergency circumstances, no competitive advantage (30+ seconds lost), marshal-initiated assistance, and inconsistent enforcement history.”

7.6 Dissenting Opinion#

The agent acknowledged counter-arguments:

“A push did occur. Marshals physically moved Senna’s car, and his engine restarted. The escape road was used. However, these arguments fail because the regulations acknowledge ambiguity, emergency circumstances distinguish this case, no advantage was gained, and inconsistent enforcement undermines strict application.”

7.7 Neurosymbolic Verdict (Prolog Trace)#

The neurosymbolic agent produces a deterministic verdict with a complete reasoning trace:

PROLOG REASONING RESULTS (DETERMINISTIC)
═══════════════════════════════════════════════════════════════════════

VERDICT: not_guilty

DEFENSES FOUND:
[defense(article_56, ambiguity),
 defense(article_56, force_majeure),
 defense(article_56, no_advantage),
 defense(article_56, third_party),
 defense(article_56, inconsistent_enforcement),
 defense(appendix_j_2_3, ambiguity),
 defense(appendix_j_2_3, force_majeure),
 defense(appendix_j_2_3, no_advantage),
 defense(appendix_j_2_3, third_party)]

CONSISTENCY CHECK (Inconsistent Enforcement):
[inconsistent(article_56)]

FULL REASONING TRACE:
Rules examined: [rule(article_56), rule(appendix_j_2_3)]
Defenses found: 9 defenses across 2 rules
Inconsistent enforcement detected: [inconsistent(article_56)]
Verdicts: [verdict(article_56, not_guilty, ambiguity),
           verdict(appendix_j_2_3, not_guilty, ambiguity)]

Key insight: The Prolog reasoner found 9 valid defenses across the two alleged rule violations. With multiple defenses available, the verdict is deterministically NOT GUILTY. The same facts will always produce the same verdict.

8. Generalization#

8.1 Adapting for Other Domains#

The Truth Resolution Agent can be adapted for:

Domain	Adaptations Needed
Soccer/Football	VAR review, offside rules, handball regulations
Olympic Sports	Doping cases, judging disputes, qualification rules
Esports	Tournament rules, cheating allegations, bracket disputes
Contract Disputes	Agreement terms, breach definitions, damage calculations

8.2 Key Customization Points#

Regulations - Replace with domain-specific rules
Evidence categories - Adapt to available evidence types
Precedent database - Build historical case repository
Judicial principles - Adjust for domain conventions

8.3 Limitations#

Cannot interview witnesses - Relies on provided testimony
No physical evidence examination - Works with descriptions only
Regulatory interpretation - May differ from official interpretations
No enforcement power - Provides recommendations only

9. Conclusion#

The Truth Resolution Agent demonstrates that structured AI reasoning can provide valuable judicial analysis for sports disputes. By encoding fundamental principles of fairness—strict legality, evidence-based reasoning, impartiality, proportionality, and consistency—the agent reached a verdict that addresses the historic injustice of the 1989 Japanese Grand Prix.

Key findings:

Senna should not have been disqualified - Regulatory ambiguity, inconsistent enforcement, and lack of competitive advantage all favor the defense
Structured graphs enable transparent reasoning - Each step of the judicial process is documented and auditable
AI can supplement human judgment - Particularly valuable when political interference or bias may affect human decision-makers
The Imola precedent is dispositive - If Prost and others cut chicanes without penalty at Imola, Senna’s emergency use of an escape road at Suzuka cannot warrant disqualification

The agent’s verdict aligns with the historical consensus among motorsport analysts and Max Mosley’s later admission that Balestre “fixed the whole thing.” After 35 years, justice—at least in principle—can finally be served.

10. References#

Appendix A: Running the Examples#

# Clone the repository
git clone https://github.com/fabceolin/the_edge_agent.git
cd the_edge_agent

# Install dependencies
pip install -e python/[dev]

# Run the LLM-only agent
tea run examples/truth-resolution-agent.yaml \
    --input @examples/cases/suzuka-1989-senna-prost.yaml -v

# Run the Neurosymbolic agent (with Prolog reasoning)
tea run examples/truth-resolution-agent-neurosymbolic.yaml \
    --input @examples/cases/suzuka-1989-senna-prost.yaml -v

Appendix B: Agent YAML Files#

Agent	Description	GitHub Link
LLM-Only	Pure language model reasoning	`examples/truth-resolution-agent.yaml`
Neurosymbolic	LLM + Prolog deterministic reasoning	`examples/truth-resolution-agent-neurosymbolic.yaml`

Appendix C: Case File#

The complete case file with regulations, evidence, precedents, and party arguments:

examples/cases/suzuka-1989-senna-prost.yaml

Appendix D: Choosing Between LLM-Only and Neurosymbolic#

Use Case	Recommended Approach
Cases requiring strict auditability	Neurosymbolic - complete Prolog trace
Cases with clear rule structure	Neurosymbolic - encode rules as Prolog
Novel cases without precedent	LLM-only - better at reasoning by analogy
Regulatory compliance	Neurosymbolic - deterministic, repeatable
Nuanced factual disputes	LLM-only - better at weighing evidence
High-stakes decisions	Neurosymbolic - explainable to stakeholders

Truth Resolution Agent: A Multi-Source Judicial Framework for Sports Disputes (Senna 1989 Case Study)

Contents