Verified Technical Solver

Give AI a calculator, a proof ledger, and a CAD kernel—not another chance to guess.

Verified Technical Solver is a Codex plugin and local MCP server that moves consequential technical claims out of free-form model reasoning and into bounded, deterministic tools. It evaluates symbolic mathematics, verifies equation solutions by substitution, checks numeric relations, audits the structure of reasoning ledgers, and builds validated parametric CAD primitives with evidence artifacts.

Why this exists

Language models are excellent at translating intent, exploring approaches, and explaining results. They are less trustworthy when exact arithmetic, unit conversions, dependency bookkeeping, or geometric validity are left entirely to token prediction. A fluent answer can still contain a sign error, a dropped constraint, an unsupported conclusion, or an invalid solid.

This plugin gives an AI a better division of labor:

• The model handles meaning: requirements, assumptions, tradeoffs, and explanation.
• Deterministic software handles exactness: symbolic calculation, substitution, comparisons, dependency-graph checks, and CAD-kernel validation.
• The user gets evidence: exact and decimal forms, residuals, pass/fail checks, measured geometry, hashes, and explicit uncertainty.

That makes the AI better not because the plugin makes it infallible, but because it makes important claims inspectable, repeatable, and easier to falsify.

What it does

Tool	Purpose	Evidence returned
calculate	Safely simplify and evaluate bounded mathematical expressions	Exact value, decimal value, LaTeX, free symbols
solve_equations	Solve bounded symbolic equation systems	Solutions, substitution residuals, verified flag
verify_relations	Check equality, approximation, and inequalities	Per-check difference and pass/fail result
audit_reasoning_structure	Find missing support, broken references, duplicate IDs, and dependency cycles	Structured issue report and conclusion coverage
build_box_part	Create plates/boxes with fillets and through-holes	Kernel validity, solid count, dimensions, volume, area, STEP/STL/SVG, SHA-256
build_cylinder_part	Create cylinders and tubes	Kernel validity, measured geometry, exports, verification report

What makes it safer

• It never uses Python eval; expressions are parsed through a small AST allowlist.
• Inputs have bounds for expression size, syntax nodes, symbols, equations, solutions, dimensions, holes, and worker output.
• Each operation runs in a subprocess with a timeout, isolating failures from the MCP server.
• Equation solutions are checked by substitution.
• CAD output must be one kernel-valid solid and is measured after construction.
• Output paths are restricted to VTS_OUTPUT_ROOT and existing artifacts are never overwritten.
• Reasoning audits are described honestly: they validate structure, not whether natural-language premises are true.

Installation

Requirements

• Python 3.11+
• Codex with plugin/MCP support
• Dependencies in requirements.txt (mcp, sympy, and cadquery)

Local setup

git clone https://github.com/ASVLCII/verified-technical-solver.git
cd verified-technical-solver
python -m venv .venv

Activate the virtual environment, then install dependencies:

python -m pip install -r requirements.txt

The included .mcp.json starts scripts/server.py with python. If Codex does not launch inside your activated environment, replace command with the absolute path to the virtual environment's Python executable.

Optionally set VTS_OUTPUT_ROOT to control where CAD artifacts may be written. Without it, output is restricted to ~/Documents/Codex.

Example prompts

Use Verified Technical Solver to solve x + y = 5 and x - y = 1. Verify every solution by substitution.

Check whether this tolerance stack remains below 0.25 mm. State assumptions and show each deterministic relation check.

Create a 60 mm × 30 mm × 4 mm plate with four 4 mm through-holes and a 2 mm edge fillet. Export STEP, STL, and a dimensioned SVG, then report kernel validity and hashes.

Build a reasoning ledger for this engineering recommendation, audit its structure, and distinguish evidence from assumptions and judgment.

Verification workflow

1. Restate the objective, known values, units, constraints, and missing engineering inputs.
2. Separate semantic judgment from mechanically verifiable claims.
3. Route quantitative dependencies through deterministic tools.
4. For difficult problems, compare materially different approaches.
5. Record conclusions in a reasoning ledger with stable IDs, dependencies, evidence, and explicit assumptions.
6. Audit the ledger, then independently inspect semantic validity.
7. Use substitution, dimensional analysis, boundary cases, or another independent route for consequential results.
8. Report failures and uncertainty alongside successful checks.

See verification-protocol.md for the evidence classes and CAD acceptance criteria.

Testing

python scripts/test_solver.py
python scripts/test_mcp.py

The tests cover exact calculation, substitution, rejected code execution, bounded factorials, visible relation failures, reasoning cycles, CAD dimensions and volume, invalid/overlapping geometry, output confinement, overwrite refusal, MCP initialization, and recovery after a rejected call.

Limits

• The reasoning audit checks graph structure and evidence bookkeeping; it is not a theorem prover or truth oracle.
• CAD generation currently supports exact box/plate and cylinder/tube primitives, not arbitrary assemblies or free-form surfaces.
• A kernel-valid model is not automatically manufacturable. Tolerances, fits, material, loads, process constraints, and regulatory requirements still need qualified engineering judgment.
• Symbolic computation can be expensive; the server applies bounds and timeouts, but it is not intended for untrusted public multi-tenant execution.

Project layout

.codex-plugin/plugin.json          Plugin metadata
.mcp.json                          Local MCP server declaration
scripts/server.py                  MCP interface and worker isolation
scripts/worker.py                  Math, reasoning, and CAD implementation
scripts/test_solver.py             Deterministic unit tests
scripts/test_mcp.py                MCP integration test
skills/verified-technical-solver/  Agent workflow and verification protocol

Contributing and security

Contributions are welcome—especially new independently verifiable operations, stronger resource bounds, additional geometry acceptance tests, and cross-platform installation improvements. Read CONTRIBUTING.md before opening a pull request. Please report security issues privately as described in SECURITY.md.

License

MIT © ASVLCII. See LICENSE.