Innovation Inventory

Innovation Inventory

Three Patterns With No Academic Precedent

“When you build a system that runs 270+ sessions, 168 scripts, 25 hooks, 33 skills, and 420+ lessons learned — and you measure it — you find things nobody has published about.”

---

What This Page Is

This series began as a practitioner’s account of applying military doctrine to AI coordination. Along the way, the vault — a live Obsidian knowledge management system operating as a multi-agent AI laboratory — produced fourteen distinct coordination patterns. Most have academic parallels. Three do not.

This page documents those three HIGH-novelty patterns: what they are, what academic field they extend, what gap they fill, and what evidence supports them. All three emerged from operational necessity, not theoretical design. They were invented under time pressure, refined through 420+ lessons learned, and measured across 270+ sessions.

The claim is specific: these patterns work in a live system, fill gaps in published research, and are independently verifiable from the vault’s git history.

---

The Full Innovation Inventory

Fourteen patterns emerged from vault operations. Novelty is assessed against published academic literature as of April 2026.

#	Pattern	Novelty	Academic Field	Gap
1	Tetris Task Primitives	HIGH	Constraint scheduling, workflow automation	No published work on composable atomic AI task operations
2	Gravity-Fed Pipeline	HIGH	Continuous improvement (DMAIC), cognitive architecture	No published work on self-improving governance-as-momentum
3	Toboggan Doctrine	HIGH	Behavioral economics, formal methods	Choice architecture applied to AI agent enforcement — novel
4	PAT Brainstorming	MED-HIGH	Ensemble methods, voting theory	Parallel orthogonal review extends Condorcet
5	OODA + IG&C	MED-HIGH	Decision theory, military science	Boyd’s advanced framework applied to AI — unexplored
6	OC AAR as Facilitated Dialogue	MED	Institutional learning, AAR doctrine	TC 25-20 extended to AI agent teams
7	Standing Task Orders	MED	State machines, workflow engines	Template-based persistent state across sessions
8	Knowledge Wells	LOW-MED	Information retrieval, RAG	Pull-based domain loading (47 wells)
9	CPI Scorecard	MED	Lean Six Sigma, dashboards	Auto-populated from gravity deposits — no manual entry
10	Multi-Agent Role Identity	MED	Agent architecture, organization theory	Named persistent specialists vs. anonymous workers
11	MDMP for AI Planning	LOW-MED	Military planning, AI planning	FM 5-0 adapted to AI task decomposition
12	Gate A/B Structural Enforcement	MED	Formal verification, CI/CD	PreToolUse/PostToolUse hooks as formal verification gates
13	DMAIC Integration	LOW-MED	Quality management, process engineering	Lean Six Sigma woven into AI operational backbone
14	Ralph Loop	MED	Batch processing, task schedulers	Serial/parallel executor with embedded DMAIC metrics

---

Pattern 1: Tetris Task Primitives

What It Is

Every task in the vault decomposes into a small set of atomic operations — primitives — that combine like Tetris pieces. Thirteen primitives cover 93% of all vault work. Six recurring combinations (“stacks”) account for the majority of sessions.

The 13 primitives: Research, Read, Classify, Enrich, Route, Create, Edit, Validate, Deploy, Monitor, Diagnose, Coordinate, Archive.

Each primitive has standardized inputs and outputs. A “Research” primitive takes a question and produces findings. An “Enrich” primitive takes a file and produces an enriched file. They compose: Research → Classify → Enrich → Route is the inbox triage stack. Research → Create → Validate → Deploy is the skill creation stack.

What Academic Field It Extends

Constraint-based scheduling (Gomes et al., AAAI 2007)¹ demonstrated that structural constraints produce optimal multi-agent schedules. Workflow automation research has produced task decomposition frameworks for decades.

Formally, Tetris task primitives define a typed workflow algebra: each primitive is a typed function (input schema → output schema), composition is function application, and the six recurring stacks are pre-compiled function chains.² Coverage(S) = |primitives_used| / |primitive_universe| ≈ 93%. Six recurring stacks account for 93% of session work — Pareto concentration tighter than typical operational distributions.

What’s New

No published work treats AI task operations as composable atomic units with standardized I/O contracts. Existing workflow systems define tasks as opaque black boxes. Existing AI agent frameworks define tasks as natural language prompts. Neither provides the composability that emerges when primitives have typed inputs, typed outputs, and measurable cycle times.

The vault demonstrates that primitive-level decomposition enables:

• Predictive routing: If a task decomposes into known primitives, execution time and agent assignment can be predicted before the task starts.
• Cross-session learning: Primitive-level metrics (cycle time, error rate, rework frequency) transfer across tasks. A “Research” primitive that takes 45 seconds in one context takes roughly 45 seconds in another.
• Dependency coloring: Primitives with shared dependencies (same file, same git index) are colored to prevent contention. This is constraint scheduling applied at the operation level, not the agent level.

Figure 1 — Tetris Task Primitives: Composition Architecture

13 atomic task primitives compose into 6 recurring stacks covering 93% of vault session work. Each primitive has typed input/output contracts, enabling Tetris-like composition.

Evidence

• 13 primitives identified from 270+ sessions of operational data
• 6 recurring stacks covering 93% of vault operations
• Task files carry a primitives: field in YAML frontmatter
• Primitive metrics tracked in task-schema.yaml
• Documented in Task-Primitives-Taxonomy.md knowledge well

---

Pattern 2: Gravity-Fed Pipeline

What It Is

The vault’s governance system improves itself without dedicated improvement effort. Each operational action — booting a session, closing a task, running an AAR — deposits metrics and artifacts into standardized locations. Downstream processes (CPI scorecard, quality dashboards, lesson learned shards) consume these deposits automatically. Governance improves because the system is designed so that doing normal work is doing improvement work.

The metaphor: water flows downhill. You don’t pump it. You build channels.

What Academic Field It Extends

DMAIC continuous improvement (Lean Six Sigma) defines a structured improvement cycle: Define, Measure, Analyze, Improve, Control. SOAR cognitive architecture (Laird, 2022)³ integrates episodic memory — experiences that inform future decisions.

The gravity pipeline implements potential energy minimization: each session deposits artifacts at lower-energy states, and the system flows toward the minimum-energy configuration.⁴ The CPI loop acts as a Lyapunov function, monotonically decreasing the system’s defect rate: V(x) ≥ 0 and dV/dt ≤ 0, where V(x) = cumulative defect frequency and x = system state.⁵

What’s New

DMAIC requires deliberate improvement effort — someone must run the cycle. SOAR’s episodic memory is pull-based — the system retrieves relevant episodes when needed. The gravity-fed pipeline is push-based and self-organizing: normal operations deposit artifacts that downstream processes consume without retrieval effort.

No published work describes a governance system where:

1. Normal operations produce governance data (boot metrics, close metrics, AAR findings) as a side effect, not a separate activity.
2. Quality improves through momentum — each session’s deposits make the next session’s analysis richer, without anyone running a DMAIC cycle.
3. The CPI loop closes inside the template — templates carry institutional knowledge from prior sessions, so any agent picking up the template inherits the system’s learning.

Figure 2 — Gravity-Fed Pipeline: Self-Improving Governance Flow

Agents enter heavy with full preparation and flow downhill through gates. Normal operations auto-deposit metrics and artifacts at each stage. The CPI loop at the bottom feeds improvements back into templates for the next session — no manual improvement effort required.

Evidence

• 168 scripts, 25 hooks, 33 skills in the operational pipeline
• Boot metrics: ~580 entries auto-deposited across sessions
• Close metrics: ~90 entries auto-deposited
• CPI scorecard reads gravity deposits — no manual data entry
• 4 fully automated pipelines, 4 cron-ready
• Documented in Gravity-Pipeline-Design-Spec.md and Vault-Operating-Architecture.md

---

Pattern 3: Toboggan Doctrine

What It Is

AI agents don’t follow rules because they understand them. They follow rules because the system makes following rules easier than breaking them. The Toboggan Doctrine applies choice architecture — the behavioral economics principle that how choices are structured determines which choice gets made — to AI agent enforcement.

Instead of walls (hard blocks that agents route around) or suggestions (soft guidance that agents ignore under pressure), the toboggan builds channels: structural paths where the compliant action is the path of least resistance. Agents enter the channel at the top and slide to the correct outcome. The channel doesn’t require understanding. It requires gravity.

Implementation: PreToolUse hooks intercept agent actions before execution. Non-compliant actions receive deny (structural block) rather than ask (advisory prompt). The hook doesn’t explain why — it blocks. The agent, unable to proceed on the non-compliant path, routes to the compliant alternative. The alternative is designed to be obvious and easy.

What Academic Field It Extends

Choice architecture (Thaler & Sunstein, Nudge, 2008)⁶ established that the design of choice environments determines outcomes more reliably than education or incentives. Differential privacy (Dwork, ICALP 2006)⁷ established that formal mathematical guarantees outperform empirical promises. AI governance (Gebru et al., FAccT 2021)⁸ argued for structural oversight mechanisms over capability-based safety.

The deny hook is jidoka applied to agent governance: errors caught at the point of production.⁹ The template is the poka-yoke: structurally preventing the error before it can be made.⁹ The compliance rate Γ(t) = 1 − |violations(t)| / |enforcement_opportunities(t)| → Γ ≈ 1.0 sustained over 270 sessions after framework deployment.

The mechanism is incentive-compatible:¹⁰ the deny hook makes non-compliance a dominated strategy (no benefit, certain cost), while the template makes compliance the minimum-effort path. Agents don’t need to understand the rules — the channel structure makes compliance the path of least resistance.

What’s New

No published work applies choice architecture to AI agent enforcement. The existing AI safety literature focuses on:

• Alignment: Training the model to want the right things (RLHF, Constitutional AI)
• Guardrails: Filtering outputs after generation (content filters, toxicity classifiers)
• Capability control: Limiting what the model can do (sandboxing, permission systems)

The toboggan doctrine operates at a different level: it structures the operational environment so that the compliant action is the default, the non-compliant action is blocked, and the agent doesn’t need to understand why. This is not alignment (the agent’s values are unchanged), not guardrails (the filtering happens before generation, not after), and not capability control (the agent retains full capability — it just can’t use it on the blocked path).

The closest parallel is Dwork’s differential privacy: a formal guarantee that holds regardless of the adversary’s strategy. The toboggan hook holds regardless of the agent’s “intent” — it fires on the action, not the reasoning.

Figure 3 — Toboggan Doctrine: Wall-Based vs Channel-Based Governance

Wall-based governance blocks non-compliant actions reactively — agents route around or fail. Channel-based governance structures the environment so the compliant action is the path of least resistance. The toboggan doctrine operates at the channel level.

Evidence

• 25 hooks in production, enforcing compliance via deny pattern
• Hook effectiveness data tracked in hook-effectiveness.jsonl — T-690 DONE: Pareto analysis (4,882 entries, 16 sessions). 6 keep, 5 noise, 12 observers. See Hook-Effectiveness-Pareto-Report.md
• Zero gate violations after compliance framework deployed (Paper 3, [^60])
• “Kill zone” channel metaphor originated in session taupe (2026-03-29)
• 420+ lessons learned, with structural fixes replacing behavioral guidance at 3-strike threshold
• Documented in Herding-Cats-Toboggan-Doctrine-DRAFT.md (48K words, Paper 8 candidate)

---

Academic Foundations

These three patterns did not emerge in a vacuum. They build on established research across multiple fields:

Researcher	Institution	Contribution	Vault Pattern It Validates
John Laird	Michigan	SOAR cognitive architecture — multiple knowledge types	Knowledge wells + skills + hooks
Maja Matarić	USC	Structure over intelligence in multi-robot coordination	OC pattern, agent teams, PAT
Carla P. Gomes	Cornell	Constraint-based multi-agent scheduling	Tetris task primitives
Benjamin Grosof	MIT/IBM	Defeasible logic — rule systems with exceptions	Hook/rule/skill architecture
Cynthia Dwork	Harvard	Differential privacy — formal guarantees	Structural deny hooks
Timnit Gebru	DAIR Institute	Structural governance over capability	Gate enforcement, oversight patterns
Andrej Karpathy	Eureka Labs	LLM OS — layered operating architecture	Three-layer vault architecture

Full citations: References & Bibliography

---

What Makes This System Different

This is not a research paper proposing theoretical patterns. This is a running implementation with measurable outputs:

Metric	Value
Total sessions	270+
Git commits	6,000+
Scripts in production	168
Hooks enforcing compliance	25
Skills (executable procedures)	33
Knowledge wells	47
Lessons learned (documented)	420+
CPI metrics (auto-collected)	~670 entries
Task primitives identified	13
Recurring task stacks	6 (93% coverage)

The system improves itself. Every session deposits data. Every AAR produces fixes. Every fix propagates through templates and skills. The gravity pipeline ensures this happens as a side effect of normal operations, not as a separate improvement program.

Three patterns have no academic precedent. The vault provides the evidence. The published series provides the narrative. This page provides the map.

---

---

---

Navigation

← Series Home | Full References | Paper 3: The Live Laboratory | Paper 7: Platform Blueprint

---

Related

• Herding Cats in the AI Age — Home
• Paper 3: The PARA Experiment
• Paper 7: MDMP Platform Blueprint
• References & Bibliography

Footnotes

1. Gomes, Carla P., van Hoeve, Willem-Jan, Selman, Bart, & Lombardi, Michele. “Optimal Multi-Agent Scheduling with Constraint Programming.” AAAI 2007, pp. 1813–1818. ↩

2. The typed workflow algebra framing: each primitive P has type signature τ(P) = (Input, Output). Composition P₂ ∘ P₁ is valid iff Output(P₁) ⊆ Input(P₂). The six recurring stacks are pre-verified composition chains. This formalizes what practitioners discover empirically: not all task combinations compose cleanly. ↩

3. Laird, John E. “Introduction to the Soar Cognitive Architecture.” arXiv:2205.03854, 2022. SOAR integrates procedural, semantic, and episodic memory — the three knowledge types mirrored by vault skills, knowledge wells, and session logs. ↩

4. The potential energy analogy: each artifact deposit moves the system from a higher-entropy state (scattered lessons, unprocessed data) toward a lower-entropy state (structured knowledge, actionable metrics). Entropy H(X) = −Σ p(xᵢ) log₂ p(xᵢ) decreases as artifact deposits constrain the action space for future sessions. Shannon, Claude E. “A Mathematical Theory of Communication.” Bell System Technical Journal, 27(3): 379–423, 1948. ↩

5. Lyapunov stability: a dynamical system is stable if there exists a scalar function V(x) ≥ 0 (Lyapunov function) with dV/dt ≤ 0. The CPI loop satisfies this: defect frequency V is non-negative and decreases monotonically as fixes propagate through templates and hooks. Lyapunov, A.M. “The General Problem of the Stability of Motion.” International Journal of Control, 55(3): 531–534, 1892 (translated 1992). ↩

6. Thaler, Richard H. & Sunstein, Cass R. Nudge: Improving Decisions About Health, Wealth, and Happiness. Yale University Press, 2008. Choice architecture: the way choices are presented systematically influences which choice is made, independent of rational deliberation. ↩

7. Dwork, Cynthia. “Differential Privacy.” ICALP 2006. DOI: 10.1007/11787006_1. The parallel to toboggan hooks: both provide guarantees that hold regardless of adversary strategy, not dependent on adversary intent. ↩

8. Bender, Emily M., Gebru, Timnit, et al. “On the Dangers of Stochastic Parrots.” FAccT 2021. DOI: 10.1145/3442188.3445922. ↩

9. Jidoka (自働化): Toyota Production System principle — machines detect defects and stop automatically, rather than passing defects downstream. Poka-yoke (ポカヨケ): mistake-proofing device that makes defects physically impossible to produce. Both from Shingo, Shigeo. Zero Quality Control: Source Inspection and the Poka-yoke System. Productivity Press, 1986. ↩ ↩²

10. Mechanism design (reverse game theory): design rules/incentives such that each agent’s self-interested strategy produces the desired collective outcome. Hurwicz, Leonid, Maskin, Eric, & Myerson, Roger B. — 2007 Nobel Prize in Economics for mechanism design theory. The deny hook satisfies incentive compatibility: compliance dominates non-compliance in expected value for any agent, regardless of the agent’s objectives. ↩