AQ-TSM-000 — Trust State Model Specification v0
DOCUMENT STATUS
Specifies the behaviour of the Dynamic Trust State layer L4 (AQ-TSE-01) inside AQ-NODE-01. Version 0 defines structure and rules only. Every threshold in it is a declared assumption with no measured value no measured value. Evidence level of the model: Theory. Nothing specified here has been implemented or exercised, and the model has never been run against a real observation stream.
1. CONFIDENCE IS NOT TRUST
The single most consequential distinction in this specification is the one most often collapsed in practice.
- Confidence is per domain, and is a statement about how well a property of the world is currently known. It is produced from observations supplied by L1, L2, L3, L6 and L7. Confidence may rise as soon as evidence supports it, and it authorizes nothing.
- Trust is system-wide, discrete, and a statement about what the node is permitted to do. It is a function of the confidence vector and of standing policy. It falls on a single qualifying observation, and it rises only through the recovery sequence in §5 (DC-4).
Consequences that follow directly, and that an implementation must not soften: a domain returning to established does not restore a trust state; a decision function at L5 reads the trust state and never a raw confidence value; and no combination of confidence values can produce a trust state above the one the recovery sequence has restored. Confidence is evidence about the world. Trust is a posture the system holds toward its own authority.
2. CONFIDENCE DOMAINS AND THEIR REPRESENTATION
Version 0 uses six domains. The set is a declared assumption in its own right (TA-D, §6) — there is no evidence yet that it is complete or that its members are independent.
| Domain | Property it is about | Primary supplier | What degrades it |
|---|---|---|---|
| D-ENT | Entropy source behaviour at the terminal | L1 | Health-test failure; a source that stops reporting; a sample whose provenance tag is missing |
| D-TIME | Integrity of the time base evidence is ordered by | L2 | Divergence beyond TH-TIME-DIV; loss of an external reference; an ordering inversion in the record |
| D-IDENT | Binding of identities to key material | L3 | Any verification failure; a compromise indication; revocation data that cannot be refreshed |
| D-CHAN | Behaviour of the physical channel at the current rung | L0 via L1 | Channel disturbance beyond TH-CHAN-DIST, whatever its cause — attribution is not required to degrade |
| D-PLAT | Integrity of the host platform carrying the node | L1, L9 | Unexpected reset; power excursion; an unattested configuration change |
| D-EVID | Continuity and integrity of the Evidence Fabric | L6 | A gap beyond TH-EVID-GAP; a failed integrity check; a write that cannot be made durable |
Confidence in v0 is an ordinal class, not a number and not a probability. A scalar would imply a calibration that no measurement supports; publishing one would violate DC-1 by presenting a MODELLED quantity in the form of a MEASURED one. The four classes are ordered: established, provisional, degraded, undetermined. undetermined is not a mild case of degraded; it is the explicit representation of DC-2 and, for transition purposes, it is treated no more favourably than degraded.
Figure 1 — The per-domain confidence record held by AQ-TSE-01
domain-confidence-record
domain : D-ENT | D-TIME | D-IDENT | D-CHAN | D-PLAT | D-EVID
class : established | provisional | degraded | undetermined
provenance : MEASURED | DERIVED | MODELLED (DC-1)
observed-at : [time-base id] + [uncertainty statement] (from L2)
age-state : fresh | stale | expired (TH-STALE)
basis : evidence-fabric reference(s) (DC-5)
advisory : L8 annotations, non-authorizing;
MUST NOT change class, order or age-state
Rules: no field may be defaulted. A record that cannot be built is
published with class = undetermined and a reason, never omitted.
An expired age-state forces class = undetermined regardless of the
last observed value.
3. STATE SET
| State | Meaning | Entry condition |
|---|---|---|
| NORMAL | Full action set, bounded only by standing policy | All domains established and fresh; commissioning complete; or restoration from RECOVERING via a second, separately authorised step |
| RESTRICTED | Reduced effect set; established sessions may continue | One qualifying observation in any domain (D1) |
| CRITICAL | No new effects; the node holds and reports | A second domain not established, or either integrity domain (D-IDENT, D-EVID) not established (D2) |
| DENY | No execution of any kind; terminal until validated | An integrity domain failed, a verification failure at L3, or a declaration by a named authority at L9 (D3) |
| RECOVERING | Observation only; the node is being re-qualified | Entry to the recovery sequence at R5; never entered automatically |
Figure 2 — Trust state transitions. Degradation is monotonic; every upward edge passes through RECOVERING (DC-4)
entry (commissioning): all domains established, policy loaded
|
v
+----------------------------------------------------------+
| NORMAL full action set, bounded by standing policy |
+----------------------------------------------------------+
| D1 one qualifying observation, any domain
v
+----------------------------------------------------------+
| RESTRICTED execute-restricted only; effects narrowed |
+----------------------------------------------------------+
| D2 second domain degraded, or an integrity
| domain (D-IDENT / D-EVID) not established
v
+----------------------------------------------------------+
| CRITICAL safe-hold and abstain only; no new effects |
+----------------------------------------------------------+
| D3 integrity domain failed, verification failure,
| or declared by a named authority at L9
v
+----------------------------------------------------------+
| DENY no execution; terminal until validation |
+----------------------------------------------------------+
|
| recovery sequence R1..R7 must complete in order
v
+----------------------------------------------------------+
| RECOVERING observation only; TB-4 gate stays de-armed |
+----------------------------------------------------------+
|
| R7 controlled restoration, one step upward only,
| authorised and recorded at L9
v
RESTRICTED (never directly to NORMAL from DENY)
Downward transitions are immediate and need one observation.
There is no upward edge that does not pass through RECOVERING.
4. PERMITTED ACTIONS PER STATE
The table is the authority for what L5 may emit. Q1-Q3 must already have passed; the trust state can only remove what they allowed.
| State | execute | execute-restricted | safe-hold | abstain | deny | Key material | TB-4 gate |
|---|---|---|---|---|---|---|---|
| NORMAL | permitted | permitted | permitted | permitted | permitted | Full use per policy, including new key agreement | May arm |
| RESTRICTED | not permitted | permitted, within the narrowed effect set | permitted | permitted | permitted | Established sessions only; no new key agreement | May arm for the narrowed set only |
| CRITICAL | not permitted | not permitted | permitted | permitted | permitted | Verification only | Not armed |
| DENY | not permitted | not permitted | permitted (hold and report only) | permitted | permitted | No use | Not armed |
| RECOVERING | not permitted | not permitted | permitted | permitted | permitted | Verification only, for re-qualification | Not armed |
ABSTAIN IS AN ANSWER
abstain means "insufficient confidence to authorize" (DC-3). It is a completed decision, is recorded as one, and must never be implemented as a timeout, a retryable error, or a condition that a caller can clear by repeating the command. A retry loop around abstain converts a permitted output into a denial-of-service against the node's own operator.
5. THE MANDATORY RECOVERY SEQUENCE
No step may be skipped or reordered. A new qualifying observation at any step returns the node to R3 and restarts the sequence.
| Step | Requirement | Evidence artefact (L6) | Threshold |
|---|---|---|---|
| R1 anomaly | The condition exists. By definition it is not yet observable to the node. The interval R1→R2 is detection latency and is currently undefined. | None at the time; reconstructed at R6 | — |
| R2 detection | An observation qualifies against a declared threshold, or a required observation is absent past its age limit | The raw observation with provenance and time | TH-ENT-QUAL, TH-TIME-DIV, TH-CHAN-DIST, TH-EVID-GAP, TH-STALE |
| R3 state transition | Immediate, on the single observation, with no averaging or grace period (DC-4) | Transition record naming the causing observation | TH-CONCUR |
| R4 restriction | L5 narrows the action set and the TB-4 gate is de-armed before the next command is evaluated | Restriction record and gate-state record | — |
| R5 recovery observation | The affected domain is observed over a dwell window with independent confirmations. Absence of the anomaly is not evidence of recovery (DC-2). | The observation series, in full, not a summary | TH-REC-DWELL, TH-REC-COUNT |
| R6 validation | A named authority at L9 confirms the cause is characterised, not merely no longer visible, and files it against the F-1..F-14 register | Validation record carrying the cause statement and register entry | — |
| R7 controlled restoration | One step upward only, with the authorising identity and the evidence set relied on both named. DENY restores to RESTRICTED, never to NORMAL. | Restoration record; reconstructable in isolation (DC-5) | — |
6. THRESHOLDS AS DECLARED ASSUMPTIONS
No value below has been measured. Each row states the parameter, the assumption that justifies having such a parameter at all, and the method by which a value will later be established. Until the method has been executed, the parameter is undefined and any implementation must refuse to start rather than substitute a placeholder.
| ID | Parameter | Gates | Declared assumption and rationale | How a value will be established |
|---|---|---|---|---|
| TA-D | The six-domain set | The whole model | Assumes these six domains span what must be known before authorizing. Rationale: each maps to a distinct supplier layer and a distinct failure surface. | Bench fault-injection: any injected fault that no domain represents falsifies the set |
| TH-ENT-QUAL | Entropy health qualification boundary | D-ENT class | Assumes source misbehaviour is detectable at the terminal before dependent key material is used. Rationale: L3 depends on L1 at generation time. | Characterisation runs on the bench source, including deliberately degraded sources |
| TH-TIME-DIV | Divergence bound between local time base and reference | D-TIME class | Assumes a divergence exists that is distinguishable from normal hold-over. Rationale: without it, spoofing and drift are indistinguishable. | Hold-over characterisation, then reference-manipulation trials on the bench |
| TH-STALE | Observation age at which a domain becomes undetermined | Every domain | Assumes a stale observation stops being evidence at some age. Rationale: DC-2 — a value that is merely old must not read as benign. | Derived per domain from the measured natural update interval of its supplier |
| TH-CHAN-DIST | Channel disturbance qualification | D-CHAN class | Assumes a disturbance metric exists whose definition survives the change of rung. Rationale: the contract must not change between bench and orbit. | Defined from L0/L1 bench measurement, then re-validated at each rung before use |
| TH-IDENT-FAIL | Verification-failure count for domain degradation | D-IDENT class | Assumes counting is meaningful at all. Rationale: it is not a tolerance — a single failure is already a hard-constraint failure at L5, so this parameter only governs the domain class, never authorization. | Set from observed benign failure rates on the bench link; provisionally 1 |
| TH-EVID-GAP | Tolerated continuity gap in the Evidence Fabric | D-EVID class | Assumes some gap is attributable to benign scheduling. Rationale: if it is not, the parameter collapses to zero, which is an acceptable outcome. | Measured write-latency distribution under bench load |
| TH-CONCUR | Count of non-established domains forcing CRITICAL | D2 transition | Assumes concurrent degradation is more significant than the sum of its parts. Rationale: correlated causes are likelier than coincidence. | Fault-injection with deliberately correlated faults (shared clock, shared power) |
| TH-REC-DWELL | Dwell window length at R5 | Recovery | Assumes an anomaly that has truly cleared stays cleared for a characteristic time. Rationale: intermittent faults must not be recovered through. | Measured recurrence intervals of injected intermittent faults |
| TH-REC-COUNT | Independent confirming observations at R5 | Recovery | Assumes independent confirmations exist. Rationale: repeated reads of one stuck sensor are one observation, not several. | Established only after independence between sources has itself been measured |
7. WHAT THIS MODEL DOES NOT YET JUSTIFY
STATED LIMITS OF v0
Each item below is a gap in justification, not a to-do. None may be closed by argument; each needs an artefact.
- No calibration. The four ordinal classes are a convention. There is no measurement that says
provisionalanddegradedare separable in any observation stream. - No completeness argument for the domain set, and no evidence that the six domains are independent. Common-mode failure — one clock, one power rail, one host — is unmodelled, and TH-CONCUR is a guess about a dependence structure nobody has measured.
- No error characterisation. False-positive and false-negative behaviour of every transition rule is unknown, so no claim about proportionate response can be made, in either direction.
- Monotonic degradation is a policy choice, not a proven optimum. DC-4 buys reconstructability at an availability cost that has never been quantified. A rule that reaches DENY on a single false observation is functioning as specified, and the operational consequence of that is unmeasured.
- No detection-latency bound. The R1→R2 interval is undefined, so the model says nothing about how long a node acts on stale trust after an anomaly begins.
- The adversary model is untested. Transition rules assume observations are trustworthy until flagged. An adversary who suppresses observations is handled only by TH-STALE, and TH-STALE has no value.
- RECOVERING under continued anomaly is underspecified beyond the prohibition on upward transition.
- Recoverability is not established at every rung. R6 needs a named authority and R7 needs an authorised action; on a flight platform with intermittent contact, neither may be executable within a useful interval. Whether DENY is recoverable above rung 3 is an open question that must be answered before rung 4, not during it.
- The L8 exclusion is asserted structurally and not yet tested. The check that no advisory annotation can influence a transition needs to exist as an L9 test artefact.