fix: context health note — signal omitted details instead of false assurance#434
Merged
Conversation
…surance The context health note previously said 'details are preserved in distilled observations' — which is inaccurate when observations are heavily compressed. The model trusts this and doesn't invoke recall, missing specific details like rejected alternatives and exact error messages. Updated to explicitly state that distilled observations are lossy summaries with specific details likely omitted, and that recall should be used to verify claims before answering questions about specifics. Eval score: 4.62-4.69 at 400K inflation (avg ~4.65, variance-dependent).
BYK
added a commit
that referenced
this pull request
May 21, 2026
…indow (#435) Updates marketing copy with the latest eval results from the recall quality + distillation transparency work (#428, #430, #431, #432, #433, #434). ### README.md - Context retention table: Medium 2.3→4.1, Hard 3.3→4.8, Average 3.9→4.6 - Lore vs tail-window delta: +50%→+77% - Added footnote: Lore scores averaged across multiple runs; TW/compaction baselines from a prior eval run with the same scenarios - Added v6 to version history ### docs/index.html - Hero stat: +50%→+77% vs tail-window - Detail retention: 4.8→4.6 (overall average across difficulty levels, multiple runs) ### Review corrections - Fixed Medium from 4.3→4.1 (honest multi-run average, not cherry-picked) - Average row (4.6) now self-consistent with column values: (5.0+4.1+4.8)/3=4.63≈4.6 - Added footnote clarifying that TW/compaction columns are from a prior eval run
This was referenced May 21, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Updates the context health note injected at gradient layer 1+ to honestly signal that distilled observations are lossy summaries with details omitted, rather than falsely assuring the model that details are preserved.
Problem
The context health note said:
This gave the model false confidence. When it saw a compressed distillation mentioning "flock locking" but not the rejected "staleness check" alternative, it answered "no alternatives were considered" without invoking recall — because the note said details were preserved.
Fix
Updated both variants of the context health note to:
Key changes:
Eval Results (CM-1, 400K inflation)
Two runs: 4.62 and 4.69 (average ~4.65, up from 3.69 baseline).
m3 (the persistent failure) scored 5.0 on one run when the distillation observer preserved the alternative, and 2.2 on the other when it didn't. The remaining variance is in distillation quality (non-deterministic LLM output), not in the context health note.
Tests