diff --git a/assets/benchmark-agent.svg b/assets/benchmark-agent.svg
index 66b0b15..35c7090 100644
--- a/assets/benchmark-agent.svg
+++ b/assets/benchmark-agent.svg
@@ -1,171 +1 @@
-
- Share of tasks where the model named a file that actually had to change, by condition, for each repository. Whiskers are 95% confidence intervals.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
- Did the model name a file that actually had to change?
- Tasks from real merged commits, indexed before those commits existed. Whiskers: 95% CI.
- no context (memorisation control)
- grep, tripled budget
- reify, three rounds (same cost)
- perfect context (ceiling)
- 0%
- 25%
- 50%
- 75%
- 100%
- tasks with a correct file
- 22%
- none
- 50%
- grep ×3
- 75%
- reify ×3
- 100%
- perfect
- ERPNext (Python/JS)
- n = 40
- 0%
- none
- 28%
- grep ×3
- 78%
- reify ×3
- 100%
- perfect
- OFBiz (Java+XML)
- n = 40
- 0%
- none
- 32%
- grep ×3
- 59%
- reify ×3
- 100%
- perfect
- OpenMRS (Java)
- n = 22
- 0%
- none
- 24%
- grep ×3
- 26%
- reify ×3
- 100%
- perfect
- Medusa (TypeScript)
- n = 39
- Overlapping whiskers mean the difference is not established. On Medusa, reify and grep overlap.
-
+Share of tasks where the model named a file that actually had to change, by condition, for each repository. Whiskers are 95% confidence intervals. Did the model name a file that actually had to change? Tasks from real merged commits, indexed before those commits existed. Whiskers: 95% CI. no context (memorisation control) grep, tripled budget reify, three rounds (same cost) perfect context (ceiling) 0% 25% 50% 75% 100% tasks with a correct file 22% none 50% grep ×3 75% reify ×3 100% perfect ERPNext (Python/JS) n = 40 0% none 28% grep ×3 78% reify ×3 100% perfect OFBiz (Java+XML) n = 40 0% none 32% grep ×3 59% reify ×3 100% perfect OpenMRS (Java) n = 22 0% none 24% grep ×3 26% reify ×3 100% perfect Medusa (TypeScript) n = 39 Overlapping whiskers mean the difference is not established. On Medusa, reify and grep overlap.
- Retrieval quality with no model involved: the share of tasks where the tool put a changed file in front of the agent, per repository.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
- Retrieval only, no model: was the changed file offered at all?
- Every condition held to the same 4,000-token budget.
- 0%
- 25%
- 50%
- 75%
- 100%
- 10%
- grep
- 18%
- path grep
- 58%
- reify
- 75%
- reify ×3
- ERPNext (Python/JS)
- 12%
- grep
- 15%
- path grep
- 70%
- reify
- 78%
- reify ×3
- OFBiz (Java+XML)
- 32%
- grep
- 18%
- path grep
- 41%
- reify
- 55%
- reify ×3
- OpenMRS (Java)
- 18%
- grep
- 18%
- path grep
- 18%
- reify
- 28%
- reify ×3
- Medusa (TypeScript)
- Path grep offers 88 files to reach 18%. Offering everything is not retrieval.
-
+Retrieval quality with no model involved: the share of tasks where the tool put a changed file in front of the agent, per repository. Retrieval only, no model: was the changed file offered at all? Every condition held to the same 4,000-token budget. 0% 25% 50% 75% 100% 10% grep 18% path grep 58% reify 75% reify ×3 ERPNext (Python/JS) 12% grep 15% path grep 70% reify 78% reify ×3 OFBiz (Java+XML) 32% grep 18% path grep 41% reify 55% reify ×3 OpenMRS (Java) 18% grep 18% path grep 18% reify 28% reify ×3 Medusa (TypeScript) Path grep offers 88 files to reach 18%. Offering everything is not retrieval.
- Reify
- Three rows of scattered fragments on the left - code, documents, history - become the same three rows aligned into one solid column on the right. Thin seams remain between them, because compiled knowledge keeps its sources traceable.
-
-
-
-
-
- reify
-
+Reify Three rows of scattered fragments on the left - code, documents, history - become the same three rows aligned into one solid column on the right. Thin seams remain between them, because compiled knowledge keeps its sources traceable. reify
- Reify
- Three rows of scattered fragments on the left - code, documents, history - become the same three rows aligned into one solid column on the right. Thin seams remain between them, because compiled knowledge keeps its sources traceable.
-
-
-
-
-
- reify
-
+Reify Three rows of scattered fragments on the left - code, documents, history - become the same three rows aligned into one solid column on the right. Thin seams remain between them, because compiled knowledge keeps its sources traceable. reify
- Reify
- The Reify mark: scattered fragments resolving into one aligned column.
-
-
-
+Reify The Reify mark: scattered fragments resolving into one aligned column.
- Reify
- The Reify mark: scattered fragments resolving into one aligned column.
-
-
-
+Reify The Reify mark: scattered fragments resolving into one aligned column.
-
-
- reify
-
- The business logic is
- in one person's head.
- Reify compiles it out.
-
- Local
- knowledge engine for AI coding agents
-
- Found the file
- that had to change, same token cost
-
- reify ×3 rounds
- 75%
-
- grep ×3 budget
- 50%
- ERPNext,
- 40 real merged commits, indexed before them
- 3 of 4
- codebases win by 25–50 points; the 4th ties
-
- github.com/lambiengcode/reify
- 11 languages · zero network calls · Apache-2.0
-
+reify The business logic is in one person's head. Reify compiles it out. Local knowledge engine for AI coding agents Found the file that had to change, same token cost reify ×3 rounds 75% grep ×3 budget 50% ERPNext, 40 real merged commits, indexed before them 3 of 4 codebases win by 25–50 points; the 4th ties github.com/lambiengcode/reify 11 languages · zero network calls · Apache-2.0