All videosContext engineering2:15

Budget and provenance

That leaves 48 thousand for dynamic information. This is not a universal model property.

Links containing ?t= open the video at a specific second.

Video summary

The ideas to retain

01

The maximum context window is not your operational budget

Assume a provider exposes a total limit \(W\). An application may protect an output allowance \(O\), keep operational headroom \(M\), and reserve fixed mandatory content \(F\), such as…

02

A context item needs more than text

Selection becomes easier to reason about when each candidate is represented as more than a string:

03

Prioritization is not sorting by similarity score

A selector can estimate the utility of candidates, but a production system also needs hard constraints.

Key moments

Jump directly to a section

  1. The window is not the budget
  2. Do not evict the specification
  3. A summary can lose the rule
  4. A derivative needs its source
  5. Caching is not summarising
  6. A pointer must recover the version
Read the reviewed transcript

This video has no narration. This transcript reproduces its on-screen text; it does not invent a spoken track.

The window is not the budget

In this example, total capacity is 64 thousand tokens.

We reserve 8 thousand for output, 4 thousand for margin and 4 thousand for mandatory instructions.

That leaves 48 thousand for dynamic information. This is not a universal model property.

Do not evict the specification

A 42-thousand-token log crowds out necessary files within a 48-thousand-token budget.

We extract relevant failures and keep a reference to the complete log.

The contract and files now fit. The extraction must retain the error that matters.

A summary can lose the rule

The source says: return within fourteen days, except opened items.

A summary keeps the deadline but drops the exception. It looks correct but changes the decision.

The compaction check tests both conditions and rejects this loss.

A derivative needs its source

Summary S derives from document revision A through a lossy transformation.

When the revision changes, S is no longer validated for the current state.

Provenance lets us recover the affected passage without regenerating everything by default.

Caching is not summarising

Reusing a cached prefix avoids repeated processing work, subject to the provider contract.

The content remains part of the logical input. Nothing has been selected or summarised.

Compaction replaces the earlier representation. That transformation can lose information.

A pointer must recover the version

The context stores a short ID for file A rather than every line.

When detail is needed, the read requests revision A and verifies its identity.

If only B is returned, it is not presented as A. Saving context creates a recovery obligation.