ovr.md

Reformatted Content:

GDAL (Geospatial Data Abstraction Library) AI Tool Policy vs. TorchGeo Projected Stance

The GDAL AI Tool Policy serves as a template and philosophical anchor for Adam Stewart's proposed policy in #torchgeo. This comparative analysis highlights areas where adaptation is necessary due to differences in project scale, community size, and technical domain.

Core Philosophical Alignment: Maintenance as the Scarce Resource

Both projects share the fundamental premise that LLMs have decoupled code creation from code value. GDAL explicitly states that maintenance time is the scarce resource, while TorchGeo echoes this by citing the increase in "AI slop" and verbose PRs as a threat to maintainer sanity.

Implication for TorchGeo: The policy must frame itself as pro-maintenance, protecting the project's longevity by filtering out contributions that degrade signal-to-noise ratio without adding proportional value.

Structural Adaptation: From "Limited Use" to "Strict Permissive with Disclosure"

Feature	GDAL Policy (Template)	TorchGeo Projected Stance (Adaptation)	Rationale for Difference
Primary Authorship	Human must be primary author; LLM is tool only.	Human must hold copyright and understand every line.	Aligns with #torchgeo's existing strict copyright/header requirements (MIT License).
Scope of Use	Limited to auto-completion/mechanical tasks. Vibe-coding banned.	Allowed if disclosed, concise, and high-signal. "Vibe coding" discouraged but not explicitly defined as a ban in the same way; focus is on outcome (quality/conciseness).	#torchgeo contributors are more likely to use AI for complex logic generation than GDAL's conservative community allows. A total ban would alienate potential contributors.
Disclosure	Mandatory labeling of LLM usage in tickets/commits.	Mandatory disclosure via checklist or explicit comment/collapsible section.	Isaac Corley pushed for this as a compromise to allow utility (e.g., tables/metrics) while maintaining transparency.
Agent/Bot Use	Fully banned (human-coordinated or uncoordinated).	Banned if fully autonomous; human must be responsible.	Similar stance, but #torchgeo emphasizes the legal aspect of copyright ownership for agents more heavily than GDAL's operational focus.
Verbosity	Actively discouraged; high verbosity may lead to closure.	Must be concise and high-signal. Verbose PRs rejected on quality grounds.	Direct alignment. Both projects view verbose output as a hallmark of low-quality AI generation.

Key Differences in Enforcement and Tone

A. Legal vs. Operational Focus

• GDAL: The policy is heavily operational, focusing on the impact of LLMs on maintenance workflows.
• TorchGeo (Adam Stewart): While also concerned with maintenance, Adam places significant weight on legal liability and copyright verification (Thaler v. Perlmutter).

B. Community Size and Enforcement Feasibility

• GDAL: A large, established project with many maintainers who can enforce strict rules without overwhelming the core team.
• TorchGeo: Smaller core maintainer group (Adam Stewart + Isaac Corley). Adam's comment about being unlikely to review AI-assisted PRs >10 lines unless pre-approved by another maintainer indicates a need for delegated enforcement.

C. Organizational Tiering

• GDAL: Single-project focus (though part of OSGeo).
• TorchGeo Org: Adam Stewart explicitly mentions that the policy is org-wide but enforcement varies by project (torchgeo core vs. torchgeo-bench). This requires a modular policy structure where the "Core" principles are fixed, but specific thresholds (e.g., line counts, review requirements) can be adjusted per sub-project.

Recommended Refinements for TorchGeo's Policy Document

1. Explicit "Maintenance Burden" Rationale: Include a section similar to GDAL's rationale that explicitly states AI-generated content increases maintenance load without proportional value, justifying strict review thresholds.
2. Clear Disclosure Mechanism: Adopt GDAL's labeling idea but adapt it to #torchgeo's GitHub workflow (e.g., PR template checkbox + collapsible section for verbose outputs). This satisfies Isaac Corley's need for utility while maintaining Adam Stewart's demand for transparency.
3. Copyright Responsibility Clause: Retain the strong legal language from GDAL's influence but expand it to match #torchgeo's existing MIT License header requirements, ensuring contributors warrant they hold copyright and have verified no infringement.
4. Tiered Enforcement Guidance: Clarify that while the policy is org-wide, core projects (torchgeo) will enforce stricter thresholds (e.g., <10 lines for AI-assisted code without pre-review) than research/benchmarking projects (torchgeo-bench). This prevents a one-size-fits-all approach from stifling innovation in experimental areas.
5. Conciseness as Quality Gate: Explicitly link verbosity to quality standards, mirroring GDAL's stance that high verbosity is an indicator of LLM-generated content and may be rejected on quality grounds alone.

Conclusion

The GDAL policy provides a robust structural and philosophical foundation for #torchgeo. By adapting its operational focus (maintenance conservation) and legal defensiveness (copyright responsibility) to fit TorchGeo's smaller, more agile community structure, the project can establish a "Strict Permissive" stance that protects maintainers while allowing responsible AI-assisted contributions. The key is balancing Adam Stewart's need for strict control with Isaac Corley's desire for practical utility through disclosure and conciseness rather than outright prohibition.