Open Specification · CC-BY 4.0
TPMN
Truth-Provenance Markup Notation — an open specification language for structuring and auditing AI reasoning.
TPMN defines how AI claims are classified, verified, and traced back to evidence. Anyone can implement it.
Deep dive video — coming soon
Three layers, one verification stack
TPMN — Truth-Provenance Markup Notation
The open specification. Defines five epistemic claim states, the Structural Prohibition Taxonomy (SPT), and the Epistemic Evidence Framework (EEF). Structures how AI reasoning should be classified and audited.
TPMN-PSL — Platform Specification Language
The formal grammar within TPMN. Compiles natural-language prompts into MANDATE — computable, verifiable specifications. Defines the three-phase verification protocol.
TPMN Checker — Reference SAS
The first Sovereign AI Service in the GEM²-AI ecosystem. Runs the TPMN-PSL three-phase pipeline and returns a truth_score for any AI output. Works today via MCP.
Analogous to HTTP (spec) → nginx (implementation). TPMN defines the rules. The Checker enforces them.
Five epistemic claim states
Every AI claim is classified into one of five states. This is the core of TPMN.
⊢
Grounded
Supported by evidence, input context, or verifiable fact
⊨
Inferred
Logically derived from grounded claims, chain visible
⊬
Extrapolated
Beyond available evidence, basis must be stated
⊥
Unknown
Knowledge gap detected, stops inference chain
?
Speculative
Plausible but unverified, possible but uncertain
Three-phase verification protocol
TPMN-PSL structures verification across the full AI generation lifecycle.
P-Phase
Before generation
Validate the prompt contract. Compile NL input into a MANDATE — a verifiable specification the AI must satisfy.
Inline
During generation
Apply epistemic tagging at the claim level. Mark each statement as grounded, inferred, extrapolated, unknown, or speculative.
O-Phase
After generation
Verify output against the original specification. Detect SPT violations, flag extrapolation, score reliability.
Structural Prohibition Taxonomy (SPT)
SPT formalizes three categories of prohibited reasoning transitions that AI systems commonly make. These aren't opinions — they're structural logic errors.
S → T
State → Trait
Treating something mutable as permanent. "Revenue grew 20%" → "This company always grows."
L → G
Local → Global
Treating a local truth as universal. "This benchmark shows X" → "X is universally true."
Δe → ∫de
Short → Mass
Making large claims from thin evidence. One data point → sweeping conclusions.
Architectural lineage
Panini grammar compression
Ontological discretization — categorizing knowledge claims before reasoning begins, drawn from the oldest known formal grammar tradition.
TLA+ formal modeling
Specifications as executable contracts. TPMN borrows notation and rigor from Lamport's temporal logic of actions.
Mathematical logic notation
Epistemic symbols (⊢ ⊨ ⊬ ⊥ ?) derive from standard mathematical logic for provability and truth.
Natural-language annotation
TPMN is designed to be readable by both machines and humans. Epistemic tags attach to natural-language claims, not just code.
TPMN is open and evolving. CC-BY 4.0.