The Landscape

Everything in this directory has a shelf life. Everything outside it does not.

The lessons in curriculum/ and the definitions in concepts/ never name a product, a vendor, or a benchmark score. That is enforced by tools/validate_quarantine.py, not by good intentions: a build fails if a name from registry.yml appears in the conceptual spine outside a marked block.

Why the wall exists

While researching this course, the same memory system was found cited at two very different scores on the same benchmark suite — one figure from an independent comparison, one from the vendor's own blog. Neither was flagged as contested. Both were published within months of each other.

That is not unusual, and it is not necessarily dishonest. Benchmark numbers depend on the backing model, the retrieval budget, the prompt, and which subset was run. But it means a number quoted without its conditions is close to meaningless, and a number embedded in a lesson is a number that will be wrong by the time someone reads it.

So the rule is: the mechanism goes in the lesson; the vendor goes here, dated. A learner who understands supersession can evaluate any system that claims to do it. A learner who memorised a leaderboard row has learned something with a half-life of weeks.

How to read anything in this directory

• Check last_verified before trusting a page. High-volatility pages fail CI at 180 days precisely so they cannot rot silently.
• Treat claims_are_vendor_sourced: true as "this is marketing until reproduced."
• Run the numbers yourself where it matters. The eval harness built in end-to-end eval scores any system against the same corpus this course teaches from, which is a more useful comparison than a leaderboard you did not run.

Contents

Pages land here as the course reaches the modules that reference them:

• tools/ — how shipping systems structure extraction, storage, and conflict resolution
• benchmarks/ — what each public memory benchmark actually measures, and what it misses
• standards/ — proposed portability and interop work, and how far to trust it
• papers/reading-list.md — the surveys and primary sources worth reading directly