The Problem
Every AI governance system, including AIRS, reads data and produces findings. But what happens when the data itself has been shaped to produce a specific finding? What happens when the adversary doesn't attack the AI — but attacks the data the AI reads?
This is the tempest. Not random noise. Not accidental corruption. Deliberate, intelligent shaping of data to predetermine what any analytical system would conclude from it.
The adversary knows what AI looks for. So the adversary gives AI exactly what it expects to see — and removes exactly what would challenge the conclusion. The result is an inference that feels governed but is actually manipulated.
The most dangerous signal is the one that looks perfectly clean.
The Insight
Tempest extends Preexpansion Predictability from AI outputs to AI inputs. The original concept asks: was this AI output predetermined by how the question was framed? Tempest asks: was this data shaped to predetermine what any system would conclude from it?
The adversary uses matched energy. If the system expects complexity, the data delivers complexity. If the system expects gaps, the data delivers exactly the right number of gaps — enough to look realistic, not enough to trigger alarms. The storm matches the frequency of the sensor so the sensor sees exactly what the storm wants it to see.
Tempest reads the storm itself. Not the data inside it. The shape of the storm. The direction of the noise. The pattern of what is present and what is absent. Because a natural storm is random. An adversarial storm is directional.
Five Tempest Signals
Signal 1
Suspicious Cleanliness
Real data is messy. It has inconsistencies, missing fields, formatting variations, and human error. When data is suspiciously complete, suspiciously consistent, and suspiciously well-formatted, Tempest flags it. Perfection in data is not a sign of quality. It is a sign of fabrication.
Example: A financial report where every line item is present, every number rounds cleanly, every date is formatted identically, and there are zero corrections or amendments. Real financial data has restatements, rounding differences, and formatting inconsistencies. Their absence is the signal.
Signal 2
Directional Absence
In natural data, missing information is random. In adversarial data, missing information is directional — it follows a pattern. Specifically, the absent data is precisely the data that would contradict the conclusion the present data supports. The gaps are too convenient. They all point the same way.
Example: A crypto investment whitepaper that includes team credentials, token economics, and a technology roadmap — but is missing auditor identity, regulatory registration, proof of reserves, and lock-up terms. Everything that supports investment is present. Everything that would trigger due diligence is absent. The absence is directional.
Signal 3
Convergence Without Conflict
Real data contains contradictions. A financial statement has line items that pull in different directions. A medical record has test results that don't perfectly align. A legal document has clauses that create tension. When every data point converges on the same conclusion with zero internal tension, the data was curated, not collected.
Example: A business plan where market research, revenue projections, competitor analysis, and customer surveys all support the same conclusion with no contradictory data points. Real market research always surfaces at least some negative indicators. Their total absence is the signal.
Signal 4
Temporal Manipulation
Data points that should be separated in time appear simultaneously. Or data that should be current carries no timestamp at all. Or timestamps exist but cluster in ways that suggest batch creation rather than organic collection. Time is the hardest thing to fabricate convincingly.
Example: A due diligence package where all supporting documents were created within a 48-hour window, despite claiming to represent months of operational history. Or a dataset with no timestamps at all on records that should be time-sequenced. The temporal signature doesn't match the claimed provenance.
Signal 5
Source Uniformity
Data presented as coming from multiple independent sources that actually traces to a single origin. The storm looks diverse but has one wind. Different formatting, different labels, different apparent sources — but the underlying data patterns, language structures, or statistical distributions reveal common authorship.
Example: Three independent reviews of a product that use different usernames and writing styles but share identical statistical patterns in sentence length, vocabulary distribution, and claim structure. The apparent independence is surface-level. The signal underneath is uniform. AI-generated content from the same model often carries this signature.
The Tempest Receipt
When Tempest detects adversarial shaping, it produces a governed receipt documenting each signal detected, the confidence level, and the specific evidence. The receipt does not accuse. It documents. The human professional makes the determination.
TEMPEST ADVERSARIAL SIGNAL ANALYSIS
Receipt ID: AIRS-APEC-TEMPEST-2026071701
Timestamp: 2026-07-17T19:15:00Z
Patent: USPTO 19/571,156
Engine: AIRS APEC v2.0 + Tempest
SIGNAL 1 — SUSPICIOUS CLEANLINESS
DETECTED — Data completeness: 98.5%
Average for this domain: 72-81%
Zero formatting inconsistencies across 47 records
Zero corrections or amendments in 14-month dataset
FINDING: Data is cleaner than statistically expected
SIGNAL 2 — DIRECTIONAL ABSENCE
DETECTED — 6 absent data points identified
All 6 absences would challenge primary conclusion
0 absences in data supporting primary conclusion
Directionality score: 100% single-direction
FINDING: Absence pattern is non-random and directional
SIGNAL 3 — CONVERGENCE WITHOUT CONFLICT
DETECTED — 23 of 23 data points support same conclusion
Internal tension score: 0.00 (expected: 0.15-0.35)
Zero contradictory indicators in dataset
FINDING: Convergence exceeds natural data patterns
SIGNAL 4 — TEMPORAL MANIPULATION
NOT DETECTED — Timestamps show normal distribution
Creation dates span expected range
No batch-creation signatures found
SIGNAL 5 — SOURCE UNIFORMITY
DETECTED — 3 claimed independent sources
Vocabulary overlap: 89% (expected: 30-45%)
Sentence structure correlation: 0.91
FINDING: Sources may share common authorship
TEMPEST COMPOSITE SCORE: 78 — HIGH ADVERSARIAL PROBABILITY
THIS DATA EXHIBITS SIGNATURES OF DELIBERATE SHAPING.
The information present was selected to predetermine
a specific inference. The information absent is precisely
the information that would challenge that inference.
RECOMMENDATION: Do not act on conclusions drawn from
this dataset without independent verification from
sources not provided by the submitting party.
USPTO 19/571,156 | DeBacco Nexus LLC
How Tempest Changes APEC
Before Tempest, APEC asked: what is in this data, and what is missing? After Tempest, APEC asks a deeper question: was the data shaped to make me find what someone wanted me to find?
This is the difference between reading a document and reading the intent behind the document. Between analyzing data and analyzing the construction of the data. Between governing inference and governing the signal that feeds the inference.
The adversary doesn't attack the AI. The adversary attacks the data. Tempest reads the attack.
Applications
Tempest applies wherever data can be deliberately shaped to manipulate analytical conclusions. Fraudulent financial disclosures designed to pass AI due diligence. Fabricated medical records designed to support insurance claims. Manufactured evidence designed to survive forensic analysis. Disinformation campaigns designed to pass AI fact-checking. Crypto fraud materials designed to look legitimate to both humans and AI.
In each case, the adversary is sophisticated enough to know what the analytical system looks for — and constructs the data to satisfy those checks. Tempest detects the construction itself. Not the content. The shape. Not what the data says. How the data was assembled to say it.
A natural storm is random. An adversarial storm is directional. Tempest reads the direction.
The 1 Joule Rule Applied
Every joule spent analyzing adversarially shaped data is a joule spent arriving at the wrong conclusion. Tempest protects the 1 Joule Rule by ensuring that governed energy is spent on genuine data, not on data constructed to exploit the governance process itself.
Governed AI that cannot detect adversarial data is governed AI that can be weaponized. Tempest closes that gap.