SELF-RUN ADAPTER EVIDENCE · NOT LEADERBOARD-IDENTICAL

Context-Bench pilot: context compilation beats stuffing, at 97% fewer tokens

A 15-question pilot on the public letta-evals filesystem suite, judged by the official strict rubric. The Perseus DAG assembly (perseus#962) outscores full-context stuffing while rendering 2.8% of its tokens, on par with naive RAG arms.

0.167
full-context
mean rubric, n=15 · 20,187 tokens/question rendered
0.333
naive RAG k=3
mean rubric, n=15 · 267 tokens (−98.7% vs full-context)
0.233
naive RAG k=5
mean rubric, n=15 · 360 tokens (−98.2%)
0.267
Perseus DAG
mean rubric, n=15 · 573 tokens (−97.2%) · beats full-context by +0.100

What this measures

Context-Bench is Letta's filesystem-reasoning suite: an agent gets a set of files and strict short-answer questions (rubric.txt, 0 / 0.5 / 1.0). The official target is the Letta Code agent; this pilot adapts the identical dataset and judge to Perseus assembly arms. Canonical results: leaderboard.letta.com.

Armsfull_context: every file stuffed into the answer prompt at the official 8,000-char window. naive_rag_k3/k5: cosine top-k over chunk embeddings. perseus_dag: the auditable context-compilation DAG (perseus#962) — query root → typed evidence gates → selective expansion under a hard token budget, every edge versioned and digest-sealed.

Judge: the official rubric.txt verbatim (SHA-256 9c124fc9…), gpt-5-mini, strict 0/0.5/1.0 parse. Disclosed deviation: the rubric YAML pins temperature 0.0, which the gpt-5-mini API rejects (HTTP 400) — the judge runs at temperature 1.0. Answer model: gpt-5-mini, temperature 1.0.

Tokens: rendered tokens = chars ÷ 4 of what each assembly passes the answer model. Provider-reported usage is captured per call; the full-context arm's real prompt usage averaged ~30k provider tokens per question.

Honest scope. This is a self-run adapter pilot: public dataset questions (not the hidden holdout), n=15, one answer model, and the DAG harness is not the official Letta Code agent. These numbers are adapter evidence — they do not validate any leaderboard claim and are not statistical validation. Absolute scores are low across every arm: strict-judge reference questions are hard for a single-call gpt-5-mini, regardless of context strategy.

Per-question rubric (n=15)

questiontypefullrag3rag5dag
Among all people who live in the same state as t…comparison_tiebreak0.00.50.50.5
Among people living in the same state as the own…comparison_tiebreak0.00.00.00.5
How many total records (bank accounts, vehicles,…cross_file_counting0.00.50.00.0
Who owns more vehicles: the person with the high…multi_hop_chain0.50.00.00.0
Who owns more vehicles: the person with the most…multi_entity_comparison0.50.00.50.5
Among the 15 people with the highest total bank …set_intersection0.50.50.50.0
Who has a higher total bank account balance: the…multi_hop_chain0.01.00.50.5
What is the total bank balance of the person wit…aggregation0.00.50.00.5
Who has more bank accounts: the person with the …multi_entity_comparison0.50.00.50.0
What is the combined bank balance of the 3 peopl…aggregation0.00.50.50.0
Who has more credit cards: the person with the h…multi_hop_chain0.00.00.00.0
Who has more credit cards: the person with the h…multi_entity_comparison0.50.50.01.0
Who has more internet accounts: the person with …multi_hop_chain0.00.50.50.0
Among people living in the same state as the own…temporal_reasoning0.00.00.00.0
Who has more credit cards: the person with the m…multi_entity_comparison0.00.50.00.5

n=15 pilot · gpt-5-mini answer + judge · run 2026-08-15 · 60 answer calls, 60 judge calls, 538,450 provider tokens total (≈ $1 of API spend)

Custody

Every input and the results envelope are digest-sealed and verifiable from the source tree:

pilot sha256 12f842cf270b0bfe3323ea59fb842e2091e6d98fdc6881a84a8961fd7085339a
rubric sha256 9c124fc9cc34841d1bc0c601d13a96246986ca4caaa7730eb3abbd6e8761b385
results digest defee0f5ddcffcd0d027979ff1a2f7a732d8c2faa29bf8cb523e3be19a74df65

Verify with python3 benchmark/context-bench/custody.py in Perseus-Computing-LLC/perseus. Rerun the live pilot with python3 benchmark/context-bench/run.py (requires an OpenAI-compatible key; --dry-run reproduces the full pipeline at zero spend).