Political World — when AI documentation became stale

✅ Reproduced 🤖 AI failure prevention

Political World already had unusually AI-friendly repository files:

AI_START_HERE.md
llms.txt
API references
RU/EN documentation
examples/
templates/

That was good.

It still developed version drift.

What drifted

In one inspected GitHub tree:

PoliticalWorldAPI.cs → API 1.9.0
README.md            → API 1.9.0
AI_START_HERE.md     → API 1.6
llms.txt             → API 1.6.0

A separate newer runtime log from development later reported API 1.14.0.

Why this is dangerous for coding assistants

An AI can follow instructions perfectly and still produce the wrong result if its entry file is stale.

For example, an assistant reading only AI_START_HERE.md could correctly obey:

Check IsCompatible(1, 6)
Read API_REFERENCE_1_6.md

while ignoring capabilities already present in the inspected 1.9 source.

Conversely, an assistant given a newer runtime log might assume 1.14 APIs exist in the older public GitHub source.

Both errors come from missing provenance, not from syntax.

Fix at the documentation-system level

AI entry files should carry machine-readable version metadata.

Example:

project_version: 1.7.0
api_version: 1.9.0
source_commit: ce0c91754722dd1e88e8eaf116c4c667ae020204
verified_worldbox: 0.51.2-build-719
generated_from: source
last_verified: 2026-08-16

And CI should eventually compare the declared AI version against the actual code constant.

Future automated check

For Political World, a simple verification script could extract:

public const string ApiVersion = "..."

and compare it against:

README
AI_START_HERE.md
llms.txt
API reference index

If they disagree, CI should fail or at least emit a documentation warning.

Lesson

Creating llms.txt is not the finish line.

AI documentation needs the same maintenance discipline as public API documentation:

• version binding;
• automated consistency checks;
• explicit snapshot identity;
• stale markers;
• runtime evidence kept separate from source evidence.

The best AI-friendly documentation is not the one with the most text.

It is the one that makes it difficult for a model to combine incompatible facts.