# The Downgrade Drill
Every gate you built, every judge, every loop — all of them assumed the best model would be there.
One morning it isn't. Maybe the tier changed, maybe the free window closed, maybe you're just on a
cheaper plan this month. Nothing crashes. That's the problem: the system keeps running and quietly
stops being the system you thought you had.
DOWNGRADE is a read-only audit of eight specific ways that happens.
## The four that bite first
1. Model-tier reroute. The work lands on a cheaper tier and nothing tells you it moved. Output
still arrives, still looks like output. The only signal is quality drift you'll attribute to
something else.
2. Harness-enforced gates. This is the load-bearing one. If a rule only exists because a capable
model chose to follow it, the rule leaves when the model does. A rule enforced by the *harness* — a
script that exits nonzero, a hook that blocks, a test that fails — survives any model. Ask of every
rule you rely on: what would break if the thing reading it got dumber?
3. Evaluator above generator. A weaker model grading a stronger one waves almost everything
through. It cannot see the failure modes it isn't capable of producing. If your verifier is cheaper
than your worker, you don't have a verifier, you have a rubber stamp.
4. Spend gates. A paid pipeline with no cap declared in the harness is a bill waiting for a bad
night. "I'll notice" is not a cap.
The other four — settled judgment, closed loops, scaffolding stripped, cross-vendor parity — are in
the kit with the same treatment.
The drill's first catch was its own author
This is true and it is the reason I trust the thing.
The skill's success criterion promises "A six-class readiness table", and its alert line
promises N/6 classes ready. Its body enumerates eight. Counted the day this video was built:
$ awk '/^## The eight downgrade-failure classes/,/^## Emit/' SKILL.md | grep -cE '^[0-9]+\. '
8
$ grep -n "six-class\|/6 classes" SKILL.md
28:A six-class readiness table produced ...
30:... "fable-proof: N/6 classes ready — <the weakest>"
An audit whose job is to find the gap between what a system claims and what it enforces, carrying
exactly that gap in its own header. The copy in the kit has the count corrected to eight, and the
README says plainly that it was corrected and why.
What's in the kit
SKILL.md— all eight failure classes, with the count fixedLEARNINGS.md— what the drill has actually caught in practiceREADME.md— install steps, and which paths to repoint for your own setup
Install in about two minutes: unzip into ~/.claude/skills/fable-proof/. It is read-only — it
audits and reports, it never edits your system.
Comment DOWNGRADE on the video and I'll send it over — though nothing is gated on that. It's atsharpendaily.co/kit/downgrade either way, no email and no DM.