CK·Knowledge
Governed knowledge · built on BRaaS & the BAMP harness

Knowledge that knows
its own boundaries.

A governed knowledge platform that keeps your sovereign substrate private by design — and proves a 4 KB structured capsule gives an agent the same task performance as 154 KB of raw context.

What it is

Search your knowledge — without leaking your moat.

Most knowledge tools publish everything you give them. CK Knowledge classifies every document deny-by-default: product and narrative content is searchable; your governing substrate — the rules, decisions, and mechanisms that are your competitive moat — is demand-paged to trusted agents and never published.

🛡️

Governed by design

A deny-by-default classifier + an adversarial review pass keep mechanism out of the public surface. The boundary is enforced in code, re-proven every cycle.

Provably efficient

A ~4 KB interpretive capsule matches full-context performance at ~1/25th the tokens — measured across 8 models, not asserted.

🔎

Semantic + narrative

Vector search over your product corpus plus the captured development story — the arc, decisions, and lessons, queryable.

The proof

4 KB > 154 KB. Measured.

In a pre-registered evaluation (367 cells, 8 models, $7.66), a small structured capsule out-performed every larger-context regime — and collapsed the gap between model tiers (a small model + the capsule matched a frontier model).

4 KB capsule
3.29
154 KB full context
2.42
158 KB raw narrative
2.09
Empty (no context)
1.59

Mean task score (1–5). Independent evaluation; the methodology is portable to any long-running project facing context-rot. Tier-collapse measured: Haiku ≈ Opus once the capsule is present.

25–40×
context compression
0
moat rows published (proven)
8
models benchmarked
$7.66
total eval cost

Sovereignty

Your substrate stays yours.

The principle: demand-page the moat, don't publish it. Your governing rules are the reason customers choose you — so they belong in a contract zone, never on a public surface.

  • Deny-by-default: unknown content is excluded until explicitly approved.
  • Defense-in-depth: a deterministic classifier and an adversarial semantic review — they catch different failure modes.
  • Continuously re-proven: a moat invariant runs every cycle; drift raises an alarm.
  • Auditable: every classification + every publish decision is recorded and re-derivable.