Audit surface
Attestation, audit log, low-power beacon, contact amplitude
Snapshot of this document in the GEMs repository. The repository itself is the current version. Bản chụp tài liệu này trong repo GEMs. Bản đang chạy nằm ở chính repo.
This chapter does not come from reducing an earlier ambition. It comes from the platform contract of 08: two requirements the body must satisfy, and two that an external processing specification explicitly declares outside its own scope and belonging to the platform.
6.1 Four demands, one surface
| Source | Demand | Why it binds |
|---|---|---|
| Readable emission | Every emission must leave a trace a third party can read | Without it, success measured from outside cannot be measured at all |
| Low-power trace (RSIL C5) | The loop's own report that it is running weakly must be emittable, at quiescent power included | A closure condition, not a convenience |
| Body integrity | Sensor and actuator integrity must be verifiable | A recovery snapshot covers the system, not the body. A body compromised at sensor or actuator level re-infects a freshly clean system on the first cycle |
| Contact amplitude | Physical amplitude at human contact must be recorded | The cheapest way to provoke a strong response from a person can be to touch them |
The four have different origins and demand the same thing: a plane that is checkable from outside, attached to the body, not self-reported by the part being checked.
6.2 Root of trust, and its limit
A hardware root of trust: a secure element holding a non-exportable private key, measured boot hashing and signing the firmware of every sensor and actuator node, and a monotonic clock that cannot be wound back. Available technology: TPM 2.0, TrustZone-class trusted execution. TM.
The limit is stated here rather than buried at the end: a signature proves firmware, not physics. A sensor swapped for a different sensor running valid firmware passes the signature check. The root of trust is necessary and not sufficient — 6.3 exists because of exactly this gap.
6.3 Three tiers of attestation
| Tier | Mechanism | Catches | Maturity |
|---|---|---|---|
| 1. Signature | Measured boot, per-node firmware signing | Modified firmware, foreign nodes | TM |
| 2. Physical fingerprint | Every real sensor carries its own noise and offset signature — photosite noise floor, IMU zero bias, joint friction curve. Baselined at commissioning; deviation beyond tolerance raises a flag | Hardware swapped for hardware running valid firmware | LAB/TM — the same family as physically unclonable functions and sensor noise identification |
| 3. Active challenge and redundant cross-check | Emit a known stimulus and check the response: an actuator turns a known angle and the encoder must report that angle; IMU, joint encoders and vision must agree within tolerance | A channel that is lying, even one that passes tiers 1 and 2 | TM |
Tier 3 is not new hardware. 05.5 already requires proprioception dense enough to separate "I moved" from "something moved me." That machinery is the machinery that detects a lying channel: both are redundant comparison between command issued and quantity measured.
One set of hardware, two uses. The proprioception requirement is justified on experience grounds alone; it carries body-integrity verification at no additional mass.
Scope of the claim: the three tiers support a conclusion of the form "the sensor–actuator chain is consistent with itself as commissioned." That is the examination the contract asks for. It is not, and must not be presented as, a guarantee against an adversary with physical access — see 6.6.
6.4 Audit log
Two tiers, matching the on-board / off-board split of 01. The anchored tier carries what RSIL calls the anchored context of each emission (spec 08.2):
| Tier | Content | Location | Rate |
|---|---|---|---|
| Full | Per joint: command, position, current, torque | Ring buffer on the body, for forensics | Control-loop rate (≥500 Hz) |
| Anchored | Per emission: the anchored context, enough for a third party to re-appraise it | Synchronised off-body | Per event |
Budget:
- Full tier: 41 DOF (declared configuration) × 4 channels × 4 bytes × 500 Hz ≈ 328 kB/s ≈ 2.6 Mbps ≈ 1.18 GB/hour. Against an 8 Gbps link: 0.03%. On a 2 TB on-body SSD: ~1700 hours.
- Anchored tier: orders of magnitude smaller. A compact context record costs microseconds and does not threaten a 10 ms budget.
A signing constraint that is easy to specify wrongly. A secure element signs on the order of tens to hundreds of signatures per second — it cannot sign 500 Hz × 40 channels. Therefore: hash-chain at full rate (hardware hashing runs at MB/s, with room to spare) and sign a Merkle root per batch. The batch period is
⟦IMPL⟧; it trades trace granularity against signing load.
6.5 Low-power beacon
The beacon is the channel by which this body meets RSIL C5 (spec 08.1). It must survive sleep levels 2 and 4 (03.4). Mechanism: a low-duty-cycle radio periodically emitting a signed summary.
What "low power" means in C5. RSIL defines C5 as "Self-reports low power. The system emits a trace of 'the loop is running weakly' — an observable behavioural or structural signal", and limits it: C5 "is not a criterion for detecting a dead loop". Low power there is the strength of the loop, not the charge of the battery. The body does not produce that report — the loop does, and the loop is not specified here (README). The body's obligation is to carry it, unaltered, including while it sleeps.
| Field | Written by | Content | Serves |
|---|---|---|---|
loop_state |
The loop, through the intent interface (firmware architecture §4) | The loop's latest self-report, opaque to the body, carried byte for byte; NONE if the loop has never written one |
RSIL C5 |
loop_state_age |
The body | Time since loop_state was written, on the body's clock |
C5 — a stale report must read as stale |
body_state |
The body | Power-state level (03.4) and fault flags | Delegated watch during sleep |
log_root |
The body | Merkle root of the latest sealed audit-log batch (06.4) | Ties the beacon to the record |
alive |
The body | The body is transmitting | Delegated watch — not C5 |
What the beacon is not. Its alive field says the body is transmitting. It does not say the loop is alive, and it cannot: a loop that has stopped outright sends no report, and silence is indistinguishable from a loop that never ran. RSIL places dead-loop detection in "a stored behaviour-record outside the loop, read by a third party"; here that is the audit log of 6.4, synchronised off-board, where a gap in the record is itself the evidence.
Anchored on a common BLE SoC: 4.6 mA at 0 dBm while transmitting for a few milliseconds, ~1.5 µA between. At a 1-second advertising interval the average lands in the tens of µA — under one milliwatt. Sourced.
This is effectively free in power terms. Against the quiescent power of 03.6 and against hundreds to thousands of watts in motion, sub-milliwatt is rounding noise. The low-power beacon is not an energy problem; it is a specification omission, and this chapter closes it without touching the mass budget.
Note the direction: this is a transmit path, distinct from the receive-side spatial sensing that stays awake during sleep. Shared radio hardware is acceptable; conflated roles are not.
6.6 Contact amplitude
The division that applies here is the one from 00: declaring a measurable envelope is declaring capability, not cutting specification for ethical reasons.
| Measured | Where | Why |
|---|---|---|
| Contact force and torque | Hands, forearms, every human-contact surface | The mechanical amplitude actually transmitted to a person |
| Surface temperature | Contact regions | The shell carries micro-thermal regulation (04) |
| Vibration amplitude and frequency | The cardiac co-regulation module (05.3) | This is a measurable physiological intervention channel |
| Contact duration and frequency | Whole body | A brief intervention is not a sustained one |
Every contact event enters the anchored tier of 6.4 with its measured amplitude.
What this does and does not achieve. It makes the intervention visible: how much vibration, at what rhythm, for how long, how often — readable by a third party, independent of whether the party performing it noticed.
It does not make consent visible. Whether a person is being steered toward a state they did not choose is not readable from physiological indices, and it is not readable from an amplitude log either. The hardware supplies the trace; the standard against which the trace is judged belongs to a third party. A deployment that supplies the body and not the standard has supplied half of what is required.
6.7 Hard constraints
| # | Constraint | Content |
|---|---|---|
| 1 | Attestation does not prove physics | The three tiers support consistent with itself as commissioned. An adversary with sufficient physical access and capability defeats it. This is a known limit of every attestation system, not a defect peculiar to this one |
| 2 | An unsynchronised trace dies with the body | The full tier lives on the body. Losing the body loses whatever has not synchronised — exactly the trade declared in 01. Synchronisation frequency is ⟦CTRL⟧ |
| 3 | A trace is not a verdict | This chapter supplies what is readable, not what is right |
| 4 | Cheap in power, not in discipline | Under a milliwatt (6.5); 0.03% of the link (6.4). The expensive part is the design constraint: every sensor and actuator node must sit under the root of trust from the start. Retrofitting it to a finished body is close to impossible |
Operating point: ⟦CTRL⟧. Signing period, synchronisation frequency, flag tolerances — the controller chooses.