⬡ TPMN-PSL v0.1.5-p · Public · CC-BY 4.0

Don't write prompts.
Write specifications.

For complex, high-stakes AI workflows: the open epistemic markup standard for AI reasoning. Platform-agnostic. Formally grounded. Built to make AI know what it doesn't know — and mark it at every step.

Read the Spec See the 5 Symbols

The difference TPMN makes

Same information. One version is auditable. One isn't.

Without TPMN
This architecture scales well and will handle production load. The latency will stay under 50ms at peak traffic. The team agrees this is the right approach.
With TPMN
This architecture handled 10k RPS in staging ⊢.
Production load is estimated at 15k RPS ⊨ (Q3 traffic model).
Latency under 50ms at peak ⊬ (no production benchmark yet).
Team consensus on approach ⊨ (3 of 4 engineers).

The 5 epistemic symbols

Every non-trivial claim in a TPMN-governed output carries exactly one tag.

⊢
Grounded
Directly supported by input or verifiable fact
⊨
Inferred
Derived from grounded claims; inference chain visible
⊬
Extrapolated
Beyond evidence; basis must be explicitly stated
⊥
Unknown
Knowledge gap; stops inference chain
?
Speculative
Possible but unverified

Three-phase checking protocol

P-phase validates before. Inline tags during. O-phase verifies after. All three against the same CONTRACT.

P-phase · Before
Pre-flight
Extract CONTRACT from prompt. Establish A_Priori_Grid. Define the computable MANDATE before inference begins.
Inline · During
Generation
Tag every non-trivial claim with its epistemic symbol as it is written. Continuous recalibration against the MANDATE.
O-phase · After
Post-flight
Verify output against CONTRACT. Check SPT violations. Score via TruthReport: epistemic_score ∈ [0,1], contract_score ∈ [0,1], truth_score ∈ [0,100] (integer %). Gate or pass.
Prompt → [P-phase] → LLM → [Inline] → Output → [O-phase] → Trust

The MANDATE

TPMN's core transformation: NL ambiguity → computable specification space.

NL prompt → MANDATE = CONTRACT + A_Priori_Grid + Panini_Categories

Once the MANDATE is established, the LLM reasons within it —
not against hoped-for intent, but against a bounded, decidable structure.
Computable
Category membership rules are decidable: ∀x ∈ D: x ∈ Sᵢ ∨ x ∉ Sᵢ
Provable
Outputs can be formally verified against the CONTRACT in O-phase
Traceable
Every claim traces back to a P-phase category or carries ⊬ escalation
Bounded
Scope is explicit. Out-of-scope claims are structurally prohibited without ⊬

Three operational modes

v0.1.5-p formalizes three workflows. Each runs the same P-phase / O-phase protocol — they differ in what happens with the result.

Strict
Diagnose only
Evaluate content. Flag SPT violations and extrapolation. Halt on ⊬ — no rewrite. Pure audit.
→ TruthReport
Refine
Strip overclaims
Run diagnosis, then compose: remove SPT violations and ungrounded ⊬ claims using only the original input. Re-verify the result.
→ TruthReport + ComposedOutput
Interpolate
Suggest grounding
Run diagnosis, then generate concrete search queries for each ungrounded claim. The caller fetches evidence, then re-runs as refine.
→ TruthReport + SearchSuggestions

Structural Prohibition Taxonomy

Three inference patterns that TPMN-governed output must never produce.

S → T
State → Trait
"This model is slow" (benchmark result) → "This model is inherently slow" (permanent property)
L → G
Local → Global
"Works in our staging environment" → "Works in all production environments"
Δe → ∫de
Incremental → Mass
One data point → broad universal claim without stated evidence count

Start with the spec.

Read the full TPMN-PSL v0.1.5-p specification. Extend it for your domain. Build on it.

Read Full Spec Extension Template