GPD Physics Verification

MTF integrates with Get Physics Done (GPD) to shift hypothesis selection from chi-squared toward physical correctness. Rather than reimplementing physics verification, MTF uses GPD’s existing MCP servers as callable tools — the same way it uses arxiv and Semantic Scholar.

Install with pip install -e ".[gpd]". Controlled by config.enable_gpd_mcp (default True; no-ops gracefully if the package is missing).

GPD servers

Server	What MTF gets	Which agents / phases call it
verification	Structured checks: dimensional (5.1), symmetry (5.2), limiting cases (5.3), fit-family (5.18)	_run_phase_physics_checks (fitting phase), _screen_hypothesis_plausibility (literature phase), DebateEngine postscript, ReviewerAgent
errors	104 curated error classes with detection strategies (sign errors, missing 2π factors, gauge artifacts, etc.)	LiteratureAgent, _prefetch_fitting_warnings (fitting phase), ReviewerAgent
protocols	Step-by-step methodology with checkpoints for 47+ physics domains	LiteratureAgent, FittingAgent, _classify_domains (orchestrator)
conventions	Canonical defaults for 18 subfields (Fourier convention, metric signature, natural units, gauge choice)	Convention locking (literature phase), FittingAgent pre-exec check
patterns	Persistent cross-session error pattern library in ~/.gpd/	_prefetch_domain_patterns (literature phase), _prefetch_fitting_warnings (fitting phase), LiteratureAgent, FittingAgent, ReviewerAgent
skills	Programmatic discovery of available GPD capabilities and physics domain routing	_classify_domains (orchestrator pre-flight)

Physics-first ranking

When GPD is active, DebateEngine.synthesize() adds a physics-first ranking criterion to the system prompt for fitting and review phases:

A model with χ²=1.5 that passes all verification checks ranks above χ²=0.9 with a dimensional analysis failure.

For fitting and review phases, DebateEngine also extracts LaTeX/dimensional expressions from the synthesis text and calls dimensional_check as an objective postscript — appended to the synthesis and stored as PHYSICS_VERDICT.

Auto domain classification

Before the literature phase, MTFOrchestrator._classify_domains() calls route_protocol and route_skill with the phenomenon description, parses known GPD domain names from the responses, and overwrites config.physics_domains for the run (ephemeral — no persistence). Falls back to the configured default if no domains are detected.

Controlled by:

• config.auto_detect_domains: bool = False

• config.gpd_domain_detection_max_domains: int = 3

The detected domains (or fallback notice) are written to MemoryKind.DOMAIN_CLASSIFICATION as an audit trail.

Convention locking and pre-exec validation

At the start of the literature phase, MTF calls subfield_defaults once per domain in config.physics_domains and stores the result as MemoryKind.CONVENTIONS. Every subsequent agent sees these locked conventions in its prompt context, preventing silent mismatches (Fourier sign, metric signature, natural-unit choices) between agents.

Additionally, FittingAgent.fit() calls convention_check on generated fitting code before exec() — a phase-level check separate from the agent’s own subfield_defaults call. If the check returns FAIL, the violation is written to PHYSICS_VERDICT and the agent retries once with the violation text in context.

Controlled by:

• config.fitting_convention_check: bool = True

• config.fitting_max_convention_retries: int = 1

Literature plausibility screening

After each literature debate synthesis, _screen_hypothesis_plausibility() runs limiting_case_check on each candidate hypothesis (extracted from the synthesis text) with generic limits: classical_limit, zero_coupling, large_N. Results are shown to the user as [PASS] / [WARN] / [FAIL] badges before the approval gate, and stored as PHYSICS_VERDICT.

If config.auto_reject_physics_failures = True, hypotheses receiving a CRITICAL FAIL verdict are filtered from the approved list (with a non-empty fallback in case all hypotheses fail).

Controlled by:

• config.literature_plausibility_screen: bool = True

• config.auto_reject_physics_failures: bool = False

Cross-session pattern memory

GPD’s ~/.gpd/ pattern store is the only persistent cross-session memory in the MTF pipeline. Patterns are used at three points:

1. Literature pre-fetch: _prefetch_domain_patterns() calls lookup_pattern(category="convention-pitfall") per domain before the first literature fan-out. Results are stored as DOMAIN_PATTERNS and appear in LiteratureAgent prompt context.

2. Fitting pre-fetch: _prefetch_fitting_warnings() calls lookup_pattern(category="sign-error"), lookup_pattern(category="convergence-issue"), and check_error_classes per hypothesis before the fitting fan-out. Results are stored as FITTING_WARNINGS and appear in FittingAgent prompt context.

3. Review and recording: ReviewerAgent calls lookup_pattern before reviewing and add_pattern after finding new errors. LiteratureAgent and FittingAgent also have add_pattern as a tool to record systematic errors found during their work.

Usage

# Recommended (GPD enabled by default)
mtf "anomalous resistivity plateau"

# Disable for faster iteration
mtf "anomalous resistivity plateau" --no-gpd

# Cross-domain
mtf "neutron star cooling anomaly" --physics-domains gr nuclear amo

# Specific GPD servers only
mtf "..." --gpd-servers verification errors

Python API

config = MTFConfig(
    enable_gpd_mcp=True,
    physics_domains=["condensed_matter", "qft"],
    # Domain auto-detection
    auto_detect_domains=False,
    gpd_domain_detection_max_domains=3,
    # Literature plausibility screen
    literature_plausibility_screen=True,
    auto_reject_physics_failures=False,
    # Pre-exec convention validation
    fitting_convention_check=True,
    fitting_max_convention_retries=1,
)