agent-harness / docs /HARNESS_CATALOG.md
cuber12's picture
Publish agent harness research code and paper artifacts
d61821a verified
|
Raw
History Blame Contribute Delete
3.2 kB

Harness catalog

Dimension legend

Code Dimension Values
L Advanced lexical retrieval 0 = exact/regex only; 1 = BM25 plus fuzzy retrieval
S Syntax representation 0 = raw lines; 1 = Tree-sitter symbols and chunks
D Dense retrieval 0 = off; 1 = code embeddings
G Structural graph expansion 0, 1, or 2 dependency hops
Q Query policy one-shot or iterative reformulation
I Tool interface unified search tool or specialized tools
P Context packing ranked snippets, skeletons, whole files, or role summaries

Exact path, substring, and regex search remains available in H000-H015 and H020. It is removed only from explicitly labeled controls. When two or more of L, S, and D are active, their ranked lists use reciprocal-rank fusion (RRF).

The 21 initial treatments

ID L S D G Q I P Purpose
H000 0 0 0 0 one-shot unified snippets Exact/regex baseline
H001 1 0 0 0 one-shot unified snippets Lexical main effect
H002 0 1 0 0 one-shot unified snippets Syntax main effect
H003 0 0 1 0 one-shot unified snippets Dense main effect
H004 1 1 0 0 one-shot unified snippets Lexical Γ— syntax
H005 1 0 1 0 one-shot unified snippets Lexical Γ— dense
H006 0 1 1 0 one-shot unified snippets Syntax Γ— dense
H007 1 1 1 0 one-shot unified snippets Full retrieval, no graph
H008 1 1 1 1 one-shot unified snippets One-hop graph reference
H009 1 1 1 2 one-shot unified snippets Two-hop graph
H010 1 1 1 1 iterative unified snippets Iterative-query effect
H011 1 1 1 1 one-shot specialized snippets Interface effect
H012 1 1 1 1 iterative specialized snippets Query Γ— interface
H013 1 1 1 1 one-shot unified skeletons Skeleton packing
H014 1 1 1 1 one-shot unified whole files Whole-file packing
H015 1 1 1 1 one-shot unified role summaries Summary packing
H016 – – – – one-shot unified – No-search negative control
H017 – – – – one-shot unified random Token-matched random-context control
H018 – – – – one-shot unified gold files Oracle-file upper bound
H019 – – – – one-shot unified gold functions Oracle-function upper bound
H020 1 1 1 1 iterative unified snippets Adaptive source router

The unrestricted seven-dimension Cartesian product has 384 combinations. It is not the primary design because most cells would answer no new clean causal question while greatly multiplying model cost. The 21 treatments form focused factorial and ablation blocks. If later evidence justifies another combination, it receives a new file and ID.

Both a full configuration hash and a causal treatment hash are computed. The latter excludes the ID, name, and description so that relabeling cannot conceal a duplicate treatment.