Prometheus-Research

Truth-driven, auditable research for evolving agents in an execution_world.

Chinese · Docs Index · V11 Index · V10 Index · V12 Index

---

What this repository is (project goals, not technical details)

This is the research record of the Prometheus project. The goal is to preserve reviewable evidence chains rather than narratives: when we claim a result, a third party should be able to trace it to concrete artifacts.

V11’s core positioning is: in an execution_world (exchange truth chain), build an auditable, ablation-friendly, reproducible evolutionary research workflow, where every conclusion must point back to evidence.

---

Foundation (methodology SSOT)

This repository is built on a single, shared measurement method:

• V10 — A Minimal Method for Measuring Complex Systems (Evolutionary Probing): docs/v10/V10_METHOD_MEASURING_COMPLEX_SYSTEMS.md
• Whitepaper edition (narrative entry, bilingual):
  • English: docs/whitepaper/README.md
  • 中文: docs/whitepaper/README_CN.md

This document defines how we measure (measurability gate, NOT_MEASURABLE discipline, measurement-bias dimensions), and is treated as a foundational reference for V11/V12 research contracts.

---

V11 goals (what we aim to deliver)

• A reviewable truth chain: any statement about positions / capital / fills / fees must be traceable to exchange-retrievable JSON (or explicitly marked as unknown / NOT_MEASURABLE).
• Evidence-first, fail-closed: if critical evidence is missing or the chain breaks, the run must downgrade the conclusion (NOT_MEASURABLE) or stop (fail-closed)—no silent defaults.
• Ablation comparability: every new mechanism (e.g. C-dimension social probes) must be evaluable via controlled ablations under comparable conditions, without pre-written narratives.
• No legacy semantic contamination: V11 execution_world must enforce “physical inaccessibility” of legacy/v10 execution semantics in the dependency closure.
• Sustainable research record: SSOT documents are additive-only; breaking changes require a major bump and re-running minimal PROBE acceptance.

---

V11 current focus (what we measure)

• Evidence closure & audit consistency: run manifests, evidence_refs, paging closure, orphan detection—so claims remain verifiable.
• Minimal, fixed-dimension probes in execution_world: strict mask discipline; unknown is never fabricated as zero.
• Ablation templates: e.g. C_off vs C_on, emphasizing comparability and fact-only outputs (not interpretations).

---

V12 direction (world modeling + life system)

V12’s positioning is:

• V10 = evidence chain is born
• V11 = evidence chain is industrialized
• V12 = world modeling + life system comes online

V12 primary deliverables (research-side, ordered)

1. Modeling tool: World Feature Scanner

   • Scan “world facts” (starting from E/exogenous market information) under execution_world
   • Output machine-verifiable evidence bundles and a modeling report

2. Modeling documents from scanner results

   • Turn scanner facts into SSOT: parameter spaces, enums, limits, availability, NOT_MEASURABLE conditions

3. Refactor Agent genome dimensions

   • Align expressible dimensions to exchange interface parameter spaces (no invented knobs)

4. Event-driven execution (initial)

   • Market data: WebSocket (push/event-driven)
   • Trade execution: REST (request/response)

5. Individual metabolism (replace “death judgment”)

   • Agents adapt internal states/constraints based on truth-backed feedback, instead of binary kill switches

6. Split reproduction by capital doubling (replace “reproduction judgment”)

   • Reproduction becomes a consequence of auditable capital growth, rather than a subjective committee decision

Two hard additions (must be frozen early)

• WS evidence discipline: if market data becomes WebSocket-driven, we must persist subscription + message streams as evidence (otherwise decisions cannot be replayed/audited).
• Dual resource gates in Proxy Trader: enforce both (a) per-agent resource limits and (b) system trading resources (exchange rate limits / endpoint availability / degraded modes) fail-closed, with reason codes.
• Modeling acceptance: the scanner must output a machine-readable “genome alignment table” (field→dimension mapping + mask/NOT_MEASURABLE rules), otherwise genome refactors will drift into subjective design.

V12 observation goal (what we want to see, under limited resources)

Under limited compute, we do not aim to “cover the evolutionary space”. We aim to observe repeatable structure under an auditable world contract:

• Clustering: agents/behaviors form stable clusters across runs/windows (not one-off plots).
• Dimension collapse: effective degrees of freedom shrink over time (e.g., effective rank / PCA variance concentration), with clear controls to rule out “collapse by missing data / gates / defaults”.

Methodology note:

In a high-dimensional genotype space, even millions of agents over millions of generations cover only a negligible fragment of the evolutionary landscape. Prometheus does not aim to exhaust this space, but to detect whether selection introduces statistically detectable bias in observed evolutionary trajectories.

---

Start here

• V11 entry (recommended): docs/v11/V11_RESEARCH_INDEX.md
• V10 entry (legacy mainline): docs/v10/V10_RESEARCH_INDEX.md
• Docs index: docs/README.md

Note: V11 documents have been physically separated into docs/v11/. Any V11 files under docs/v10/ are kept as stub pointers to avoid breaking historical links.

---

Research vs Quant (responsibility boundary)

• Prometheus-Research: project goals, acceptance criteria, decision records, SSOT docs, and evidence entrypoints.
• Prometheus-Quant: implementation and run artifacts (code, logs, run_dir evidence bundles, auditor outputs).

---

License

MIT License. See LICENSE.