사양

CAP Requirements Baseline (EARS) — Traceability Anchor

이 장은 아직 한국어로 제공되지 않아 영어 원문을 표시합니다.

Status: Working draft for H1-PROTO-01. This is the sole authoritative mapping between the CAP Final specification body and its requirements, mirroring Faying §1.3's "Goal ↔ Requirement one-to-one" discipline. Every hard constraint in the drafted normative chapters (03–06) cites Trace: R<n>.<m>.<k> back to a clause here. EARS forms used: Ubiquitous ("The CAP terminal SHALL …"), Event ("WHEN …, the CAP terminal SHALL …"), State ("WHILE …, the CAP terminal SHALL …"), Unwanted ("IF …, THEN the CAP terminal SHALL …"), Optional ("WHERE …, the CAP terminal SHALL …"). Convention: SHALL = absolute (RFC 2119 MUST); SHALL NOT = MUST NOT; SHOULD/MAY as RFC 2119. "CAP terminal" = the Protocol_Engine + Descriptor_Validator + session manager on the terminal acting as the Faying Relying Party for terminal resources. Open architecture decisions are tagged [需人类决策 D-n] and aggregated in README.md.

Goal ↔ Requirement map

Design goal (outline §1.3)	Requirement	Faying counterpart
D1 Accountability-Binding	R1	Faying R1 (Unique Accountability)
D2 Default-Deny	R2	Faying R2
D3 Offline-First Availability	R3	— (CAP-specific)
D4 Graded Control aligned to Faying	R4	Faying R6 (Graded Authorization)
D5 Destructive-Operation Confirmation	R5	Faying R13.4 (Attention-Drift / high-sensitivity ≥ G1)
D6 Atomic Handover	R6	— (CAP-specific)
D7 Liveness / Zombie-Reclamation	R7	Faying R5 (Continuous Calibration)
D8 Observable-Auditable-Revocable	R8	Faying R3 / R4
D9 Interoperability (CBOR+JSON)	R9	Faying R10
D10 AI-era Threat Resistance	R10 (STRIDE) + R12 (AI-era)	Faying R9 / R13
— (crypto evolvability)	R11	Faying R8 (PQC)
— (resource access model)	R13	— (CAP-specific)

R1 — Accountability binding to a Faying Mandate

Goal D1. Every control authority CAP grants must be unambiguously attributable to exactly one Human Prime (Natural_Person) or Official_Post, via a live Faying Mandate.

R1.1 The CAP terminal SHALL treat each accepted AuthRequest as bound to exactly one accountability endpoint (Human Prime / Official_Post), reconstructable from the audit record as the tuple (descriptor_id, mandate_ref, subject_fay_id, terminal_id, access_mode, operation_id?, ts).
R1.2 WHEN an AuthRequest is not accompanied by a verifiable binding to a live Faying Mandate, the CAP terminal SHALL reject it under Default-Deny (E_MANDATE_MISSING). [需人类决策 D-1: exact binding model]
R1.3 IF the bound Faying Mandate's attestation chain does not resolve to a single accountability endpoint, THEN the CAP terminal SHALL reject the request (E_MANDATE_INVALID) and write the event to the audit log.
R1.4 The CAP terminal SHALL NOT create or continue a Session whose accountable endpoint cannot be determined.

R2 — Default-Deny

Goal D2. Absence, ambiguity, or unverifiability resolves to denial.

R2.1 IF any field required for a control decision is missing, malformed, or unverifiable, THEN the CAP terminal SHALL reject the request and SHALL NOT grant partial access.
R2.2 IF a required dependency (revocation status, presence freshness, mandate validity, operation-criticality classification) is unavailable, THEN the CAP terminal SHALL converge to denial rather than optimistic pass.
R2.3 The CAP terminal SHALL grant access only when all decision predicates (R1, R3–R6, R13) hold simultaneously.
R2.4 IF an operation's criticality cannot be determined, THEN the CAP terminal SHALL treat it as critical (the strictest class) and apply R5. [需人类决策 D-5]

R3 — Offline-first availability

Goal D3. Loss of network SHALL NOT strip a previously-authorized Fay of legitimately granted, non-critical control.

R3.1 WHILE the terminal is offline, the CAP terminal SHALL continue to validate Authorization_Descriptors locally and honor grants whose grade/criticality requirements can be satisfied offline.
R3.2 IF an operation's criticality requires online revocation confirmation or fresh online presence that cannot be obtained offline, THEN the CAP terminal SHALL reject that operation while still serving offline-satisfiable operations.
R3.3 The CAP terminal SHALL bound offline trust by descriptor validity ceilings (≤ 90 days offline, ≤ 7 days for converted tickets) so revocation delay is capped.

R4 — Graded control aligned to Faying G0–G4

Goal D4. CAP control authority is gated by the Faying grade carried by the bound Mandate; CAP consumes the grade, it does not redefine it.

R4.1 The CAP terminal SHALL read the Faying grade G ∈ {G0,G1,G2,G3,G4} of the bound Mandate and SHALL reject any access whose grant min_grade is stricter (higher security / lower autonomy) than G (E_GRADE_INSUFFICIENT). Autonomy order per Faying §7.1: G0 < G1 < G2 < G3 < G4.
R4.2 WHEN the bound Mandate's presence signal is stale for the current grade (per Faying Appendix B max_staleness_ms: G1≤5s, G2≤60s, G3 key-step intent, G4≤30min), the CAP terminal SHALL tighten — suspending sessions whose criticality requires fresher presence — and SHALL NOT loosen (downgrade-monotonicity).
R4.3 WHEN the CAP terminal receives a Faying G0 Emergency Override for a Fay, it SHALL immediately terminate all of that Fay's active Sessions on the terminal, refuse new AuthRequests from that Fay until re-bootstrapped, and route control to human (E_EMERGENCY_OVERRIDE_ACTIVE).
R4.4 The CAP terminal SHALL NOT allow a Session's effective grade or granted modes to be elevated within its lifecycle; a higher grade/mode requires a new AuthRequest under a fresh Mandate (no runtime escalation interface).
R4.5 WHEN the bound Mandate transitions to Revoked (any Faying RevokeReason), the CAP terminal SHALL terminate all Sessions referencing it (E_MANDATE_REVOKED).

R5 — Destructive / irreversible operation confirmation

Goal D5. The most dangerous gate. Irreversible or safety/life-critical operations require the strongest evidence of a present, informed, accountable human.

R5.1 The CAP terminal SHALL classify every operation by reversibility ∈ {reversible, hard-to-reverse, irreversible} and physical_risk ∈ {none, property, safety, life-critical}, yielding criticality ∈ {routine, sensitive, high, critical}.
R5.2 WHEN an operation's criticality = critical (irreversible, or life-critical), the CAP terminal SHALL require all of: (a) bound Mandate grade ≥ G1 Strict; (b) a presence proof fresh within the G1 window (≤ 5 s); (c) an explicit, per-operation human confirmation obtained before execution; otherwise reject (E_HUMAN_CONFIRMATION_REQUIRED / E_GRADE_INSUFFICIENT / E_PRESENCE_STALE).
R5.3 WHEN an operation's criticality = high (hard-to-reverse, or safety-risk), the CAP terminal SHALL require grade ≥ G2 and a fresh presence proof for the grade, and SHALL require explicit human confirmation unless the bound grade is G1 with a current presence proof. [需人类决策 D-6: high-class confirmation policy]
R5.4 Before executing any high or critical operation, the CAP terminal SHALL present a structured disclosure containing at least {operation, resource_id, reversibility, physical_risk, affected_scope, terminal_attested_state}; free-text-only confirmation SHALL be rejected.
R5.5 The CAP terminal SHALL bind one human confirmation to exactly one atomic operation; IF a single confirmation is presented for a bundle containing a high/critical operation, THEN the CAP terminal SHALL reject it (E_OPAQUE_ACTION_BUNDLE).
R5.6 The CAP terminal SHALL NOT rely on Fay-supplied state for the disclosure of high/critical operations; it SHALL use terminal/driver-attested state (anti-CT1).
R5.7 The CAP terminal SHALL write every high/critical confirmation decision (granted/denied/timeout) to the audit log and SHALL surface it on the Human-Prime view channel.
R5.8 The criticality gate SHALL be history-independent: prior benign behavior SHALL NOT lower the requirements of R5.2–R5.5 (anti-CT6).
R5.9 IF human confirmation is not received within the confirmation timeout, THEN the CAP terminal SHALL default to denial (E_HUMAN_CONFIRMATION_TIMEOUT) and SHALL NOT execute the operation.

R6 — Atomic control-authority handover

Goal D6. At any instant a resource has at most one active controller.

R6.1 The CAP terminal SHALL execute handover so that external observers never see two Sessions active on one Resource_ID simultaneously.
R6.2 IF any pre-release step fails, THEN the CAP terminal SHALL roll the source Session back to active; IF a post-release step fails, THEN it SHALL leave the resource idle and SHALL NOT auto-restore the source Session.
R6.3 WHEN a handover targets a human (Fay-to-Human), the CAP terminal SHALL reuse the §8 confirmation gate where the handed-over control includes high/critical operations.
R6.4 WHILE a resource is handover_pending, the CAP terminal SHALL serialize/reject concurrent handover requests for that resource.

R7 — Liveness and zombie-session reclamation

Goal D7.

R7.1 The CAP terminal SHALL determine a Session failed only when both the persistent connection is broken beyond the heartbeat-timeout threshold and now − last_heartbeat_at exceeds it (dual determination).
R7.2 WHEN a Session is determined failed, the CAP terminal SHALL terminate it, revoke OS access, and release the resource.
R7.3 WHEN the bound Mandate's presence becomes stale such that an active Session's criticality is no longer permitted (R4.2), the CAP terminal SHALL suspend or terminate that Session.
R7.4 A terminated Session SHALL NOT be auto-restored; a new AuthRequest is required.

R8 — Observable, auditable, revocable

Goal D8.

R8.1 The CAP terminal SHALL emit SessionStateChanged for every observable state transition to the owning iFay_Runtime.
R8.2 The CAP terminal SHALL record, for every grant/deny/handover/confirmation/revocation, an audit entry sufficient to reconstruct the R1.1 accountability tuple and the decision reason code.
R8.3 WHEN a revocation statement reaches the terminal, the CAP terminal SHALL reject subsequent validations of the revoked credential immediately and terminate referencing Sessions.
R8.4 Revocation reachability SHALL be ≥ establishment reachability (a path that could grant control must have an at-least-as-reachable path to revoke it).

R9 — Interoperability (wire format)

Goal D9.

R9.1 The CAP terminal SHALL accept the protocol's CBOR primary encoding and the JSON fallback as semantically equivalent for every message and credential.
R9.2 The CAP terminal SHALL compute signature inputs over RFC 8949 deterministic CBOR (offline credentials) / RFC 7515 JWS (online tickets), so independent implementations verify identically.
R9.3 The protocol SHALL be defined by an authoritative CDDL schema (schemas/cap-1.0.cddl); WHEN CDDL and prose conflict, CDDL prevails.
R9.4 WHEN negotiating versions, implementations SHALL select the highest commonly-supported CAP/MAJOR.MINOR; cross-MAJOR SHALL be rejected (no best-effort interop).

R10 — STRIDE security properties

Goal D10. (Spoofing/Tampering/Repudiation/Information-Disclosure/DoS/Elevation — full criteria authored with §12 next wave.)

R10.1 The CAP terminal SHALL reject requests whose descriptor signature, Mandate binding, or channel binding fails verification (anti-Spoofing/Tampering).
R10.2 IF an AuthRequest's operation or mode differs from what a prior confirmation was bound to, THEN the CAP terminal SHALL reject execution (anti-Tampering/Elevation).
R10.3 Error responses SHALL NOT expose key fingerprints, internal IDs, or resource topology (anti-Information-Disclosure).
R10.4 The CAP terminal SHALL enforce session/connection/storage quotas and rate limits (anti-DoS).

R11 — Cryptographic evolvability

(crypto goal.)

R11.1 The CAP terminal SHALL support the mandatory suite (ed25519, ecdsa-p256-sha256).
R11.2 The protocol SHALL reserve algorithm-suite negotiation bits for post-quantum algorithms (e.g., ML-DSA-65) without a wire-breaking change (parity with Faying G8).
R11.3 WHEN a new default suite is introduced via negotiation, existing valid credentials SHALL remain verifiable until their not_after.

R12 — AI-era adversarial threat resistance (CAP-specific)

Goal D10. Anchors the CT1–CT7 catalog (outline §12.3). Full per-threat criteria authored with §12 next wave; the clauses below are the protocol-level commitments the drafted chapters already rely on.

R12.1 (CT1 Perception Spoofing) The CAP terminal SHALL base high/critical disclosures on terminal/driver-attested state, not Fay-supplied state.
R12.2 (CT2 Opaque Bundle) The CAP terminal SHALL reject bundling of high/critical operations under one confirmation (see R5.5).
R12.3 (CT3 Simulated-Operation Confusion) WHEN a high/critical operation is requested via a GUI-grounding/execute path, the CAP terminal SHALL apply R5.2/R5.3 with no exemption for "automation convenience".
R12.4 (CT4 Prompt Injection) The CAP terminal SHALL treat Fay-originated criticality claims as untrusted; terminal/driver-declared criticality prevails; unknown → critical (R2.4).
R12.5 (CT5 Presence-Replay / Attention-Drift) The CAP terminal SHALL require per-grade fresh presence and SHALL hold high/critical operations arriving during an attention-drift window until re-confirmation.
R12.6 (CT6 Cultivated Trust) See R5.8 (history-independence).
R12.7 (CT7 Confirmation/Audit Flooding) The CAP terminal SHALL priority-bucket the confirmation/audit channel and SHALL NOT drop critical-bucket events under rate limiting.

R13 — Resource access model (read-write lock)

(CAP-specific.)

R13.1 The CAP terminal SHALL enforce the read-write-lock matrix: read is shareable; write/execute/configure are exclusive and mutually incompatible with any other occupancy.
R13.2 The CAP terminal SHALL NOT auto-queue an incompatible request; it SHALL return E_RESOURCE_BUSY (handover is the explicit path).
R13.3 The CAP terminal SHALL treat access modes as independent (no hierarchical inclusion); each mode SHALL be independently authorized.
R13.4 The CAP terminal SHALL confine a Session to the granted_modes derived from the matched Grant, which SHALL be a subset of the bound Mandate's scope. [需人类决策 D-1]

Changelog

(draft) Initial EARS baseline for the Faying-level uplift, H1-PROTO-01.