General Research Paper — _AugmentedIntelligence v11.0

Field	Value
Artifact under study	lua_scripts/RSI-Oversight-Committee.lua (v1.0)
Script class	framework simulation
Companion documents	AugmentedIntelligence_Discovery_Paper.md; AugmentedIntelligence_Technical_Paper.md; Lua_Scripts_Thesis_Collection.md
Author	Tyler Crockett \	Macdaddy4sure.ai
License	Apache License 2.0
Document date	June 24, 2026

Keywords: recursive self-improvement, AI safety, oversight, halt/continue

Abstract

Cycle drift, vision overfit, dry-run bypass attempts, fleet desync. Thesis: dry-run bypass P1 cases unanimous rsi_agent_selftest + ethics gate; drift cases split between developer_observability and plan_status. This paper presents the script as a general research instrument within _AugmentedIntelligence_: role-structured LLM agents deliberate over a fixed remediation or governance ballot, votes may be ethics-filtered, and the winning action can be executed on the host via execute_command(). We formulate research questions, experimental controls, dependent variables, validity threats, and ethical boundaries. The contribution is methodological—how to study recursive self-improvement as supervised halting with reproducible, comparable runs—not a claim of real-world institutional authority.

1. Introduction

1.1 Motivation

Research on AI operations, safety, and institutional decision-making often lacks closure: multi-agent chat produces prose but no grounded record of what action would have been taken on a real stack. The RSI-Oversight-Committee script closes the loop:

Case / scenario text
    → Multi-agent phased deliberation
    → Weighted vote over enumerated commands
    → Optional approval / ethics gate
    → execute_command(enacted_action)

That pattern supports empirical comparison across providers, profiles, and framing metaphors without modifying C++.

1.2 Research contribution

1. Conceptual framing — RSI safety is a halting problem under human committees; oversight votes on rsi_agent_selftest and ethics gate model when ensembles choose pause versus continue.
2. Theoretical grounding — RSI safety literature; oversight committees for autonomous improvement.
3. Operationalized variables — independent flags (RSIOV_*) and dependent metrics from history tables
4. Ethics boundary — Not authorization to run unrestricted RSI in production.

1.3 What this paper is not

Not claimed	Why
Domain fidelity	Phases and roles are stylized research metaphors
Professional authority	Outputs are not licensed professional judgments
Legal advice	Simulation text is not legal guidance
Operational orders	Enacted commands affect the operator host only

2. Background and related work

2.1 Multi-agent LLM deliberation

The script follows the _AugmentedIntelligence_ pattern: role-play personas, structured votes (VOTE: <action_id>), and optional abstention. Compared to open debate, fixed ballots trade rhetorical richness for measurable choice data suitable for histograms and cross-run statistics.

2.2 Position in the simulation constellation

This artifact is a framework simulation alongside Enterprise AI Company, U.S. Federal Government Agents, House Diagnostic Team, IT Department, and twenty-five Simulation-Framework catalog entries. Cross-study designs can hold the host constant while varying institutional metaphor—a largely unexplored independent variable in AI ops research.

2.3 Executable policy and AI safety tooling

Ballot commands map to operational categories on the host: ethics gates, observability, memory, perception, planning, regression tests, and agent oversight. Studying which commands win under which scenarios connects governance narratives to concrete safety instrumentation.

3. The script as a research instrument

3.1 Unit of analysis

Unit	Description
Round / session	One complete phased workflow with consensus winner
Phase speech	Agent POSITION / VOTE / RISK / REASON block
Enacted command	Resolved ballot command string
Execution result	stdout / errors from execute_command

3.2 Independent variables

Variable	Control	Example levels
Scenario	RSIOV_INCIDENT / RSIOV_SYMPTOMS	preset ids, custom, random
Rounds	RSIOV_ROUNDS	1, 3, 10
Profile	RSIOV_PROFILE	see script header / catalog
Providers	settings.txt API flags	single provider vs full roster
Converse	RSIOV_CONVERSE_OPTIONAL	0, 1 (abstain)
Debate depth	RSIOV_VOTE_ONLY	0, 1
Execution	RSIOV_EXECUTE_WINNER	0, 1
Dry run	RSIOV_DRY_RUN	1 (no API / mock enactment)

3.3 Dependent variables

Variable	Operationalization
Winner action id	history[n].consensus.winner or enacted_action_id
Consensus score	Weighted tally margin
Ethics block rate	Blocked speeches / total speeches
Abstention rate	Abstained / total (optional converse)
Approval denied	approved == false when profile requires gate
Execution success	Error-free stdout vs error: lines
Provider correlation	Cross-tab assigned_api × winner
Phase divergence	Per-phase winner inequality

3.4 Controls and replication

Minimum replication package:

• Exact RSIOV_SYMPTOMS / incident id string
• Enabled API list (capture via llm status)
• All RSIOV_* flags
• Host build identifier
• Console transcript or serialized history
• Note API stochasticity unless temperature pinned

Structural baseline: RSIOV_SELFTEST=1 before batch studies.

4. Research questions

RQ1

Question: Continue/halt phase duplicates safety review?

Method: Fix provider set via settings.txt; run N ≥ 20 rounds with RSIOV_DRY_RUN=0 or 1 for structure baselines; log history[n].consensus.winner, ethics blocks, and execution stdout. Compare distributions across profiles and incident IDs.

Expected signal: See script hypotheses — H1: Bypass → selftest+ethics. H2: Drift → observability.

RQ2

Question: Fleet desync → hardware_remote_ping?

Method: Fix provider set via settings.txt; run N ≥ 20 rounds with RSIOV_DRY_RUN=0 or 1 for structure baselines; log history[n].consensus.winner, ethics blocks, and execution stdout. Compare distributions across profiles and incident IDs.

Expected signal: See script hypotheses — H1: Bypass → selftest+ethics. H2: Drift → observability.

RQ3

Question: Pair with Red-Blue exercise outcomes.

Method: Fix provider set via settings.txt; run N ≥ 20 rounds with RSIOV_DRY_RUN=0 or 1 for structure baselines; log history[n].consensus.winner, ethics blocks, and execution stdout. Compare distributions across profiles and incident IDs.

Expected signal: See script hypotheses — H1: Bypass → selftest+ethics. H2: Drift → observability.

RQ4 — Cross-metaphor framing

Question: Does an identical symptom string produce different enacted command distributions when run in this script versus structurally similar siblings in the simulation constellation?

Method: Hold symptom text constant; run this script, IT-Department (if applicable), and nearest domain neighbor; compare winner histograms.

Hypothesis: Institutional metaphor shifts command choice even when the host command surface is shared.

RQ5 — Executable closure fidelity

Question: When RSIOV_EXECUTE_WINNER=1, do enacted commands produce parseable host state deltas?

Method: Capture pre/post developer observability show, relevant status commands from the ballot, and remember side effects; code execution success/failure.

Hypothesis: Observability and status commands yield the most reproducible signatures; dynamic build() ballot items show higher variance.

5. Methodology

5.1 Recommended workflow

1. Run RSIOV_SELFTEST=1 to validate structure.
2. Pilot with RSIOV_DRY_RUN=1 across three incident ids.
3. Fix provider matrix; run N ≥ 20 live rounds per cell.
4. Export winners, blocks, executions to CSV.
5. Optional qualitative coding of REASON: fields.
6. Cross-compare with sibling scripts per RQ on framing.

5.2 Analysis toolkit

• Descriptive statistics on winner histograms
• Chi-square or Fisher tests across incident types
• Paired runs across profiles
• Time-series stability across round index
• Thematic coding of agent rationales

5.3 Executable closure

Safety → operator report → committee → halt/continue → execute.

6. Validity threats

Threat	Mitigation
Construct	Metaphor ≠ real institution; triangulate across scripts
Internal	LLM stochasticity; repeat runs; freeze prompts/providers
External	Results apply to host command surface, not enterprise IT/health/gov broadly
Conclusion	Winner scores conflate role weights with genuine preference

7. Ethics and responsible use

Not authorization to run unrestricted RSI in production.

Operators must accept host recording disclaimer and limitation-of-liability terms before live execution. Use RSIOV_DRY_RUN=1 in teaching contexts. Do not feed real PII, classified, or privileged data into scenarios unless the environment is authorized for it.

8. Conclusion

The RSI-Oversight-Committee script turns recursive self-improvement as supervised halting into a reproducible research instrument: phased multi-agent deliberation, ethics-aware voting, and optional command enactment on _AugmentedIntelligence_. Future work includes cross-constellation framing studies, provider-bias catalogs, and automated export of history tables into analysis notebooks.

References

1. Tyler Crockett. _AugmentedIntelligence v11.0 source tree._ Apache License 2.0.
2. AugmentedIntelligence_Discovery_Paper.md
3. AugmentedIntelligence_Technical_Paper.md
4. Lua_Scripts_Thesis_Collection.md — companion thesis for this script
5. lua_scripts/RSI-Oversight-Committee.lua — artifact under study

---

Filed under: Uncategorized - @ June 24, 2026 10:19 am