EVEMISSTechnology

Open framework · Agent Method Execution Protocol

Turn methodologies into executable agent protocols.

AMEP packages a human methodology as a deterministic, auditable protocol that an agent runs — not a prompt it reads. Each method pack is a small Python kernel plus schemas, a CLI, trace memory, and a conservative claim-state vocabulary.

View on GitHub ↗ The five packs Quickstart

Open source · Preparing for release

Why a prompt isn’t enough

Prompts are flexible but weakly binding. Workflows fix order but can’t carry a methodology’s judgment and downgrades. Tools extend reach but never say when, why, or how far. AMEP is the layer in between.

Prompt

Flexible, but no stable state, versioning, audit, or failure protocol.

Workflow

Stable order, but it describes sequence — not a method’s judgment criteria.

AMEP method pack

Repeatable, verifiable, auditable, versionable — a methodology compiled into a protocol.

The method-pack model

Every pack is decomposed into the same twelve layers and runs a bidirectional loop: the agent generates forward, the pack audits backward, and the coupling step revises the next action.

  • 01 Intent contract
  • 02 Input contract
  • 03 Domain vocabulary
  • 04 Action rules
  • 05 Constraint rules
  • 06 Audit rules
  • 07 Output schema
  • 08 Memory rules
  • 09 Tool permission
  • 10 Failure protocol
  • 11 Iteration loop
  • 12 Boundary policy

Bidirectional loop

  1. → Forward

    The agent proposes, drafts, formalizes, calls tools.

  2. ← Backward audit

    The pack traces gaps, requires witnesses, flags risks, marks what cannot be claimed.

  3. ↻ Coupling update

    Backward pressure revises the next forward action; gaps become tasks.

Five method packs

One umbrella protocol, five disciplines. RigorLoop is the reference implementation; the others share its I/O flow, trace memory, and claim-state discipline.

Rigor · Audit

RigorLoop

The reference pack. A bidirectional rigor-audit protocol for mathematical and research agents: claim normalization, gap registers, a citation ledger, and Lean formalization-prep — with conservative claim states instead of overclaiming.

View source ↗ Creation · Routing

GCPR-CreativeOps

Turns a creative, product, research, or strategy task into a GCPR seven-tuple, diagnoses self/problem state, routes a creation strategy (sketch / refine / dimension-lift / …), runs an iterative operation loop, and projects the result back into a feasible domain.

View source ↗ Causal · Prediction

FDCS-CausalOps

Cross-scale, context-sensitive dynamic causal modeling with bounded virtual interventions and a scale-horizon uncertainty guard that downgrades predictions beyond observer capacity — no false precision.

View source ↗ Metacognition

CognitiveDeconstructionOps

A metacognitive pack: routes a minimal sufficient set of reasoning modules, strips a concept to its origin-point candidate, recompiles it, and integrates the module outputs into one reconstruction.

View source ↗ Strategy · World-rules

BoundlessStrategyOps

Hypergame and world-rule reasoning: identifies the game, extracts real vs. false rules, classifies the user’s deeper intent, and routes a safe, lawful, non-coercive strategy class — with handoffs to the other packs.

View source ↗

Not a prompt. Not a skill file.

A skill is a way to package and load a capability into an agent. A method pack is the deterministic, audited engine underneath. You can expose a pack through a skill — but the pack is what runs.

Deterministic & tested

The kernel runs without any LLM call. A unit-test suite across the five packs keeps behavior reproducible — not a prompt that drifts run to run.

Auditable by design

Append-only JSONL trace memory, a conservative claim-state vocabulary, and structured audit findings. Every run leaves an inspectable trail.

Schema-enforced

Outputs are validated against JSON schemas at runtime, and every pack implements the same 12-layer method-pack contract.

Runtime-agnostic

Runs as a local CLI inside any agent — Claude Code, Codex, and others — or standalone. Pure-stdlib Python, no platform lock-in.

Quickstart

Each pack is a self-contained, dependency-free Python package. Run one as a CLI and it writes a structured artifact bundle plus an audit trail.

git clone https://github.com/kakon77777-commits/amep
cd amep/packs/boundless-strategyops

# route a real task — deterministic, offline, no API key
PYTHONPATH=src python -m boundless_strategyops.cli \
  run --input "I want to build an AI tool but can't position it"

# -> output/strategy_report.md, strategy_route.json, safety_note.md, …
View on GitHub ↗ Read the papers

From a methodology you read to a protocol an agent runs.

AMEP is part of the EVEMISS Technology open-source layer — infrastructure for disciplined, auditable agentic work.

Star on GitHub ↗ Talk to us