The cache itself attests when AI behavior drifts from its commitments —
measured per commit, on the deployer's own laptop, in nanoseconds, against a
conservative 3σ threshold with a binary renewal trigger.
why ShortLex · the moat IS the algorithmC4
ShortLex is the
sole known addressing scheme where the
hierarchical semantic coordinate IS the physical memory address at every
scale (patent v20 §16). Five properties simultaneously required and
uniquely satisfied:
positional semantics · scale invariance
· nested positional equivalence · compositional nesting
· sparse O(1) allocation. Hash tables destroy spatial locality;
B-trees order by numeric key, not semantic weight; vector databases use
flat embeddings; HNSW's graph position isn't the memory address. ShortLex
is the structural-impossibility-result moat — any competitor
preserving Codd's logical/physical separation cannot achieve it.
scale invariance · map of maps · one user-facing top-level setB6 · FRAC-1
The 12×12 lattice below is
one top-level user-facing set
— endlessly subdivided (each cell expands recursively into its own
12-axis sub-lattice;
A2.B3.C1.… addresses are stable at
every depth) and reconnected across the stack:
local pmu-onchip walks
the cells · SimHash distance fills them · the cloud bridge
superimposes them across deployers. Demoed at N=12, 32, 64; the chip
doesn't notice N. Compliance-taxonomy depth is a display choice, not an
architectural commitment.
The 12 things that drift — ShortLex ladder
Length-1 axes first (the coarse layers),
length-2 next (the fine kinds). Each names a time-themed drift
between intent and reality. Recursively expandable to length-3+.
🏛️A · Strategy
top-of-plan intent
⚡B · Tactics
execution layer
🔧C · Operations
flow layer
⚖️A1 · Law
non-negotiable commits
🎯A2 · Goal
named target
💰A3 · Fund
resource envelope
🎿️B1 · Speed
cadence of action
🤝B2 · Deal
unit of exchange
📡B3 · Signal
observable indicators
🔌C1 · Grid
fixed scaffolding
🔄C2 · Loop
repeating cycles
🌊C3 · Flow
continuous throughput
Pricing — insurance-shaped, three componentsF2
- Base rate · flat per-deployer monthly
access ($250–$500 / mo). Covers daemon support, cloud
bridge, σ-baseline maintenance. ~95% margin on the
flat — pure platform fee.
- Per inference · $0.005 / receipt at
tier-2 VOLUME, $0.002 at tier-3 ENTERPRISE >1M/mo. Marginal
compute is ~0.011¢F1,
so the per-receipt fee is ~99% margin against compute alone.
- Premium · 25 bps (0.25%) of dollar
value of decisions or transactions flowing through PMU-attested
agents — derived as 10–15% of underlying primary at typical
cyber/AI rates (Marsh / Coalition / Armilla benchmarks). Scales
with underwritten exposure, not receipt volume. 75% margin
(25% = 12% IFRS-17 actuarial reserve + 8% case-study co-authoring
time + 5% dollar-level cloud scaling)F3.
- Funding ask: $1.5M to first paid conversionF5/F6
+ M-D research (direct PMU counters for sub-3σ resolution)M-D1.
The pilot · deployer + underwriter commitments
- 6 months · 3 deployers · all receipts
free (tier-1 SEED).
- Deployer: run daemon, opt-in cloud-send, monthly
drift review, honest reporting on flagged zones.
- Underwriter: read stream weekly, name drift
threshold by month 2, co-author case study by month 4.
- No source code, no equity, no exclusivity.
Daemon measures cache state, not agent payloads.
Renewal trigger — binary, named at start
Pilot converts to paid tier-2/tier-3 when the first claim
against a flagged-zone deployer is paid by the underwriter.
No claim during pilot → CONFIDENCE EVENT (six months of clean
attestation is itself risk reduction); terms re-open.
The chip in numbers (this laptop, today)
- 0.54 ns / XOR-gateA1
· 155 ns / full 12×12 walkA3
(one DRAM access). M-series, arm64.
- CV < 0.3% on L1 / gate / walk across N=20
stability runsA5
— the signal IS the silicon.
- 3.4σ above time-local baseline distinguishes
two software-identical agent actions (
read
10-byte vs 2.7 MB JSON)D8.
3/3 reproducible · 0/5 negative controls flagged.
- N×N scale-invariant. 12×12 today;
32×32 and 64×64 also rendered; same codeB6.
⚡ the moment in the room
Two agent actions a software classifier calls identical
(
read .json); cache fingerprints separate by 3.4σ
above this laptop's noise floor measured 90 seconds ago
D8.
The chip notices what software cannot.
This is NOT search.
It is
O(1) reach IS verify — the
same fetch the agent had to do to act
IS the verification. No
second pass. No traversal. No similarity score. No model. The receipt is
one cache line read at one coordinate, and the property is
scale-invariant from 12×12
(the human-readable compression) to N×N at any N — same one-
cycle gate.
For the underwriter: priceable
per-inference because there is no per-claim search cost to amortize.
For the CTO: zero verification overhead at the
revenue path. For the AI-safety lead:
nothing for the model to fool — there is no inference loop in the
verifier. An immeasurably stronger claim than “fast lookup,”
and the only one that prices.
Audited, replicable on your Mac in 90 seconds:
/pmu-simulator/demo §F.
The 12 axes are not arbitrary.
Modeled on the Six Human Needs — three meld pairs, three cardinals,
each cardinal has two parents:
A·Strategy (long-term) = Connection × Significance ·
B·Tactics (medium-term) = Contribution × Growth ·
C·Operations (short-term) = Uncertainty × Certainty.
Your real lattice is your problem-space taxonomy — same gate,
different N.
Moat — four layers
- Patent: the cache notarizes what software
cannot distinguish (S≡P≡H anchored 2026-05-23)C4/C5.
- AC0 architecture: one combinational
op; not refactorable from a Python stackC3.
- First-mover network effect: the first
underwriter writes the receipt format every later one
translates fromLIVE-1.
- Switching cost: 6 months of σ-baselines
and a co-authored case study don't transfer.
If the cache fingerprint distinguishes software-identical agent actions at 3.4σ in 155 ns —
— and the same 144-cell receipt is portable across model, vendor,
and box — then the next five facts follow without further claim:
- The 144-cell receipt becomes a verifiable résumé
for any agent or human worker. The cells that light tell
you the competence pixel; the cells that drift tell you the
rumble strip. Hire what fits this cell; stop searching
for “someone who might.”
▸ receipt D4
· C3
- A receipt that prices is what an underwriter writes against.
The same row that anchors 25 bps insurance anchors 50–100 bps
EU AI Act compliance, 5–10 bps counterparty exposure, capital
gating, and ultimately reputation — by construction, not by
new claim. The visa stack is the answer to “what does the
chip license you to participate in, beyond insurance.”
▸ receipt §10.3
· F2-F6
- The job requirement compiles into a competence pixel; the cell match IS the placement.
The Goldilocks assignment finds the worker — no query, no
traversal, no similarity score. The marketplace runs on
verified, not searched. Zero latency. Zero friction.
▸ receipt C4
· B6
- The receptionist / voice-agent companies running probabilistic guardrails cannot pass actuarial review.
Their post-mortem rulebook (“don't quote above $500”)
lowers probability; it cannot delete the semantic possibility.
When the first claim lands against an ungrounded deployer, your
structural cure — S=P=H + cache fingerprint — is the only
thing they can buy to keep operating. The Wild West becomes the
roll-up; the holding company that owns the substrate owns the
market share that becomes insurable.
▸ receipt F2
· C5
- So the chip is not a feature.
It is the substrate of the Zero-Latency, Zero-Friction
Economy — the Visa standard for verified competence,
agent or human, deterministic at the cache line, priced per
inference, settled on a hardware event the carrier can read.
▸ receipt §10 — the meaning closer
You don't have to be told. The cells tell you.
➞ map of maps · visa of agentic competence · reach IS verify.
REACH IS VERIFY
Pay the regulator.
Pay the premium.
Or pay us.