For Engineers

This is not a prompt framework.
It's a state-transition pipeline.

PE → STME → RSTA → OSD is a sequential toolchain: clarify the problem, structure the decision space, stabilize semantic continuity, then observe drift — with each stage producing structured output the next stage consumes.

The Pipeline

Each stage has a single responsibility and a bounded output contract. None of the stages make the decision for the user — they constrain and expose the structure the decision happens inside.

INPUT

PE — Problem Explorer

Clarifies the problem before any decision logic runs. Tracks six dimensions of problem clarity. Outputs only what the user stated — no AI-inserted assumptions.

in: raw question · out: clarified problem statement (6-dim)

STRUCTURE

STME — Structured Multi-State Transition & Evaluation

Decomposes the clarified problem into states and transitions, flags structural pressure, and ranks transitions. Does not output a recommendation.

in: clarified problem · out: state graph + transition ranking

STABILIZE

RSTA — Recursive State Transition Architecture

Formal model of how semantic states transition across turns. Provides the theoretical constraints that keep a long interaction from drifting off its own state graph.

in: state graph · out: semantic stability constraints

OBSERVE

OSD — Observable Semantic Dynamics

Makes semantic state evolution visible in real time — not prediction, observation. Distinct from drift detection, intent tracking, and mechanistic interpretability.

in: live interaction · out: SAI trace + case classification

OSD Probe — the observability tool

OSD Probe is the empirical instrument behind the OSD framework. It runs a multi-judge pipeline (GPT / Claude / Gemini) against a target conversation, computes an SAI (semantic stability indicator) trace turn-by-turn, and classifies observed behavior against a growing library of documented case types.

Multi-judge scoring across three independent model families, reducing single-model judge bias
SAI display per turn, so drift is visible at the point it happens, not reconstructed after the fact
JSONL import / export for reproducible offline analysis
Subject Consistency Check — flags when the same probed subject behaves inconsistently across repeated runs

[ GitHub · osd-behavioral-probe ][ Live Tool · osd-behavioral-probe.vercel.app ]

Before / After

What changes when OSD + RSTA sit on top of a normal decision or agent pipeline:

BEFORE

Semantic drift is noticed only after an output fails or a user complains
No turn-level record of how the state changed — only the final output
Judge bias from relying on a single model to evaluate itself
Decision structure and decision content are entangled in one response

AFTER

SAI trace shows drift turn-by-turn, before it reaches an observable failure
Every state transition is logged and reconstructable, not just the endpoint
Three independent judge models cross-check each observation
PE/STME separate problem structure from problem content, RSTA/OSD separate state from output

Repos & Live Tools

STME

Decision structuring engine. Preprint on SSRN + USPTO provisional patent pending.

[ SSRN ][ GitHub ][ Demo ]

RSTA

Semantic state transition theory. Preprint on Zenodo, DOI: 10.5281/zenodo.20603119.

[ Zenodo · DOI ][ GitHub ]

OSD

Observability layer. Preprint on Zenodo, DOI: 10.5281/zenodo.20758240.

[ Zenodo · DOI ][ GitHub ][ Live Tool ]

PE

Problem Explorer. Supports Claude, GPT, and local Ollama models.

[ Launch ]

Integration notes

Practical detail for anyone wiring this into their own pipeline:

OSD Probe's judge pipeline currently targets claude-sonnet-4-6 as the primary evaluation model, cross-checked against GPT and Gemini judges
Case data is stored and exchanged as JSONL — each line is one turn-level observation, making diffing and replay straightforward
STME's state/transition output is a plain structured object — it can sit in front of any agent loop as a pre-decision structuring step without touching the agent's own reasoning
None of the tools require the target model to be instrumented internally — OSD observes at the output/interaction layer, not inside model weights. It complements, but does not replace, mechanistic interpretability

PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance · PIDA-LAB · AI Identity · Decision Structure · Human Autonomy · Long-Term Human–AI Relations · Responsibility & Governance ·

This is not a prompt framework.It's a state-transition pipeline.