Skip to content

Say what the graph is for: it writes, steers and checks the answer - #30

Merged
punnerud merged 1 commit into
mainfrom
feat/positioning
Aug 12, 2026
Merged

Say what the graph is for: it writes, steers and checks the answer#30
punnerud merged 1 commit into
mainfrom
feat/positioning

Conversation

@punnerud

Copy link
Copy Markdown
Owner

The positioning undersold the point — "see the steps drawn as a graph" describes a picture, but the graph is load-bearing. The README now leads with a table where every row is a pinned, CI-checked measurement:

what the graph does measured
writes the answer (strongest path → synthesis) 54.9% → 91.2%
steers exploration (drift/repeat filtered by embedding) tested in explore/settle
gates the answer type (numeric-or-not is code's decision) model chose a number for "capital of France" 3/3 without it
hands arithmetic to mpeqs +32.5 points, 95% CI [+15.3, +49.7], replicated
accumulates knowledge (per-session SQL + RDF) sessions isolated, hints opt-in

And the boundary is stated: not a general knowledge-graph builder for documents/RAG — the graph is built from the model's own reasoning and fed back into it.

README stays short: 83 lines.

The positioning undersold the point. "See the steps drawn as a graph" describes
a picture; the graph here is load-bearing, and every load it bears has a
measured number behind it, pinned in docs/claims/ and re-checked by CI.

The README now leads with that: the strongest path writes the answer (54.9% to
91.2%), embeddings steer exploration away from drift and repetition (decisions a
model cannot make about itself), an embedding gate decides whether a numeric
answer is even on the table, arithmetic settles exactly through mpeqs (+32.5
points where arithmetic is the task, replicated), and every run accumulates into
per-session tables and RDF.

And the boundary is stated rather than implied: not a general knowledge-graph
builder for documents or RAG -- the graph is built from the model's own
reasoning and fed back into it.
@punnerud
punnerud merged commit 17c2051 into main Aug 12, 2026
6 checks passed
@punnerud
punnerud deleted the feat/positioning branch August 12, 2026 15:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant