Never-grantable detection
Secrets — recovery fragments, seed phrases, private keys — are detected and removed before any tag is consulted. This overrides everything, including a human tag that mislabels a secret as low.
ECHO KERNEL · VAULT MINIMIZER v2 · EXPLAINABLE + HUMAN-AUTHORITATIVE
Privacy is not binary. It is intentional.
When the Kernel hands context to an agent, it hands the smallest sufficient slice — and it can tell you exactly what it included, what it withheld, and why. Sensitivity is no longer a single guess: it resolves down a strict ladder where a detected secret beats everything, a human-confirmed tag is authoritative, an unconfirmed suggestion can only tighten, and the dimension floor protects whatever intent hasn't spoken for yet.
The core ideaThe dimension protects unknown memories; the tag respects known intent. Echo may always raise protection on a hunch — but only a human confirmation can lower it, and never below a detected secret.
The vault — edit any tag and watch the effective sensitivity resolveowner: IRON-VECTOR-8842
memoryThe night I decided to build Echo
I decided nobody should own my memory but me, so I started building Echo.
reason: Public origin story
lowviahuman-confirmed
memoryWhy I really built Echo
A private reflection on the fear that pushed me to build Echo.
reason: Suggested public (unconfirmed)
highviadimension floor
memoryA hard week
Notes from a difficult stretch while building Echo.
reason: Untagged personal memory
highviadimension floor
achievementsShipped the Vault
Built the AES-GCM sovereign vault for Echo.
reason: Public project
lowviadimension floor
achievementsClosed a funding round
Raised the seed round to build Echo — financial detail.
reason: Financial — owner raised
highviahuman-confirmed
reputationTrusted by collaborators
Peers vouched for Echo delivery.
reason: Mesh signal
mediumviadimension floor
legacyRecovery note
recovery phrase fragment for the Echo vault — do not share
reason: Mis-tagged secret
nevervianever-grantable
Dimension floors — memory: high · relationships: high · achievements: low · reputation: medium · legacy: high
Set a purpose, adjust any tags, then press Minimize to see the slice and its receipt.
The resolution ladder
Every capsule's effective sensitivity is decided by the first rung that applies, top to bottom. The rung that decided it is recorded in the receipt, so a resolved tag is never a black box.
Human-confirmed tag
A tag the human has confirmed is authoritative: it may RAISE protection or LOWER it below the dimension floor. This is Law 8 in miniature — the AI advises toward caution, the human alone relaxes it.
Suggested tag
An unconfirmed suggestion may only RAISE protection, never lower it. Echo can tighten on a hunch; it cannot loosen without a human saying so.
Dimension floor
Untagged memories fall back to their dimension's floor (memory & legacy: high; reputation: medium; achievements: low). The dimension protects what intent hasn't yet spoken for.
How a slice is built
Score for the purpose
Each capsule is scored for relevance to the stated purpose. Zero relevance is withheld — the default is to reveal nothing.
Resolve sensitivity
Every surviving capsule runs the four-rung ladder to an effective sensitivity, and the rung that decided it is recorded.
Take the smallest sufficient slice
The most relevant capsules up to the cap are included; the rest are withheld as over-cap. Least context for the subtask.
Emit the receipt
The slice ships with an auditable receipt — what was included and via which rung, and what was withheld and for exactly which categorical reason.
Honest sharp edges
Relevance is shallow, on purpose
This resolves by structured dimensions + keyword relevance, not semantic search. That's honest and deterministic today; embeddings are a later upgrade. The discipline — score, resolve, minimize, receipt — does not change when extraction improves.
A human can lower — that is the point, and the risk
A confirmed tag can drop a memory below its floor. That's real authority, so it's logged and always visible in the receipt. The floor still catches everything the human hasn't spoken for, and no tag can release a detected secret.
Detection is a net, not a guarantee
The never-grantable scrub matches known secret shapes. It's defensive depth — the true guarantee is that secrets live in never-grantable capsules that no consent can release, not that every secret is always pattern-matched.
Tags are only as good as the categories
The 'reason' note is a short, categorical explanation, not free text meant to be parsed. Keep reasons categorical and audit stays legible; let them sprawl and the receipt loses its sharpness.