Experimental

Recursive Self-Improvement

Governed recursive cycles that continuously optimize workflow topology, agent placement, and role-bound specifications. Every iteration feeds evidence into the next — never uncontrolled.

UNIFIED DECISION INGESTION

Every Decision Source, One Substrate.

Normalize approvals, meetings, chat, PR reviews, and execution logs into a single event substrate with immutable timeline reconstruction.

Approvals

Ringi & sign-offs

Meetings

Minutes & notes

PR Reviews

Code decisions

Chat Logs

Slack, Teams

Exec Logs

Execution traces

Unified

Event Substrate

Immutable timeline

Audit-grade retention

Replay capability

Branch Detection

Value Attribution

Conflict Analysis

DECISION BRANCH DETECTION

Find the Hidden Judgment Points.

Analyze ringi approval logs to identify decision branch points where outcomes diverge. Heatmap intensity reveals value conflict at each junction — the meta values behind each branch are the real architecture of your organization.

Decision branch point
Conflict intensity (low → high)
meta_a vs meta_bCompeting values
DECISION BRANCH EXTRACTION PIPELINE

From Approval Documents to Judgment Architecture.

Every branch point, meta value tag, and value context extracted from approval documents feeds directly into the recursive improvement loop. MARIA OS services are internally connected — each extraction cycle sharpens the next.

scan
build
extract
complete
Approval Policy
Procurement SOP
Budget Matrix
Meeting Minutes
Risk Framework
Scanning... 0%

Connected Services

Decision Scanner

Approval documents parsed into structured branch topology

Value Scanner

Meta value tags extracted at every judgment point

Workflow Scanner

Branch patterns fed back to rewrite workflow DAGs

Recursive Loop

Improved workflows generate new decisions — cycle repeats

Feeds next cycle
THREE-GATE CONTROLLED AUTONOMY

Minimal Gates, Maximum Control.

Three gates is enough. Too many gates suffocate. Too few invite chaos. Each gate narrows the decision funnel — only governed actions pass through.

ALL DECISIONS ENTER

Industry Safety Gate

MANDATORY

Regulatory, compliance, legal constraints. Non-negotiable.

BLOCKED

Regulatory violations, safety hazards, legal exposure

PASSES

All compliant decisions, standard operations

Company Value Gate

ALIGNMENT

MVV alignment check. Blocks decisions that contradict practiced values.

BLOCKED

Value contradictions, strategic drift, culture violations

PASSES

MVV-aligned decisions, value-consistent actions

Structural Change Gate

HUMAN REQUIRED

Workflow redesign, role changes, parameter shifts. Always requires human approval.

BLOCKED

Unauthorized scope changes, privilege escalation, design rewrites

PASSES

Approved structural changes only

GOVERNED EXECUTION

Audited, traceable, reversible

WORKFLOW SELF-REWRITE

Before Recursion vs. After Recursion.

Each recursive cycle identifies bottlenecks, removes redundant nodes, and optimizes approval topology. Workflow rewrites are treated as structural-change proposals — always gate-approved.

BEFORE — 9 nodes, 2 bottlenecks
AFTER — 6 nodes, streamlined

Nodes Reduced

9 → 6

-33%

Bottlenecks

2 → 0

-100%

Avg. Lead Time

5.2d → 1.8d

-65%

Repeat Rejections

28% → 4%

-86%

RECURSIVE IMPROVEMENT — BEFORE / AFTER

Full-Stack Recursive Rewrite.

再帰的改善はプロンプトとワークフローだけではない。知識ベース・DB接続・権限スコープも含めた全層を自動分析・改善提案する。 すべての変更はStructural Change Gateを通過 — 人間承認なしには適用されない。

System Prompt

Before (v1.0)

role: Audit Evidence Collector
goal: Collect audit evidence from sources
−strategy: Collect all available documents
−validation: Manual review after collection
−error_handling: Report errors to human
output_format: structured_json
−retry_policy: none
−coverage_check: disabled

After (v1.1)

role: Audit Evidence Collector
goal: Collect audit evidence from sources
+strategy: Priority-ranked collection by risk score
+validation: Auto hash-verify + schema check
+error_handling: Auto-retry 3x → escalate to CEO gate
output_format: structured_json
+retry_policy: exponential_backoff(max=3)
+coverage_check: assert >= 95% before pass

Coverage Rate

72%97%
+25%

False Negatives

18%3%
-83%

Retry Success

0%89%
+89%

Human Escalations

100%12%
-88%

Workflow

Before (9 nodes)

[Request] → [Manager Review] → [Legal Check]
− → [Director Approve] → [Compliance Review]
− → [CFO Gate] → [Execute] → [Audit Log]
− → [Complete]
−nodes: 9 | bottlenecks: 2 | avg_lead: 5.2d
−repeat_rejections: 28% | human_touches: 5

After (6 nodes)

[Request] → [Auto-Check + Risk Score]
+ → [Value Gate (CEO if risk > 0.6)]
+ → [Execute ∥ Audit Log] → [Complete]
+nodes: 6 | bottlenecks: 0 | avg_lead: 1.8d
+repeat_rejections: 4% | human_touches: 1

Node Count

96
-33%

Lead Time

5.2d1.8d
-65%

Rejections

28%4%
-86%

Human Touches

51
-80%

Knowledge Base

Before (static)

knowledge_sources:
− - static_docs: /shared/policies/*.pdf
− - update_cycle: manual (quarterly)
− - embedding_model: text-embedding-v1
- vector_store: pinecone
− - chunk_strategy: fixed_512
− - freshness_check: disabled

After (live)

knowledge_sources:
+ - live_docs: /shared/policies/** + Slack + Jira
+ - update_cycle: event_driven (on_change)
+ - embedding_model: text-embedding-v3-large
- vector_store: pinecone
+ - chunk_strategy: semantic_split(overlap=0.15)
+ - freshness_check: stale_alert > 7d

Source Count

1247
+292%

Staleness

90d< 1d
-99%

Retrieval F1

0.610.92
+51%

Hallucination

14%2%
-86%

DB Connection

Before (fixed)

data_connections:
− - primary: postgres (read-only, single)
− - pool_size: 5 (fixed)
− - failover: none
- audit_log: enabled
− - query_timeout: 30s
− - schema_migration: manual_apply

After (adaptive)

data_connections:
+ - primary: postgres (read/write, replica-aware)
+ - pool_size: auto(min=5, max=50, by_load)
+ - failover: auto_switch_replica(health_check=5s)
- audit_log: enabled
+ - query_timeout: adaptive(p95=2s, max=10s)
+ - schema_migration: zero_downtime_rolling

Avg Latency

120ms18ms
-85%

Uptime

99.2%99.97%
+0.77%

Pool Util

82%34%
-59%

Failed Queries

3.1%0.05%
-98%

Permissions

Before (global)

authority_scope:
− - access_level: global_read
− - write_scope: all_zones
− - approval_required: never
- coordinate: G1.U1.P*.Z*.A*
− - risk_budget: unlimited
− - escalation: manual_only

After (scoped)

authority_scope:
+ - access_level: zone_scoped(own_zone + adjacent)
+ - write_scope: own_zone_only(cross_zone=request)
+ - approval_required: if risk_score > 0.4
- coordinate: G1.U1.P*.Z*.A*
+ - risk_budget: 0.8/day(auto_throttle)
+ - escalation: auto_CEO_gate(severity >= high)

Overprivilege

67%3%
-96%

Unauthorized Ops

22/mo0/mo
-100%

Blast Radius

全Zone1 Zone
Scoped

Audit Pass

71%99%
+28%

Recursive Improvement Cycle — 全5層を対象

OBSERVE全層KPI収集ANALYZEボトルネック特定PROPOSEPrompt/WF/KB/DB/権限改善案VALIDATE3-Gate検証APPROVE人間承認DEPLOY段階的適用次サイクルへ
AGENT PLACEMENT & SPEC EVOLUTION

Role-Bound Agents, Evolving Specs.

Agents are placed by dependency pressure, not static quotas. Authority scope, risk budget, and approval thresholds evolve as versioned specs — never autonomously.

Zone Topology

3

Power

4

Capital

5

Institution

8

Mass

3

Narrative

2

Chaos

Spec Evolution

v1.0

Read-only analytics

Risk ≤ NoneApproval: N/A
v1.1

+ Low-risk execution

Risk ≤ < 0.2Approval: Auto
v1.2

+ Cross-zone coordination

Risk ≤ < 0.4Approval: Peer review
v2.0

+ Structural proposals

Risk ≤ < 0.6Approval: Human gate

Invariant: No agent can autonomously expand its own spec version. All upgrades require human approval through the Structural Change Gate.

RECURSIVE SCOREBOARD

Measure Every Recursive Cycle.

Drift Detection

84

Rejection Prevention

72

Lead Time Reduction

91

Audit Explainability

88

Value Consistency

66

Agent Utilization

77

Run a Recursive Cycle