Context-Window Limits
It read everything you gave it. It did not attend to everything it read.
This page covers one specific way AI gets things wrong at work, and what to do about it.
It runs in order. What goes wrong, why it happens, where you'd notice it on an ordinary day, who takes the blame, roughly what it costs, and the check that catches it. Then one thing to try this week.
The dollar figures are estimates, not measurements. The assumptions behind each one are printed right there, so you can swap in numbers that fit your job. Anything actually measured carries an OBSERVED tag.
What is context-window limits?
Attention across long inputs is not uniform. Information at the beginning and end of a long context is used more reliably than information in the middle — a positional effect that persists even in models with very large stated context capacity.
Two consequences that matter operationally:
1. Position determines reliability. The same fact, placed at token 400 versus token 40,000, does not have the same probability of influencing the output. 2. Nothing reports the omission. A summary built from partial attention is not flagged as partial. It reads complete, because it is complete as a summary — it's the input coverage that was partial, and the output has no way to say so.
You cannot detect this by reading the output. There is no gap where the missing thing would have been.
What do people assume?
That a large context window means uniform attention across it. That if the document fits, the whole document counts equally.
Fitting and attending are different operations.
Where does it show up at work?
A project manager pastes a 60-page requirements document and asks for a risk summary. The response is well-structured and covers eight risks.
The dependency on a third-party API deprecation — one paragraph, page 31 — is not among them. It's the one that slips the schedule by six weeks.
Who carries the downside?
Vendor: none. Executive: approved the timeline. Manager: committed the date. You: produced the risk summary. The retrospective asks how the dependency was missed.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Long-doc analyses / year: 50 Rate omitting a material item: 15% (~7.5 / year) Rate caught by other means: 65% Material misses: ~2.6 / year Cost per miss: $10,000 – $60,000 (schedule slip, rework, scope surprise)
Annualized exposure: ~$26,000 – $156,000
How do you control for it?
Chunked processing with explicit coverage accounting. Split long documents into sections. Process each. Require a per-section output. Then reconcile: every section must produce a line, including "no material items."
The control is not better reading. It is making omission visible — converting a silent gap into an explicit "none found" that someone signed.
CONTROL COST Long-doc analyses: 50 / year Additional time: 35 minutes each Annual: 29 hours Fully loaded rate: $70 / hour
Annualized control cost: $2,030
What should you do this week?
RECOMMENDATION
Stop pasting whole documents. Chunk them, and keep a coverage table — section, processed yes/no, items found.
Two effects, one of which is the real one. First, you catch more. Second, and more valuable: you now have an artifact proving what was reviewed and by whom. When something is missed anyway — and something will be — the difference between "we missed it" and "here is the coverage record showing it was reviewed and here's the gap in the method" is the difference between a performance problem and a process improvement.
That artifact is the entire thesis of Part IV compressed into one table.
Evidence
RESEARCH Positional attention degradation in long contexts — the "lost in the middle" effect.
ANALYSIS The "omission is invisible in the output" argument is the author's framing. Retain the label.
⚠ LIVE EVIDENCE LIBRARY ENTRY — DO NOT PRINT AS A STATIC ENDNOTE This is the single most perishable claim in Part II. Long-context handling is an active area of model improvement, and the positional effect may weaken or disappear in current-generation models between manuscript and print. Required treatment: book text states the mechanism and the control, not a specific performance figure. The empirical status lives at the Evidence Library URL with a visible last-reviewed date and a 90-day review cadence. If the effect no longer replicates: the control still holds. Coverage accounting is justified by the invisibility of omission, not by the positional effect specifically. Update the evidence, keep the chapter.
This is one of 24 failure modes. The book gives you all of them — plus the controls that catch each one and a 90-day plan to prove you ran them.
Preorder the Book