Summary and key takeaways

The one picture. Security is judged on its minimum, against an adversary who sees the design and then picks the attack. There are three control surfaces, applied in order of desperation: deny the request; if you cannot, obscure the bytes; if you cannot, destroy the signal. A fourth, detection, runs underneath because the first three fail silently. Each surface answers a different adversary, and the order is the threat model.

The five adversaries. External attacker, malicious insider, careless insider, platform operator, authorized analyst. Nothing stops all five. A control without a named adversary cannot be evaluated.

By Part, one sentence each:

1. Security has an adaptive adversary and an outside specification; every control must name its adversary, and disclosure cannot be undone.
2. Authentication ladder: do not transmit secrets, use short-lived and context-bound credentials, use workload identity. Authorization compresses a huge matrix: roles explode, attributes lose the inverse query, relationships make revocation a write. Expressive policy is hard to audit.
3. Enforce at the narrowest waist: row and column predicates, views, vended short-lived scoped credentials, careful delegation, egress control. Grants accumulate because controls decay toward permissive.
4. Encryption at rest, in transit, envelope keys, rotation versus crypto-shredding, masking, tokenization, salted hashing. Every property you preserve to keep utility is a property you publish.
5. Privacy: when everyone is authorized, the only lever is destroying information. k-anonymity fails on composition and dimensionality; differencing breaks aggregate interfaces; DP gives composable bounds with noise 1/epsilon and a depleting budget; minimization, erasure, consent, clean rooms and synthetic data each have a price.
6. Weak points: stale principals, service accounts, caching of decisions, egress, tenant side channels, integrity and nonces, key custody, residency, and lack of detection.
7. The mechanism forbids thirteen things, and five risks sit outside all three surfaces.

Laws used (old): granularity mismatch, fixed cost sets the batch size, Little's Law, costs are relocated never removed, immutable storage grows an indirection layer plus a collector, independence is assumed and rarely true.

Laws added by this mechanism:

The three that travel best: controls decay toward permissive, only bounds compose; states do not, and prevention fails silently.

What this mechanism forbids. You cannot enforce above the narrowest waist; have stateless verification with immediate revocation; keep authorization context across an asynchronous boundary; list effective access under an expressive policy; protect data from its own processing path; compress ciphertext; encrypt a column and keep pruning; anonymize deterministically; bound inference with cell-level access; compose syntactic privacy; delete cheaply from a model; combine key custody with server-side compute; or exceed the jurisdiction of the enforcement point.

Five residuals. Availability (backups versus erasure, legal hold); arbitrary code running with pipeline credentials; data that has left the organization; people as the first step in (phishing); friction that selects for workarounds. Security posture is a rate, not a state.

The story. Wei's table grew from "control x adversary" to include time to detect and a final row, "friction", with an owner. Hollis's board note states a position and a named adversary, never "we are compliant".

The one line. Utility and disclosure are the same operation seen from opposite sides. That is why this mechanism has no solved state, only a chosen position and a named adversary it is chosen against.