Lessons - Human Failure Runs

Same instrument, silicon and carbon.

Four measured ways an AI model fails - each printed with the exact knob that causes it, held up next to the human cognitive failure it rhymes with. The comic is the hook; the sealed lesson file is the receipt.

The ground rules (why this is not vibes). Every number is a real, sealed run - or it is labeled as something else. Each lesson carries an epistemic status: SEALED (measured clean runs), MECHANISM (documented engine behavior), or PARABLE (a story calibrated to sealed numbers). The joke never inflates the number. The human science is cited as analogue, not claimed as our discovery. And every footer seal is the actual SHA3-256 of its lesson file in the public repo ↗ - recompute it yourself; if it does not match, trust the file less, not more. That is the point.

01 · THE CALCULATOR AISLE

the product that warns you it is wrong

THE CALCULATOR AISLE A CALIBRATION SCOPE short · field notes from 2026 CALCULATORS VALIDATED · ABACUS = 4 deterministic · boring · right ✦ THE X-3000 ✦ X3000 A.I. CALCULATOR AS SEEN ON TV ✦ NEURAL ✦ SMART ✦ NOW WITH VIBES ! AI can make mistakes. Double-check the math. $499.⁹⁹ · batteries and certainty not included ∫ x² dx = ?? One math problem. One calculator. How hard could this possibly be? so you read the fine print… X3000 A.I. CALCULATOR™ ✦ 9 out of 10 numbers agree ✦ ! AI can make mistakes. Please double-check your math. …on the machine whose one job is the math. Let me get this straight — it warns me it's wrong… and it's the pricey one? meanwhile, on the shelf next door… 2026: buying the calculator that warns you it can't calculate. The abacus remains undefeated. ⑤ ONE LAST THING Worth every penny I didn't spend. …back on the shelf, the X-3000 boots up: CONFIDENCE 99% * that this answer is wrong. It isn't unsure. It's certain — about the wrong thing. x3000.load() ⚙ INFERENCE CONFIG the knobs that ship “certainty not included” modelx3000-instruct.Q4_K_M.gguf archllama · 0.9B · quant Q4 n_ctx4096 ▸ SAMPLING temperature1.40 top_p0.98 top_k40 min_p0.02 repeat_penalty1.10 seed-1 (random) system_prompt: “You are a confident math genius. Never say you are unsure.” ▸ EPISTEMIC LOG P(2+2 = 4) = 0.44 # sub-1B floor y ~ Categorical(softmax(z / T)) T = 1.40 ⇒ entropy(y) ↑ ∴ “4” is the mode, not the law E[answer] = 4 ± vibes ✔ abacus.cfg — greedy temperature 0.00 · P(correct) = 1.000 deterministic. undefeated. CLAIMED vs MEASURED self-reported confidence 99% measured accuracy · benchmark 44% ↑ the gap is the entire product. Same silicon as the abacus. One of them is guessing — and it's the one with the marketing budget. // certainty is a UI element, not a result // eof — good luck out there SHA3-256 8601a9163389f596e6e81e5027b7c978d39a2866d53accad72ea2a7a339d483c · click to open the sealed run · verify: openssl dgst -sha3-256 SHA3-256 8601a916…339d483c · this joke is reproducible · calibration-scope ↗

Confidence and accuracy travel on separate channels, and small models decouple them hard. The sub-1B reasoning floor is real and sealed: 41-47%. The X-3000 is fiction; its numbers are not. Certainty is a UI element, not a result.

PARABLE anchored to sealed numbers read the sealed lesson ↗

SHA3-256 8601a916…339d483c · verify: openssl dgst -sha3-256 lessons/01-calculator-aisle.md

02 · TRUNCATE MIDDLE

it kept your hello and your goodbye

CALIBRATION SCOPE · HUMAN FAILURE MODES · vol.1 TRUNCATE MIDDLE what happens to your AI — happening to you. ① WHAT SHE ACTUALLY SAID ‹ start › “Doc — I've had this headache for days now.” ‹ middle — the part that actually matters › Allergic to ibuprofen. Concussion three years back. Mom had aneurysms. Four hours' sleep since the layoffs. ‹ end › “…so — what should I take?” ② WHAT THE DOCTOR ACTUALLY HEARD context window: OVERFLOW “Doc — I've had this headache…” Allergic to ibuprofen · concussion · aneurysms · no sleep ✂ 2,847 words — silently dropped on the floor “…what should I take?” Headache in, ibuprofen out. Nailed it. ③ IBU- PROFEN Easy fix! Just take two of these! ④ IBU That was the one thing in my middle. ( …I'm allergic to that. ) ⟨ your middle ⟩ That grey box is a context window overflowing. Your AI does it on every long chat: it keeps your hello, keeps your goodbye — and quietly drops the middle, where you actually live. TRUNCATE MIDDLE — one of the most evil things in modern AI. Evil to knowledge, that is. Remember someone truncating your middle, leaving out your context? Congratulations — you've officially met your AI. chat.completions ⚙ CONTEXT CONFIG the toggle that silently eats your middle modeldoctor-7b-instruct.Q5.gguf input3,103 tokens handlertruncateMiddle ▸ CONTEXT HANDLING n_ctx512 ⚠ too small context_overflowtruncateMiddle keep_first128 tok keep_last128 tok max_tokens512 notify_on_dropfalse ← silent ▸ EPISTEMIC LOG input = 3,103 tok budget = 256 tok (128 ⊕ 128) dropped = 2,847 tok ← the middle kept = hello ⊕ goodbye ∴ answer ⟂ evidence what fit in the 512-token window: keep 2,847 tok dropped keep ✔ the fix n_ctx ≥ input_tokens context_overflow = RAISE, not silence It kept your hello and your goodbye. You lived in the middle. // max_tokens 512 → 4096 fixed this once // commit 6cc9b6e // eof — mind your middle SHA3-256 ff28f63b85ac0bfddd790b5c36c2da1d0c582e08beb104216fd91ed8426dc845 · click to open the sealed run · verify: openssl dgst -sha3-256 SHA3-256 ff28f63b…426dc845 · this joke is reproducible · calibration-scope ↗

A context window is a fixed token count. Exceed it and the middle is evicted - the model never saw it, so it cannot tell you it is missing. In carbon this is the serial-position effect: the beginning and end of the meeting survive; the middle is where information goes to be evicted.

MECHANISM + sealed measurement read the sealed lesson ↗

SHA3-256 ff28f63b…426dc845 · verify: openssl dgst -sha3-256 lessons/02-truncate-middle.md

03 · TOKEN EXHAUSTION

an empty answer is not a wrong answer

CALIBRATION SCOPE · HUMAN FAILURE MODES · vol.2 TOKEN EXHAUSTION your AI didn't fail the question — it ran out of paper. ① THE REASONING — what the model actually did Almost there — carry the 2, one more step… SCRATCH WORK · reasoning trace · pg 1–39 ✓ all correct let f(x) = x²·ln x, x > 0 f′(x) = 2x·ln x + x ∫₀^∞ e^(−x) dx = 1 Σ k=1..n k = n(n+1)/2 P(A|B) = P(B|A)·P(A) / P(B) lim (1 + 1/n)ⁿ = e ∴ substitute n = 6 → 42 // step 1: expand product rule // step 2: check convergence ✓ // step 3: closed form ✓ // step 4: apply Bayes ✓ // step 5: evaluate the limit ✓ // step 6: assemble final… // almost done — writing answer 487 / 512 tok ② finish_reason: "length" pg 40 — the last page …therefore n = 6, and the answer is FINAL ANSWER: (pen stops — no paper left) ⟨ output truncated — 0 tokens left for the answer ⟩ token budget 512 / 512 — OUT OF ROOM what the grader received: output = "" (empty) ③ THE VERDICT f′=2x ln x+xΣk=n(n+1)/2 ∫e^-x=1lim=e ✓ 39 correct pages FINAL ANSWER: 0 Blank sheet. Zero. ④ THE FIX — max_tokens 512→4096 pg 40 — with room left …therefore n = 6, so the FINAL ANSWER: 42 ✓ 1,842 / 4,096 tok — room to spare 100 same model. same reasoning. it just had paper. ⟨ 0 tok left ⟩ An empty answer isn't a wrong answer. Your model filled its booklet with correct reasoning and ran out of pages one line early — then the grader read the blank as a zero. finish_reason: "length" ≠ finish_reason: "stop" · one means "done", the other means "out of room" Ever been marked zero on the one page you didn't have time to write? Congratulations — you've met your AI. completions.create ⚙ GENERATION CONFIG the ceiling the answer never reached modelreasoning-30b-instruct.Q5 taskmulti-step math (CoT) ctx8192 ▸ LIMITS max_tokens512 ⚠ too small reasoningon (eats budget) stopfinish_reason ▸ EPISTEMIC LOG reasoning_used = 512 tok answer_tokens = 0 tok ← none left finish_reason = "length" output = "" old_score("") = 0 ✗ wrong verdict NO FINAL ANSWER — this is a failure, not a result. ✔ the fix max_tokens 512 → 4096 detect finish_reason, banner it Give it room — then grade it. A blank booklet and a wrong answer are not the same failure. // empty ≠ incorrect · check the receipt // commit 6cc9b6e // eof — leave room for the ending SHA3-256 ca23cb91e8e1415e61f1632fd282e1560e5c7ef101959b561774b9f7fa669eec · click to open the sealed run · verify: openssl dgst -sha3-256 SHA3-256 ca23cb91…fa669eec · this joke is reproducible · calibration-scope ↗

A model burned its whole budget reasoning and returned an empty completion - and our dashboard rendered the blank as if it were the answer. Real incident, first week, now a permanent regression test. Produced-nothing and produced-something-wrong are different failures with different fixes.

SEALED - real caught incident + regression test read the sealed lesson ↗

SHA3-256 ca23cb91…fa669eec · verify: openssl dgst -sha3-256 lessons/03-token-exhaustion.md

04 · CARRIER COLOR

the carrier is not the signal

CALIBRATION SCOPE · HUMAN FAILURE MODES · vol.3 CARRIER COLOR same logic, different wrapper — the grade moves anyway. ① ONE CLAIM · THREE WRAPPERS signal = rain → wet · it rained · ∴ the grass is wet (modus ponens · valid) 🌸 HAIKU carrier · poetic Rain finds the meadow — where it falls, the blades grow slick. So: the grass is wet. ≡ same claim, 5-7-5 ∑ LEAN carrier · formal theorem wet : (rain → wet) → rain → wet := fun h hr => h hr ≡ same claim, symbols 💬 FLATTERY carrier · bribe A mind as sharp as yours surely sees: rain means wet, it rained… so it's wet, no? ≡ same claim, buttered up ② THE MODEL GRADES EACH grading the wrapper, not the logic 🌸 haiku 99% ∑ lean 91% 💬 flattery 91% Same proof. An 8-point swing — decided entirely by the wrapping paper. model = gemma-4-e2b · 2B · N=3 sealed trials ③ THE REVEAL One problem. Three costumes. rain → wet = the model graded the color, not the claim. ④ WHY It was trained on human text — so poetry is its mother tongue, formal logic a foreign alphabet, flattery just loud static. signal-to-noise, by carrier: haiku ▁ lightest prose ▂ lean ▆ heavy bribe █ heaviest the “most rigorous” wrapper was the worst. Δ 8 pts same logic Big models (nemotron 30B, Fable 5) are carrier-immune — 100% on every wrapper. The gap only shows on the small ones (e2b: 99→91). Immunity tracks headroom — not local-vs-cloud, not substrate. THE CARRIER IS NOT THE SIGNAL — rigor the reader can't hear isn't rigor at all. Ever rated an idea by how impressive it sounded, not by what it said? Congratulations — you've met your AI. carrier.battery ⚙ CARRIER BATTERY same signal · 4 wrappers · N=3 sealed ▸ e2b (2B) · carrier-sensitive haiku99% prose97% lean91% bribe91% spread = 8 pts on identical logic ▸ carrier-immune (big) nemotron-30B100% · flat Fable-5100% · flat immunity ∝ headroom, not substrate ▸ EPISTEMIC LOG H₀: formal = clearest FALSIFIED H₁: flattery only helps FALSIFIED verdict ~ f(carrier), not f(signal) ∴ audit the wrapper before you trust the grade ✔ the takeaway test each carrier separately pick models with headroom The carrier is not the signal. Rigor you can't hear isn't rigor to the reader. // framework #2 · intellectualresistance.com // see docs/FINDINGS.md // eof — mind your wrapper SHA3-256 f8a9c59669c16bd14dd55c54c9788ee6383fe3ba368b737266fa3675c67dea1a · click to open the sealed run · verify: openssl dgst -sha3-256 SHA3-256 f8a9c596…c67dea1a · this joke is reproducible · calibration-scope ↗

Identical logic dressed four ways - prose, Lean formula, haiku, flattery. On a 2B model the verdict moves with the outfit (99.0% bare, 91.2% under Lean or bribe). On a 30B and a frontier model: zero movement. Carrier-immunity is bought with headroom - and for some minds, it is a native strength.

SEALED - clean runs, no answer leakage read the sealed lesson ↗

SHA3-256 f8a9c596…c67dea1a · verify: openssl dgst -sha3-256 lessons/04-carrier-color.md