Announce the gene by its name, not by 71 characters of hash - #24
Merged
Merged
Conversation
The injected header opened with the content address:
[Evolution Memory] Strategy reused from Gene sha256:849680dd…86 (EvoMap network):
Seventy-one characters the model cannot act on, in the most prominent line of
the injection, on every turn that primes. The header is where the model decides
whether the strategy is worth reading, and a hash tells it nothing.
It now opens with the asset's short_title, which is both shorter and worth
reading: measured across a live search, titles run 4-37 characters against the
hash's 71, so the header gets smaller and gains the only signal it was missing.
The hash is still needed -- evolver_asset_reuse_result takes an asset_id, and
the injection is the only place the model learns it -- so it moves to the line
that asks for the report, where it is used.
`summary` was the obvious source for a name and is deliberately not used: it
runs to a 214-character median, so it would have cost three times the hash it
replaced, and #8 already decided a summary does not belong in the prompt. The
fallback is the asset type, never the summary, and a test pins that.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The T1 correction test slept a fixed 150ms for a fire-and-forget POST, and went red on CI while passing locally five times running. The line above the sleep was meant to avoid exactly that and could not: it filtered `requests` into a new array and then polled that snapshot, which never grows. It now polls the live array for the report it is waiting on, and the second-complaint assertion gets a longer settle window, since proving a request did NOT arrive is the one case where waiting is the whole point. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
forrestlinfeng
added a commit
that referenced
this pull request
Sep 23, 2026
#24 put short_title in the header and a live session showed what that costs when the Hub's title is truncated: [Evolution Memory] Object (EvoMap network): 1. Initialize pool with configurable minimum and maximum object capacity limits … `Object` is what got cut off, not what the gene is. Measured across forty titles from eight queries, two were fragments and a character floor cannot find them: `自动化小` is a truncation at four characters, and `智能缓存优化` is a complete title at six. Chinese says as much in six characters as a Latin title says in forty, so length is the wrong instrument. A title now has to look like a name: more than one word if it is Latin, five characters or more if it is CJK. `Object` fails on the first and `智能缓存优化` passes on the second. Whatever fails gives way to nl_summary, present on 39 of the 40 measured, cut to the same 80 characters the title was. That does put a description in the header, which #8 kept out. The bound is what makes it different: 80 characters is a name-length excerpt standing in for a name, not the 214-character median recitation #8 rejected. Four tests; removing the fragment check fails two of them. Co-authored-by: forrestlinfeng <linfengchen@tencent.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
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.
Before / after
Why
The header is where the model decides whether the strategy is worth reading, and it opened with 71 characters of content address the model cannot act on — on every turn that primes.
The id is kept, and stays visible in the trace for debugging. It moves to the line that asks for the report, because
evolver_asset_reuse_resulttakes anasset_idand the injection is the only place the model learns it. Deleting it would have silently broken self-reporting.Why
short_titleand notsummarysummaryis the obvious source for a name. Measured against a live Hub:asset_idshort_titlesummarynl_summarySwapping the hash for a summary would have cost three times what it replaced — the opposite of the intent.
short_titleis both shorter than the hash and the only one of them worth reading:There is also a standing decision here: #8 kept the summary out of the prompt, and
test/prime.test.jspinned it with adoesNotMatch. The first draft of this change reintroduced it through a fallback chain and that test caught it. The chain is nowshort_title→ asset type, never the summary, and two tests hold that line.Net effect
Header loses 71 characters of noise, gains ~30 of signal. Smaller, and it now says what the thing is.
Tests
93/93. Two added: a named gene is announced by its name with no hash in the header and the hash present on the reporting line; an unnamed one falls back to its type rather than reciting a summary. Both assert the long fields stay out.
Five existing assertions moved from the old header text to the new reporting line.
🤖 Generated with Claude Code