truth-harness

The description states that Truth Harness is a local-first verification and evidence layer for Codex and other AI agents. It routes scoped mathematical claims through deterministic engines and returns conservative trust labels such as:

• exact-computed
• cross-checked
• smt-checked
• refuted
• unverified
• proved

Each result includes input, assumptions, backend, output, evidence trace, limitations, replay command, and local artifact references.

It provides interfaces including:

• CLI
• MCP server
• Local API
• Docker verification profiles
• Browser workbench

The system is built as a TypeScript monorepo with a shared verification core used by all interfaces. Verification adapters connect to exact arithmetic routines like SymPy, Maxima, Z3, cvc5, Lean 4, and SageMath.

It stores generated evidence locally under a .truth-harness workspace, using Docker services with disabled networking, dropped capabilities, and narrowly scoped writable storage.

Inference The product is described as a tool for inspecting AI-generated math claims by routing them through deterministic verifiers and returning evidence-backed results. It is not an AI model itself but a verification infrastructure layer.

The author states that Truth Harness began with the question:

“What if an AI answer were treated as a claim to verify, rather than a conclusion to trust?”

This reflects a shift from traditional AI use where outputs are accepted at face value to one where AI-generated claims must be verified.

Key positioning elements:

• It is designed to work underneath Codex, allowing agents to propose claims, call the harness, inspect receipts, correct refuted steps, and continue from evidence-backed checkpoints.
• The system emphasizes conservative trust labels, refusing to make stronger claims than its evidence supports.
• It distinguishes between:
  • Calculation vs. proof
  • Benchmark vs. theorem
  • Simulation vs. reality
  • Source citation vs. independent validation

Inference The positioning is that Truth Harness is a verification infrastructure layer for AI agents, not a replacement or enhancement of the agent itself — it’s about ensuring trustworthiness in AI-generated math.

The description does not explicitly name target customers or personas. However, based on its architecture and use case:

• Primary users appear to be:
  • Developers working with AI agents (especially Codex)
  • Researchers or engineers needing rigorous verification of mathematical claims
  • Students or educators who want to inspect how AI-generated math is derived

The system supports both humans and agents, suggesting a dual-purpose audience.

It also targets users who are building systems where correctness matters — such as game engines, scientific simulations, or engineering tools.

Inference The ICP likely includes developers, researchers, engineers, and educators working in technical domains requiring high-fidelity mathematical reasoning. It is not yet clear if there are specific commercial customers beyond the author’s own use cases.

There is no evidence of a business model or pricing structure in the description.

The project is presented as an open-source tool built during a hackathon, with no mention of monetization, licensing, or paid features.

Inference No commercial business model or pricing has been evidenced. The system appears to be a prototype or proof-of-concept, not yet a product intended for sale.

The system is described as:

• Built using TypeScript monorepo
• Uses Docker for isolation and verification
• Integrates with multiple deterministic engines: SymPy, Maxima, Z3, cvc5, Lean 4, SageMath
• Provides CLI, MCP server, API, and browser interfaces
• Stores evidence locally in a .truth-harness workspace
• Supports replayable receipts, claim lineage, validation plans, and agent handoffs

It includes:

• 753 passing automated tests across 86 test files
• A release audit with 21 passing checks and no warnings or failures
• Real Docker-isolated verification evidence
• External engine case validation for Rust geometry predicates
• Printable reports and teaching packets generated from recorded evidence

Inference The technical foundation is solid, built around deterministic tools and isolation. It supports multiple delivery methods and includes testing and audit capabilities.

The description states:

• A working demonstration that catches a realistic continuous collision error using exact arithmetic
• 753 passing automated tests across 86 test files
• A release audit with 21 passing checks and no warnings or failures
• Real Docker-isolated verification evidence
• External engine case validation for Rust geometry predicates

However, there is no evidence of revenue, customers, user adoption, or product-market fit beyond the author’s own development.

The system was submitted to a hackathon, indicating it is likely in early-stage development.

Inference The project shows technical maturity and has been tested in a real-world scenario, but there is no evidence of traction or commercial adoption.

There is no mention of competitors or direct market context in the description. The author does not reference existing tools for AI verification or mathematical reasoning.

The system appears to be positioned as a verification layer that sits beneath AI agents like Codex, rather than competing with other AI systems or math engines directly.

It seems to address a niche need: verifying AI-generated math claims in technical workflows, which may overlap with areas such as:

• Formal verification tools
• AI agent frameworks
• Scientific computing environments

But no direct comparison or competitive positioning is made.

Inference No clear competitive landscape is described. The system appears to be unique in its approach of combining AI agents with deterministic verification layers, but it lacks evidence of market presence or competition.

1. No commercial traction or adoption: The project is presented as a hackathon submission and has no evidence of real-world usage.
2. Single-person team: Only one member (Ocean Bennett) is listed, which may limit scalability or long-term development capacity.
3. Unproven trust model: While the system claims to be conservative in labeling, there is no external validation or feedback from users or mathematicians.
4. Limited interfaces: The system is described as supporting CLI, MCP, API, and browser workbench — but no evidence of integration into larger platforms or workflows.
5. No pricing or monetization strategy: No indication of how the product would be sold or monetized if it were to evolve.

Inference The main risk is that this remains a prototype or personal project with no clear path to commercial viability or market traction.

1. Has Truth Harness been used beyond the author’s own development and demo?
2. What kind of feedback have you received from users or collaborators in technical domains (e.g., engineering, research)?
3. How does Truth Harness handle edge cases where verification fails or is inconclusive?
4. Are there any plans to support more types of mathematical claims or integrate with additional tools beyond those listed?
5. What are the limitations of the current Docker-based isolation approach, and how might they scale?
6. Have you considered integrating with existing AI agent frameworks or platforms (e.g., LangChain, AutoGen)?
7. How do you plan to evolve the trust labeling system over time?