Evidence Library

The Evidence Library

Every empirical claim in this work is labeled by epistemic status and linked to a source page maintained after publication. Where a finding may not survive the next model generation, that is stated in the text and its current status is tracked here with a review date.
Last reviewed 2026-07-30 · Review cadence 90 days

Why does a static citation go stale?

A printed endnote asserts a fact permanently. But research on model behavior moves — a positional attention effect documented in 2023 may not replicate in a 2026 long-context model. An endnote cannot tell you that. A maintained source page can.

So the book states the mechanism and the control; the empirical status lives here with a date on it.

What do the labels mean?

LabelMeaning
RESEARCHDocumented finding from published research. Primary source cited.
REPORTEDPublicly reported case. Primary source, not press coverage.
ANALYSISThe author's reasoning. Not independently verified.
MODELEDArithmetic from stated assumptions. Not a measured outcome.
RECOMMENDATIONCareer advice. Judgment, not evidence.

Which claims are perishable?

Mode #6, Context-window limits is the most perishable claim in the taxonomy. Long-context handling is an active area of improvement and the positional effect may weaken. The control holds regardless — coverage accounting is justified by the invisibility of omission, not by the positional effect specifically.

Source status by mode

Failure modeExposure figuresLast reviewedCadence
01 HallucinationMODELED2026-07-3090 days
02 Knowledge CutoffMODELED2026-07-3090 days
03 No True UnderstandingMODELED2026-07-3090 days
04 OverconfidenceMODELED2026-07-3090 days
05 Training-Data BiasMODELED2026-07-3090 days
06 Context-Window LimitsMODELED2026-07-3090 days
07 No Common-Sense GroundingMODELED2026-07-3090 days
08 Prompt InjectionMODELED2026-07-3090 days
09 No Persistent MemoryMODELED2026-07-3090 days
10 SycophancyMODELED2026-07-3090 days
11 Reasoning FragilityMODELED2026-07-3090 days
12 No Real-World VerificationMODELED2026-07-3090 days
13 Verbosity BiasMODELED2026-07-3090 days
14 Uncertainty MiscalibrationMODELED2026-07-3090 days
15 Training-Data QualityMODELED2026-07-3090 days
16 Lack Of AccountabilityMODELED2026-07-3090 days
17 Temporal ReasoningMODELED2026-07-3090 days
18 Mathematical FragilityMODELED2026-07-3090 days
19 Cannot Truly CiteMODELED2026-07-3090 days
20 No Sensory GroundingMODELED2026-07-3090 days
21 Cross-Session InconsistencyMODELED2026-07-3090 days
22 Difficulty With NegationMODELED2026-07-3090 days
23 Cultural And Linguistic Blind SpotsMODELED2026-07-3090 days
24 No True CreativityMODELED2026-07-3090 days

Revision log

When something here turns out to be wrong, the correction is published with what changed and why. Publicly correcting your own claim is the highest trust-per-word content that exists, and almost nobody runs it.

DateChange
2026-07-30Evidence Library established. Source audit applied to Part III: figures previously attributed to named research firms replaced with placeholders and labeled illustrative.