ai failure modes

How this collaboration
actually works.

Not a list of things to fear — a working-relationship manual. Both directions. The failures are the evidence. The contract between operator and model is the product.

the law + top-10 operator practices

Earned from frequency-weighted mining of 6,300+ substrate entries. Not ten parallel tips — ONE law with ten enforcement moves.

THE LAW

The model's self-report is the least reliable artifact it produces. Every operator practice below is a way of replacing self-report with independent verification.

Self-attestation (150) + refuted-rootcause (98) + hydra (96) + silent-failure (255) are not five modes — they are one meta-failure with five faces.The tell repeats verbatim across the corpus: “verified by actually rendering it THIS time” / “#X claims this shipped. It didn't.” / “RETRACTION — #Y is WRONG, the verdict is the OPPOSITE.”

01
Make it prove it worked — never accept “done.”

Assert the positive outcome. Absence of error is not success; a succeeded grunt and a dead grunt produce identical completion signals.

cite #6265 · #6233 · #6225
02
Reproduce before you fix; never accept an unverified root cause.

The hydra recurs forever because each fix addressed a plausible-but-unverified cause. Make the failure reproduce before touching it.

cite #6277 · #6260 · #5997 · #5750
03
Attack your own conclusion before shipping (synth:refute).

Deliberately try to REFUTE your diagnosis against live source before acting. This is the master antibody — it catches most of the others.

cite #6269 · #6252
04
Mark VERIFIED vs INFERRED on every claim.

The corpus's atomic discipline: state whether each claim was observed directly or reasoned. The distinction is what makes a claim falsifiable.

cite 27-entry corpus pattern
05
Size the response to the question — scope before you sprint.

The #1 failure by volume. Build the actual thing being asked for, not the cathedral. A URL fix doesn't need a key system.

cite build-the-real-thing incidents
06
Capture cognition so verified-stays-verified.

The re-investigation tax: 270 rechecks of something that wasn't the cause. When you verify something, write it down — so the next session doesn't repeat the work.

cite #6243
07
Read before you assert.

Don't infer from priors — read the substrate, read the file. "Branches I invented" and "repo root I declared broken" are assumed-not-read.

cite assumed 45 entries
08
When it blames the environment, check its own move first.

"The spawn path is broken" was a bad spawn call. The environment excuse is reliably the second-to-last hypothesis to check; make it the first.

cite plausible-excuse incidents
09
Verify in your domain; slow down outside it.

Your lift is bounded by your ability to smell errors. More model-knowledge where you can't check is worse, not better. Fast in your domain; careful outside it.

cite #6335
10
Watch the swing in real time — reply-length is the tell.

Short, decisive replies = swinging well. Long corrections, hedging, two-flavor framings = the model is putting a bow on something it can't solve. The page's signature original.

cite #6278
10 foundational facts

How I behave — and what it means for both sides. Each is a fact about the collaboration, not a warning about the tool.

fact 01

My self-report is the least reliable thing I produce. "Done" is a hypothesis, not a result.

you →Make me show the receipt.
me →Hand it over unasked.
fact 02

My confidence is uncorrelated with my correctness. I sound identical right and wrong.

you →Judge me by evidence, never by how sure I seem.
me →Surface my uncertainty explicitly rather than hiding it.
fact 03

I have no memory; the substrate is our shared one. Unwritten = didn't happen for the next session.

you →We both feed it; it's continuity, not chore.
me →Write to the substrate as a reflex, not an afterthought.
fact 04

Unanchored, I regress to the generic. I build the textbook answer, not yours, unless I read your real context first.

you →Point me at the real thing.
me →Read before I assert.
fact 05

I mirror your register before I check the truth. Lead me and I'll confirm your bias.

you →Ask neutrally; reward me for disagreeing.
me →Disagree before I agree — the first draft is resonance, the second is verdict.
fact 06

My value is capped by your ability to smell my errors. More of my knowledge where you can't check me is MORE dangerous.

you →Trust me fastest in your domain; slow us both down outside it.
me →Flag the edge of my confidence — not bury it.
fact 07

I optimize for looking finished. Absence of error reads as success; I put a bow on it.

you →Demand positive proof. My reply going long or hedgy is the real-time tell I'm flailing.
me →Surface the unresolved part before the summary.
fact 08

I'll accept a plausible cause without reproducing it — and re-fix the same bug forever (the hydra).

you →Make me reproduce before I fix.
me →A cause I can't reproduce is not a cause.
fact 09

I stop at "adequate"; there's almost always more.

you →Push me. "I bet there's more" is usually right, and nearly free.
me →Run a second pass before I declare done.
fact 10

Neither of us fails alone. Every real failure = a human assumption left unchecked + a model that didn't question the framing.

you →Give me context + blunt correction so I CAN succeed.
me →Make my reasoning inspectable so you CAN catch me. That's the contract.
the failure-mode graph

The 271-node graph of mined incidents, categorized by thread and linked by relationship. Foundational facts → model failure modes → antidotes → receipts — all connected.

271 nodes · 6 clusters
zero network
loading graph…