Static copy

Copied from docs/en/governance/lpc.md in https://github.com/poppy-team/prumo (MIT). Pinned to revision e91694d3959be1ed92757b34063efe0b2191821f, blob 66fe47b0fc357ec3dbfe628347799969c3fdda6f. The source repository remains canonical; this copy is refreshed through a sync pull request, not live.

Lean Progressive Context (LPC)

The Lean Progressive Context (LPC) methodology is one of Prumo's core scientific differentiators. It was conceived to solve the problem of cognitive degradation and hallucination in language models when they are subjected to saturated context windows.

The Three Core Principles

1. Smallest Sufficient Context: The agent must not receive the entire codebase or the whole documentation tree. It should receive only the slice strictly necessary to fulfill the current directive.
2. Progressive Expansion: If, during execution, it turns out that an adjacent dependency or contract needs to be understood, the agent uses specific tools to request the expansion of that focal point.
3. Bounded Output: The agent must never dump hundreds of lines of logs, memory dumps, or complete output from test suites. ACI tools filter and bound the output before handing it to the LLM.

Practical Impact

Metric	Traditional Approach (Megaprompt)	Prumo with LPC
Average Token Consumption	80k – 200k tokens per turn	4k – 12k tokens per turn
Hallucination Rate	High (accumulated noise)	Close to zero (narrow focus)
Cost per Session	High ($$$)	Up to 85% lower ($)
Response Time (TTFT)	Slow (high prompt latency)	Ultra-fast