Repository navigation
Release 0.64.0 — the numbers are on the door, and every one has a source - #18
Merged
Merged
Conversation
Signed-off-by: Jimmy Wesley <rededown@hotmail.com>
The README opened with a 120px logo and nothing else, so the first screen
said the project's name and made no claim. The four-figure strip is the shape
that works on repositories people star: a number, what it measures, and the
condition it was measured under, in one line each.
The four are quoted, and STATS carries the source of each beside it:
11/11 multi-hop QA, same 12B model that scores 0/11 as top-k RAG
(paper section 5, abstract)
1.3 ms entry search p95 at recall@5 = 1.0 with no embedder
(paper section 5.1, the BM25 scent-weighted row)
0.66x token cost per correct answer against iterative RAG
(paper section 5: 1433 against 2175)
1,309 tests collected at this commit
(pytest --collect-only)
The condition is not decoration. 11/11 without "same 12B model" reads as a
frontier model doing well; 1.3 ms without "no embedder" reads as a vector
index somebody paid for; 0.66x without "per correct answer" inverts, because
the monkey's raw median is comparable to iterative RAG's and the ratio comes
from the questions it does not fail. A figure stripped of its condition is
the same defect as the 0.58x this branch just corrected, arriving by a
different road.
The card does not get the strip. It is read small and in a timeline, where
four columns collapse into grey texture, so it keeps the one comparison that
survives being small. Same palette, same wordmark, same tagline: two images,
one voice.
build_social_card.py becomes build_brand_art.py, since it now builds both.
Signed-off-by: Jimmy Wesley <rededown@hotmail.com>
The README opened with a 120px logo and nothing else, so the first screen
said the project's name and made no claim. docs/banner.png is the four-figure
strip that works on repositories people star: a number, what it measures, and
the condition it was measured under, one line each.
The four are quoted, and STATS names the source beside each:
11/11 multi-hop QA, same 12B model that scores 0/11 as top-k RAG
(paper section 5, abstract)
1.3 ms entry search p95 at recall@5 = 1.0 with no embedder
(paper section 5.1, the BM25 scent-weighted row)
0.66x token cost per correct answer against iterative RAG
(paper section 5: 1433 against 2175)
1,309 tests collected at this commit
(pytest --collect-only)
The condition is not decoration. 11/11 without "same 12B model" reads as a
frontier model doing well. 1.3 ms without "no embedder" reads as a vector
index somebody paid for. 0.66x without "per correct answer" inverts, because
the raw median is comparable to iterative RAG's and the ratio comes entirely
from the questions it does not fail. A figure stripped of its condition is
the same defect as the 0.58x this tree corrected last commit, arriving by a
different road.
The card does not get the strip. It is read small and in a timeline, where
four columns collapse into grey texture, so it keeps the one comparison that
survives being small. Same palette, same wordmark, same tagline: two images,
one voice.
The rename to build_brand_art.py landed alone in 44084ca, and that commit's
message describes this one. The add that should have staged the content named
the path the rename had already retired, so it failed, and the commit went
ahead with the only thing staged.
Signed-off-by: Jimmy Wesley <rededown@hotmail.com>
The social preview carried one comparison while the README banner carried four, so the two images arriving from the same project said different things about it, and the one that travels furthest said least. The card is what a timeline, a chat and a search result render; the banner is seen by somebody already reading the repository. The strip is now one function serving both sizes. STATS holds each figure once, with two lengths of note: at 1280x640 a column is 280px against the banner's 425, so "same 12B model, 0/11 as top-k RAG" does not fit and becomes "vs 0/11 as top-k RAG". What the short form drops is the half already carried by the subline above it or by the paper, never the half that makes the number mean something else: "per correct answer" stays, because 0.66x without it inverts. Sizes are what they are for a reason and now both are usable. 1280x640 is what the social preview field takes, and an image at the banner's 2000x760 is cropped to 2:1 there, which cuts exactly the strip that was added. Signed-off-by: Jimmy Wesley <rededown@hotmail.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
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.
Release 0.64.0 — the numbers are on the door, and every one has a source
Where the engine work is. The v0.64 commits — A transport method is not a
capability, The listen stream is served, The receiver is not the record —
already merged into
mainin #17, which was titled for the discoverabilitycommit it started as. So this release PR carries the version bump alone
(
pyproject.toml,src/monkeyllm/__init__.py, 0.63.0 → 0.64.0) plus the brandimages. Anyone opening this diff looking for the spec v0.64 surface should read
#17 instead.
docs/monkeyllm-spec-v0.64.mdand theCLAUDE.mdpointer arealready on
main.The strip
The README opened with a 120px logo and no claim, and the social card carried
one comparison. Both now carry the same four figures, and
STATSin thegenerator names the source of each beside it:
pytest --collect-onlyThe condition beside each number is not decoration.
11/11without same12B model reads as a frontier model doing well.
1.3 mswithout no embedderreads as a vector index somebody paid for.
0.66×without per correct answerinverts, because the raw median is comparable to iterative RAG's and the
ratio comes entirely from the questions the monkey does not fail. A figure
stripped of its condition is the same defect as the
0.58×that #17 corrected,arriving by a different road.
Two sizes, one set of claims
docs/social-preview.pngvs 0/11 as top-k RAGdocs/banner.pngsame 12B model, 0/11 as top-k RAG1280×640 is what GitHub's social preview field takes; an image at the banner's
2000×760 is cropped to 2:1 there, which cuts exactly the strip that was added.
A column is 280px on the card against 425px on the banner, so the note has to
lose something — and what it loses is the half already carried by the subline
above it or by the paper, never the half that changes what the number means.
STATSholds each figure once, in both lengths, so the two images cannotdrift apart without editing one list.
One commit is a rename with the wrong message
44084cacarries the message that describescdeef44. Thegit addnamed thepath the rename had already retired, failed whole, and the commit went ahead
with the only thing staged.
developis protected against non-fast-forward andwas being pushed to concurrently, so this was not rewritten:
cdeef44's ownmessage ends by saying what happened.
Still manual after merge
GitHub has no API for the social preview: Settings → General → Social preview
→ Upload an image,
docs/social-preview.png. Until thenog:imagestays theauto-generated card. X and LinkedIn cache aggressively; their Post Inspector /
Card Validator force a refetch.
Files
pyproject.toml,src/monkeyllm/__init__.py: 0.64.0README.md: the banner replaces the 120px logo, alt text carrying the figuresdocs/banner.png,docs/social-preview.pngscripts/build_social_card.py→scripts/build_brand_art.py, which buildsboth from one palette and one
STATSNo source under
src/monkeyllm/orapps/is touched beyond the versionstring.
🤖 Generated with Claude Code