When overlong context is truncated, what the user sees is unexplained capability loss

What it is

Early in a long chat the user uploaded a spec sheet. Halfway through, the system starts to behave as if it never saw the sheet: numbers drift, clauses have to be pasted again. The window filled; early turns were dropped or replaced by a coarse summary, and the UI said nothing like “I can no longer see that sheet.” What the person sees is unexplained capability loss: the same product suddenly went dull, and the only available attributions are “it is unstable” or “I pasted the sheet wrong.”

Truncation is an engineering necessity. Silence is the interaction failure. Capability can drop; the drop has to be named as missing material, not performed as the model forgetting how.

Why it happens

Context has a hard length. Past it, common tactics drop the earliest turns, roll a summary, or keep only spans retrieval thinks relevant. If the spec sheet sits in the dropped zone, later generation is unconditioned on it. The model still answers fluently, so the absence is fluent-filled — a “plausible spec” from the training distribution stands in for the user’s sheet. Fluency disguises truncation as “a so-so answer” rather than “the material is gone.”

The user side has no length gauge. A scrollbar implies the whole history is still present; the mental model is “it has been looking at all of it.” Gauge and source of truth split again, and failure shows up as causeless decline.

Studying it

Place material that a later task will need at the front of the session, push the window past the truncation point with filler turns, then ask questions that depend on that material. Independent variables: truncation policy (drop-head, summarise, retrieve), whether the UI states that the material is gone, how specific that statement is. Dependent variables: accuracy on material-dependent items, user attribution (model got worse / I was unclear / material was lost), time to repair after noticing the loss.

Code “fluent but factually from the wrong sheet” separately from “admits the material is gone.” The former is the typical face of silent truncation and must not count as success.

Where it stops holding

Sessions that always fit the window never truncate; this entry does not appear. When material is externalised in a tool (the file can still be pulled by reference), the window lost dialogue text, not the file; the cue should change to “the excerpt in the dialogue is gone; the file can still be read.” A summary the user confirmed, with a path back to source, turns truncation into compression and is no longer silent. Retrieval can kill early spans that look irrelevant and are in fact constraints, which is harder for users to foresee than uniform drop-head.

Applying it

• When a still-cited piece of material is about to leave, or has left, the window, put a clear state by the field: “the spec sheet is not in current context; pin it or paste it again if you still need it.” Do not only write truncating in an internal log.
• Offer pins: files, constraints, outlines the user can keep in the window. Pins have a quota; when the quota is full, say so. Do not silently evict the oldest pin.
• For spans replaced by a summary, allow a click through to whether the source the summary was based on still exists. If the source is gone, mark the summary as unverifiable.
• Check: put required material in turn one, fill until truncation is certain, then ask for a unique figure from the material. With no cue, a wrong figure plus a fluent answer is a silence incident. With a state cue, the user should choose to re-pin or re-paste before asking; if they keep asking with no cue and still get a wrong figure, the cue never entered the decision point.

Related

• Same group: L2.14.1 Context buys continuity, and it also lets an early misunderstanding keep shaping later output · L2.14.2 People usually cannot tell what the system actually remembered; the context needs an inspectable view · L2.14.3 Clearing needs an explicit act; opening a new session is only an implicit convention · L2.14.4 Memory across sessions hardens a one-shot preference into a standing setting
• Nearby: L1.06 Graceful Degradation of AI Failure · L1.12 Latency and Streaming Experience · L5.03 Trust Calibration
• Search terms: silent context truncation · context window overflow · unexplained capability drop