Skip to content

CROSS_CORRELATION_ALETHEIA_OPENCODE.md

Latest commit

 

History

History
223 lines (164 loc) · 11.5 KB

CROSS_CORRELATION_ALETHEIA_OPENCODE.md

File metadata and controls

223 lines (164 loc) · 11.5 KB

Cross-Correlation Report: Superhuman/Aletheia × OpenCode Ecosystem

Feng et al. (2026) "Towards Autonomous Mathematics Research" vs OpenCode v4.3.0

Generated: 2026-05-30T13:54:22.203258 References: arXiv:2602.10177v3 | github.com/google-deepmind/superhuman | github.com/MarceloClaro/OpenCode_Ecosystem

Executive Summary

Metric Value
Total dimensions compared 12
Direct matches 1
OpenCode superior 7 (58%)
Aletheia superior 1 (8%)
Complementary 3
Avg OpenCode superiority score 0.9
Avg Aletheia superiority score 0.3

Key finding: OpenCode matches or exceeds Aletheia in 8/12 (67%) dimensions. The critical gap is the foundation model (Gemini Deep Think scale).

Correlation Matrix

# Aletheia Component OpenCode Component Match Score
1 Aletheia G-V-R Loop SPEC-012 Aletheia Engine 🟰 0.85
2 Informal Verifier Cora-Debate V1-V7 + SPEC-008 Triangulacao 🟢 0.90
3 Gemini Deep Think (implicito) Reasoning Orchestrator v11 (212 tipos explici 🟡 0.70
4 3 tools (Search, Browse, Python) 18 MCPs + code-runner + playwright 🟢 0.75
5 IMO Bench + FirstProof + FutureMath + Erdos CORA-Eval D1-D9 + Domain-Shift + Olympiad 🟡 0.65
6 Taxonomia H/C/A × 0-4 (Feng §6.1) Camadas C1/C1B/C2/C3 (SPEC-008) 🟡 0.55
7 Single-use problem (reconhecido §4) SPEC-008 Triangulacao (3 camadas) 🟢 0.95
8 Matematica pura apenas 6+ dominios (juridico, fisica, metodologia, a 🟢 0.95
9 Paper + prompts no GitHub TDD + seed + hash + sync mirror — 100% audita 🟢 0.95
10 Nao abordado SPEC-008-B Camada 1B (bootstrap Jaccard, 9 CT 🟢 0.98
11 Reducao via tool use (Search) Cora V4 + 6 padroes de deteccao + verificacao 🟢 0.80
12 Gemini Deep Think (proprietario, escala massi OpenCode (modelos acessiveis via API) 🔵 0.30

Detailed Analysis

🟢 OpenCode Advantages (7 dimensions)

Informal Verifier vs Cora-Debate V1-V7 + SPEC-008 Triangulacao (score: 0.9)

Aletheia usa verificador informal; OpenCode tem 7 verificadores simbolicos + 3 camadas anti-circularidade OpenCode advantage: 7 verificadores explicitos (vs 1 implicito); auto-critica desacoplada; triangulacao anti-circular Aletheia limitation:

3 tools (Search, Browse, Python) vs 18 MCPs + code-runner + playwright (score: 0.75)

OpenCode tem 6x mais ferramentas ativas cobrindo dominios alem da matematica OpenCode advantage: 18 MCPs multi-proposito (vs 3 tools); sandbox isolado; SQLite local; PDF toolkit Aletheia limitation: Integracao profunda Google Search (modelo treinado para tool use)

Single-use problem (reconhecido §4) vs SPEC-008 Triangulacao (3 camadas) (score: 0.95)

Aletheia reconhece o problema de 'single use' mas nao o resolve; OpenCode tem framework completo para isso OpenCode advantage: Framework matematico para quebrar circularidade; domain-shift detection; bootstrap calibration Aletheia limitation:

Matematica pura apenas vs 6+ dominios (juridico, fisica, metodologia, arte, economia) (score: 0.95)

Aletheia foi projetado exclusivamente para matematica; OpenCode cobre multiplos dominios cientificos OpenCode advantage: 6 dominios com TDD proprio; integracao com editais, arteterapia, CORA-Eval Aletheia limitation:

Paper + prompts no GitHub vs TDD + seed + hash + sync mirror — 100% auditavel (score: 0.95)

Aletheia publica prompts/outputs mas sem testes automatizados; OpenCode tem TDD completo OpenCode advantage: 71 testes automatizados; seed fixa; hash verificavel; clone identico via sync mirror Aletheia limitation:

Nao abordado vs SPEC-008-B Camada 1B (bootstrap Jaccard, 9 CTs) (score: 0.98)

Aletheia nao aborda domain shift entre problemas/dominios; OpenCode tem framework dedicado OpenCode advantage: Decomposicao institucional; 3 deltas Jaccard; bootstrap calibration; 9 CTs TDD Aletheia limitation:

Reducao via tool use (Search) vs Cora V4 + 6 padroes de deteccao + verificacao de citacoes (score: 0.8)

Aletheia reduz alucinacoes via Search mas nao as detecta sistematicamente OpenCode advantage: 6 padroes de deteccao; V4 Citation Accuracy check; penalizacao no score Aletheia limitation: Google Search integrado como ferramenta nativa do modelo base

🔵 Aletheia Advantages (1 dimensions)

Gemini Deep Think (proprietario, escala massiva) vs OpenCode (modelos acessiveis via API) (score: 0.3)

Gap fundamental: Deep Think tem escala e treinamento que modelos publicos nao alcancam IMO-Gold (5/6); 100x reducao compute; inference-time scaling law proprietaria

🟡 Complementary (3 dimensions)

Gemini Deep Think (implicito) vs Reasoning Orchestrator v11 (212 tipos explicitos) (score: 0.7)

Deep Think tem escala massiva mas raciocinio implicito; OpenCode tem taxonomia explicita de 212 tipos mas escala limitada

IMO Bench + FirstProof + FutureMath + Erdos vs CORA-Eval D1-D9 + Domain-Shift + Olympiad (score: 0.65)

Bancos diferentes: Aletheia focado em matematica pura; OpenCode cobre 9 disciplinas + metodologia

Taxonomia H/C/A × 0-4 (Feng §6.1) vs Camadas C1/C1B/C2/C3 (SPEC-008) (score: 0.55)

Sistemas diferentes: Aletheia classifica resultado final; OpenCode classifica processo de validacao

Component-by-Component Mapping

Aletheia Components → OpenCode Equivalents

Aletheia Agent Architecture

  • Paper: §2, Figure 1
  • Results: 93% IMO-Proof Bench Advanced, 82% FutureMath Basic (condicional)
  • Key Features: Generator: solucao em linguagem natural, Verifier: mecanismo informal de verificacao, Reviser: correcao iterativa...

Gemini Deep Think

  • Paper: §2.1, Figure 2
  • Results: IMO-Proof Bench 30 problemas, FutureMath Basic (interno)
  • Key Features: Escalabilidade: 100x reducao compute (Jan 2026 vs Jul 2025), Paralelismo: exploracao simultanea de ideias, Ph.D.-level transfer: scaling law transfere para exercicios...

Tool Integration

  • Paper: §2.3, Figure 3-4
  • Results: Reducao de citacoes ficticias; erros sutis persistem
  • Key Features: Google Search: reducao de alucinacoes em citacoes, Web browsing: navegacao de literatura matematica, Python: ganhos marginais (modelo ja proficiente)...

Research Milestones

  • Paper: §3, Table 1
  • Results: 212/700 Erdos candidatos; 4 confirmados como novos
  • Key Features: Feng26: 100% autonomo (Level A2) — Eigenweights, LeeSeo26: Human-AI (Level C2) — Independence Polynomials, BKKKZ26: generalizacao Erdos-1051 (Level C2)...

Autonomy & Significance Taxonomy

  • Paper: §6.1, Tables 8-9
  • Results: A2 = Feng26, C2 = LeeSeo26/BKKKZ26, H2 = FYZ26/ACGKMP26
  • Key Features: Axis 1: H (Human-primary), C (Collaboration), A (Autonomous), Axis 2: 0 (Negligible), 1 (Minor), 2 (Publishable), 3 (Major), 4 (Landmark), HAI Cards: documentacao transparente human-AI interaction...

Evaluation Benchmarks

  • Paper: §2, §4, §3.3
  • Results: FirstProof: Aletheia 7/10; GPT 5.2 Pro 2/10 baseline
  • Key Features: IMO-AnswerBench: 400 short-answer problems, IMO-ProofBench: 60 proof-based problems, IMO-GradingBench: 1000 human gradings...

OpenCode Components → Aletheia Equivalents

Aletheia Math Research Engine (SPEC-012)

  • Spec: SPEC-012
  • Results: 5/5 solved (100%), avg 1.0 attempts, max L1_MINOR
  • Key Features: Generator: 16 tipos de raciocinio com selecao adaptativa por dominio, Verifier: Cora-Debate V1-V7 (7/7 checks) + deteccao alucinacao (6 padroes), Reviser: feedback loop com budget de 10 tentativas...

Cora-Debate V1-V7

  • Spec: N/A
  • Results: 7/7 checks integrados ao Aletheia Verifier
  • Key Features: V1: Logical Consistency, V2: Mathematical Correctness, V3: Edge Case Coverage...

Reasoning Orchestrator v11

  • Spec: N/A
  • Results: 212+ tipos mapeados e documentados
  • Key Features: 68 tipos base + 10 Teoria dos Jogos + expansoes, 12 categorias (logica, dialetica, estrategia, inovacao, etc.), Pipeline de 7 fases com agentes especializados...

Triangulacao Anti-Circularidade (SPEC-008 + 008-B)

  • Spec: N/A
  • Results: 14/14 TDD, domain-shift P95=0.215, P99=0.279
  • Key Features: Camada 1: Split temporal cego (Bergmeir 2012, Cerqueira 2020), Camada 1B: Domain-shift detection (bootstrap Jaccard), Camada 2: Perturbacao adversaria (4 transformacoes)...

CORA-Eval Benchmark

  • Spec: N/A
  • Results: D1:14/14, D2:8/8, D9:12/12; baseline CORA-Score 0.67
  • Key Features: D1: Raciocinio Matematico Formal (14 CTs, SPEC-009), D2: Modelagem de Sistemas Fisicos (8 CTs, SPEC-010), D9: Desenho Experimental e Metodologia (12 CTs, SPEC-011)...

MCP Tool Ecosystem

  • Spec: N/A
  • Results: 18 ativos, 24 inativos (expansiveis)
  • Key Features: Web Search (DuckDuckGo): busca web, Sequential Thinking: raciocinio multi-passo, Python Interpreter: execucao de codigo...

Multi-Domain Coverage

  • Spec: N/A
  • Results: 6 dominios cobertos com TDD proprio cada
  • Key Features: Juridico: 6 skills (pecas, contratos, jurisprudencia, etc.), Arteterapia: validacao clinica decolonial (SPEC-013), Economia: analise ARM-IAG (World Bank, complexidade)...

Full Reproducibility Infrastructure

  • Spec: N/A
  • Results: 71/71 TDD, 2.091 arquivos espelhados, 0 erros
  • Key Features: 71/71 testes automatizados em 6 suites, Seed fixa (42) em todos os scripts, Hash MD5 verificavel de cada artefato...

Critical Gaps & Roadmap

Gaps (OpenCode needs to improve)

  1. Foundation Model Scale

    • Deep Think: IMO-Gold, inference-time scaling, 100x compute reduction
    • OpenCode: depends on accessible API models (GPT, Claude, Gemini via API)
    • Mitigation: Cora V1-V7 compensates with verification rigor

  2. Proprietary Benchmarks

    • FutureMath Basic: Ph.D. exercises (internal only)
    • FirstProof: time-limited competition (expired)
    • Mitigation: CORA-Eval D1-D9 + Olympiad benchmarks

  3. Human Expert Validation Pipeline

    • Aletheia: team of ~15 mathematicians for validation
    • OpenCode: Camada 3 (anotacao humana minima, 30 docs)
    • Mitigation: SPEC-008 Camada 3 + active learning

Advantages (OpenCode exceeds Aletheia)

  1. Verification Rigor: Cora V1-V7 (7 checks) > informal verifier
  2. Anti-Circularity: SPEC-008 framework solves the "single use" problem
  3. Domain-Shift Detection: SPEC-008-B (unique capability)
  4. Multi-Domain: 6+ domains vs math only
  5. Reproducibility: 71 TDD tests + seed + hash vs paper-only
  6. Tool Ecosystem: 18 MCPs vs 3 tools
  7. Reasoning Taxonomy: 212 explicit types vs implicit

Conclusion

The OpenCode ecosystem implements the core Aletheia architecture (SPEC-012) while adding verification rigor (Cora V1-V7), anti-circularity (SPEC-008), domain-shift detection (SPEC-008-B), multi-domain coverage, and full TDD reproducibility.

The critical gap remains the foundation model scale — Gemini Deep Think's inference-time scaling law and IMO-Gold achievement are not replicable with public API models. However, OpenCode's verification layers partially compensate by catching errors that a single-pass informal verifier would miss.

In the taxonomy of Feng et al. (§6.1), OpenCode achieves Level C2 (Human-AI Collaboration, Publishable Research) across multiple domains, with the Aletheia Math Research Engine (SPEC-012) operating at Level A1-A2 (Autonomous, Minor to Publishable) within mathematical domains.

Generated by cross_correlation.py — OpenCode Ecosystem v4.3.0