The context engine, measured.
Proof, not promises. Every number on this page was generated by a script that runs against the live API — real embeddings, real Postgres, no mocks and no marketing math. Run it yourself.
Hard token budget
13 items compiled under a hard 1,500-token cap (tokenizer: chars4-v1, deterministic).
Byte-for-byte deterministic ordering
Two identical compiles produce the same ranked item set — same ids, same order, same scores.
Every excluded candidate carries a reason
13 candidates → 7 included, 6 excluded under a 200-token budget; every dropped one explains itself (budget, branch, staleness…).
Superseded memories never compile
A memory explicitly superseded by a replacement is excluded from every compiled context — stale context cannot silently persist.
Semantic retrieval survives paraphrasing
A paraphrased query ("handle odd input cases") ranks the exact seeded memory first — wording does not need to match.
Cross-agent handoff produces a complete briefing
checkpoint → resume hands the next agent a deterministic briefing (summary, decisions, lessons, next steps) — no re-explaining.
Compile latency (p50 of 3)
Measured live against the production API from a cold connection — includes the embedding round-trip for the task.
Frozen ranking weights
The compiler's ranking vector is frozen in code and echoed in every compile manifest — weights can never silently drift between runs.
| Signal | Weight | What it reads |
|---|---|---|
| semantic | 0.300 | Cosine similarity between the task and the item (real embeddings). |
| lexical | 0.200 | Keyword overlap (Postgres full-text), gated on a real match. |
| branch | 0.120 | Item branch vs the active session branch — mismatches are damped, not dropped. |
| freshness | 0.100 | Recency decay over the item age. |
| authority | 0.100 | Source trust: user > verified code/test > model. |
| confidence | 0.060 | The extraction confidence recorded on the item. |
| evidence | 0.070 | Whether the item is anchored to ledger evidence. |
| target | 0.050 | Direct relevance to the current task title/description. |
Reproduce it
The harness is a single stdlib-only Python script in the repository. No SDK, no setup.
What this page does — and does not — claim
- Measured: the deterministic engine — budget enforcement, ranking determinism, exclusion transparency, retrieval ordering, handoff completeness, latency.
- Not claimed: LLM answer quality, or memory-extraction accuracy — those depend on your model and are validated by your own runs.
- Numbers regenerate on demand; the committed artifact behind this page is timestamped and commit-pinned so you can audit exactly what ran.