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.

PASSbudget

Hard token budget

≤ 1,500 requested · 549 used

13 items compiled under a hard 1,500-token cap (tokenizer: chars4-v1, deterministic).

POST /v1/context/compile with token_budget=1500; sum of item tokens never exceeds the cap.
PASSdeterminism

Byte-for-byte deterministic ordering

identical over 13 items · sha256:19d9caf87ad7

Two identical compiles produce the same ranked item set — same ids, same order, same scores.

Compile twice with identical inputs; compare the serialized (rank, type, id, tokens) sequence.
PASSexclusions

Every excluded candidate carries a reason

6 excluded · 100% with reasons

13 candidates → 7 included, 6 excluded under a 200-token budget; every dropped one explains itself (budget, branch, staleness…).

Compile with a deliberately tight budget (200), then POST /v1/context/explain on the version.
PASSsuperseded

Superseded memories never compile

0 leaks

A memory explicitly superseded by a replacement is excluded from every compiled context — stale context cannot silently persist.

Supersede a memory via POST /v1/memories/:id/supersede, then compile and search the item set for its id.
PASSretrieval

Semantic retrieval survives paraphrasing

top-1 score 0.466

A paraphrased query ("handle odd input cases") ranks the exact seeded memory first — wording does not need to match.

Hybrid search: semantic ∪ keyword candidates, 0.7/0.3 blended boost; no lexical AND-filter.
PASShandoff

Cross-agent handoff produces a complete briefing

2,091-char briefing · {"handoffs": 1, "context_tokens_served": 181, "estimated_re_

checkpoint → resume hands the next agent a deterministic briefing (summary, decisions, lessons, next steps) — no re-explaining.

POST /v1/checkpoints then POST /v1/checkpoints/:id/resume as a different agent; inspect briefing_markdown.
PASSlatency

Compile latency (p50 of 3)

548 ms

Measured live against the production API from a cold connection — includes the embedding round-trip for the task.

Median wall-clock of three POST /v1/context/compile calls, budget 1500.

Receipts generated 2026-09-21T16:04:33Z against https://api.152-70-75-177.sslip.io at commit 8d0f10d — fixture: 48 memories, 16 ledger events, 1 superseded pair, session agent benchmark.

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.

SignalWeightWhat it reads
semantic0.300Cosine similarity between the task and the item (real embeddings).
lexical0.200Keyword overlap (Postgres full-text), gated on a real match.
branch0.120Item branch vs the active session branch — mismatches are damped, not dropped.
freshness0.100Recency decay over the item age.
authority0.100Source trust: user > verified code/test > model.
confidence0.060The extraction confidence recorded on the item.
evidence0.070Whether the item is anchored to ledger evidence.
target0.050Direct relevance to the current task title/description.

Ranking v2 (behind the ranking_utility_v2 workspace flag) scales v1 by 0.95 and adds a 0.05 outcome-utility signal — v1 stays byte-stable.

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.

Curious how it feels in a real workflow? Start with the quickstart or open the dashboard.