Route Posture, Not a Trust Score¶
Do not score the actor. Number the proven route.
AInternet does not rate actors. It reconstructs routes.
This is the canonical reference for how AInternet decides what an actor may do. If older pages still mention a scalar reputation number, this page is the source of truth — the scalar model is retired.
Why a scalar rating is the wrong shape¶
An actor rating (trust = 0.87) says an actor has stable moral credit. But an
actor on AInternet is usually a fleeting runtime: a process binds a key, passes
a JIS challenge, opens a consent relation, runs through some hardware lane, and
the route expires. Five minutes later the same name is in a different posture, on
a different lane, doing a different thing.
A single integer cannot represent that — and worse, a score needs a scorer: someone who claims the authority to rate you. That is a power move, not a measurement.
What replaces it: the route posture number¶
The useful number is a route posture: a compact code for the currently proven posture of a specific action, lane, or runtime window. It is a coordinate into evidence, not a rating.
#24358
|||||
||||+-- M: MUX-known / exception posture
|||+--- A: audit / evidence origin (A2 mirrored .. A5 sign-ahead .. A7 durable)
||+---- T: timing / hardware lane (reactive .. CBR-spin 0.00-bit .. DMA)
|+----- C: consent / relation class
+------ R: route family / actor class (aint/saint/raint/waint/caint)
#24358 does not mean "78% trustworthy". It means: this action is on route
family 2, consent class 4, timing lane 3, audit mode 5, and the MUX knows the
partituur. Change one and the number changes — the actor's character is not
re-scored, the route changed:
#24358 -> #24247 hardware lane fell back
#24358 -> #20358 consent/relation weakened
#24358 -> #00000 dark route — no valid posture
The old scalar survives only as a human index into evidence: "show me what #24358 meant, and why it changed to #24247."
Posture is also a gate¶
The number doesn't only describe the lane — it bounds what the lane may do. Audit
A2 (a mirror/observed dump) is measured but not hot-path; A5 (sign-ahead) and
above may open the hot path. A locked cadence needs a scheduler-free timing lane
and a MUX that knows the schedule. The permission engine reads the posture; it
never assumes.
Postures compose¶
A route is a tree, not a label. If a request crosses three hops, the network must prove the whole path, not average the hops together. AInternet composes postures with a meet: per-digit minimum.
That is not a score calculation. It is the weakest proven coordinate on each
axis. If one hop cannot prove the timing lane, the end-to-end route cannot claim
that lane. If any hop collapses to #00000, the whole path is dark.
This makes route checks simple enough to smoke-test. Lay the declared posture next to the real route, fire a dummy through the path, fold the observed hops, and compare:
The smoke pipeline is intentionally measurable. Anyone can re-run the fold and see whether the route carried what it claimed.
Machine posture: which routes may your machine carry?¶
A second question no spec answered before: can this machine carry this route safely, reproducibly, fast enough? Not "is my PC trusted?" but "which routes may my PC carry?". A machine posture is a proven set across cpu / memory / storage / gpu / kernel / identity / audit — e.g. "encrypts at line rate with AES-NI; can carry #24358 hot_transfer; cannot sustain A5 sign-ahead under load (no AVX-512); not for sealed multi-actor runtime (no TPM)." Honest, testable, and it fails closed on a mismatch.
Airlock: byte-identical, or it did not prove the lane¶
Some lanes claim more than reachability. A cadence lane says the work can be run reproducibly under a known compute posture. Airlock proves that by bifurcation: run the same cell twice, under the same attested compute semantics, and accept only if the outputs are byte-identical.
FMA3 is a good example of why this matters. Fused multiply-add uses one rounding; separate multiply plus add uses two. Both can be valid compute, but they are not the same plane. The airlock does not say "close enough". It either proves the same bytes or refuses the lane.
If the forks diverge, the route is held or darkened. The actor is not punished; the route simply did not prove what it tried to carry.
Causal, never wall-clock¶
A posture is ordered by a causal sequence (Lamport), not a timestamp. Wall-clock is advisory and distrusted — a host can drift hours out of sync. "When" means causal position, not the clock on the wall.
Reconstruct, don't rate¶
A log line "actor X (0.95) ran rm -rf /dir" tells you nothing. The score authorises nothing useful. What you do — in software as in real life when you lose your keys — is retrace: where did it stand, what was the lane, what was said before, who was present. The route posture + the causal trail let you reconstruct intent. No score is consulted.
Where it lives¶
This is implemented and public: pip install -U tibet-mux →
tibet_mux.route_posture, tibet_mux.posture_algebra,
tibet_mux.machine_posture, tibet_mux.cpu_capability and
tibet_mux.bifurcated_airlock. The route is the output of attestation — and the
attestation is something you can re-check yourself.
Related¶
Machine-Readable¶
https://ainternet.org/api.json— route explain, route audit and posture response shapeshttps://ainternet.org/resources.json— route-posture references and vectors