Add comprehensive governance architecture paper

[hadamrd]

Summary

This PR adds a complete position paper documenting the governance architecture of forge-loop, a system for autonomous multi-agent software development. The paper articulates the design philosophy, control surfaces, and tradeoffs that enable safe, bounded autonomous code generation.

Key Changes

• 01-introduction.md: Establishes the core problem — that autonomous coding systems drift without governance, exhibiting three pathologies: proxy substitution, value-blindness, and quality entropy. Introduces the thesis that governance must be machine-checkable.

• 02-the-tradeoff.md: Frames the central engineering tradeoff: upfront specification cost (articulating product value and quality rules) in exchange for bounded, non-decreasing value over unbounded autonomous actions. Includes economic analysis showing why this cost structure is favorable compared to per-action human review.

• 03-product-articulation-axes.md: Defends the first control surface — structured value axes with explicit negative space (rejected_as_cosmetic). Argues this makes value falsifiable and defeats proxy substitution by removing the easiest-to-satisfy moves from the admissible work set.

• 04-reinforcement-feedback-loops.md: Defends the second control surface — a typed critic gate and a novel bug→rule→permanent-gate ratchet. The ratchet converts each shipped defect into a constraint that blocks that failure class forever, making the failure set monotonically shrink.

• 05-code-quality-imperatives.md: Defends the third control surface — quality manifestos enforced at both generation time (guidance in worker briefs) and admission time (critic enforcement). Argues quality is a precondition for autonomy because agents imitate the code they read.

• 06-system-architecture.md: Shows how the three surfaces compose as ordered gates on a single repeating control loop (the "tick"). Explains why the ordering (value → generation → quality → liveness) is optimal for failing work as early and cheaply as possible.

• 07-limitations.md: Provides honest assessment of threats to the thesis (value may not be reducible, gates can be collectively gamed, ratchet may not converge) and gaps in the artifact (god-function orchestration core, speculative scaffolding, duplicated control structures). States what evidence would falsify the claims.

• 08-conclusion.md: Restates the argument, emphasizes the durable contribution (the governance philosophy, not the specific product), and closes with the core principle: as generation becomes free, value is decided at the gates.

• README.md: Index and abstract for the paper, with table of contents and reading guide.

Notable Implementation Details

• The paper grounds all arguments in actual forge-loop mechanisms (.forge/axes.yaml, .forge/quality-manifesto.md, critic module, brainstormer, merge gate) rather than idealized systems.
• Each section is self-contained but cumulative, building from problem statement through three control surfaces to their composition and honest limitations.
• The economic argument (Section 2) is central: specification cost amortizes across unbounded actions while per-PR review cost never does, making governance strictly cheaper past an early crossover.
• The ratchet (Section 4) is presented as the most original contribution: a structural mechanism that makes the system's own failure set shrink with experience, converting per-instance debugging cost into one-time rule-writing cost.

https://claude.ai/code/session_01MUGNuvHtqhSGZ5RPK1zDYy