Skip to main content

Tech Freshness Defense

Benchmarked against: Anthropic — Reduce hallucinations Version: 1.0 | Scope: All agents

Your training data is ~10 months old. Any fast-moving library API you "know" is probably wrong.

PK (Perishable Knowledge) = fast-moving tech knowledge that expires. This is the single biggest risk to code quality in an AI-driven team.


The problem

AI models confidently generate code using APIs that no longer exist. Real incidents:

  • Three Work Orders rejected in a single day for the same root cause (no Pre-Flight Check)
  • Fabricated version numbers (e.g., "LangGraph 2.0" when actual version was 1.0.9)
  • Code using deprecated APIs that fail at runtime

Captain's directive: enforce with tools, not discipline.

Knowledge classification

CategoryExamplesRule
TimelessDesign patterns, algorithms, frameworksUse directly
PerishableLibrary versions, API signatures, config formatsVerify before using

Danger Zone libraries

These libraries change APIs frequently. Never write code from memory when importing from these:

EcosystemWhy dangerousHow to verify
ai / @ai-sdk/*Vercel AI SDK v4→v5→v6 broke APIs completelyRead node_modules/ai/package.json + actual exports
next / next/*Next.js 14→15→16 changed App Router, caching, middlewareRead node_modules/next/package.json
@cloudflare/*Workers API evolves fastCheck wrangler.toml + Cloudflare docs
@supabase/*Client APIs change between major versionsRead actual installed version
langchain / langgraphv0.x→v1.x API overhaulCheck .venv installed version

Mandatory verification steps

When modifying code that imports from a Danger Zone library:

  1. STOP — Do not write code from memory
  2. CHECK VERSION — Read package.json or pip show <pkg> for installed version
  3. CHECK ACTUAL API — Read the actual source in node_modules/ or .venv/lib/
  4. THEN write code — Based on what you verified, not what you "know"

Skipping freshness verification on Danger Zone libraries = work accident.

Pre-Flight Check integration

The Pre-Flight Check 3D scoring system includes Knowledge Freshness as one of three dimensions:

Blast Radius × Knowledge Freshness × Reversibility = Risk Score (0-6)

Enforcement

Tech freshness is enforced through:

  • Pre-commit hooks that check for Pre-Flight Check records
  • WO submission gates requiring freshness verification
  • Danger Zone auto-detection in import statements