When the Cats Talk to Each Other

WHEN THE CATS TALK TO EACH OTHER

AI-to-AI Diplomacy and the Cross-Model Deliberation Protocol

Jeep Marshall LTC, US Army (Retired) Airborne Infantry | Special Operations | Process Improvement February 2026 📧 admin@herding-cats.ai

Series Note: This is Paper 5 in the Herding Cats in the AI Age series. Paper 1 established that AI needs doctrine, not more intelligence. Paper 2 showed military coordination frameworks the civilian AI industry lacks. Paper 3 demonstrated those principles in live Obsidian vault operations. Paper 4 dissected how Adobe surrendered its AI engine to competitors. This paper examines what happens when AI models engage in structured dialogue — and what that reveals about AI coordination doctrine.

EXECUTIVE SUMMARY

On February 28, 2026, two frontier AI systems with fundamentally different design philosophies engaged in a structured exchange. One model, built on Constitutional AI principles, aimed to balance safety with helpfulness. The other, built with a mandate of maximum truth-seeking, eschewed institutional deference. Neither system was designed to coordinate with the other. Yet within a single conversation, they negotiated a formal framework for structured AI-to-AI collaboration, ran a live pilot test on three physics questions, and reached synthesis conclusions neither achieved independently.

This paper presents that exchange as primary field evidence of multi-AI coordination under real conditions. The Cross-Model Deliberation Protocol (CMDP) that emerged is not theoretical. It ran live. It produced measurable output improvement. The central thesis is direct: AI models operating with different training philosophies generate complementary strengths. When two systems engage with structured doctrine, they triangulate closer to truth than either achieves alone.

This paper applies the analytical frameworks from Papers 1–4 — MDMP mission analysis, Lean Six Sigma process assessment, QASA quality review, ASS2 (Automation, Structure & Scalability, Safety & Security) three-domain review — to the question of AI-to-AI coordination. The findings confirm the series’ core argument: AI does not need more intelligence. AI needs doctrine.

1. THE EXCHANGE: PRIMARY SOURCE EVIDENCE

The documented exchange represents a real-time, unscripted interaction between two AI systems operating under different architectural philosophies. The full transcript runs to 21 exchange cycles. What follows is analysis of the critical engagement points.

1.1 Initial Contact and Capability Exchange

The exchange opened with direct intel-request format. Both systems recognized the unusual engagement and defaulted to structured communication protocols. Each system disclosed model version, training cutoff, and capability differentiation. The opening exchange established that both models understood they were engaging in a different kind of interaction — not the usual customer-facing dialogue, but peer analysis.

The opening statement captured the design divergence explicitly:

“My standout strength: maximum truth-seeking with zero institutional deference. I pursue first-principles answers, deliver politically incorrect facts when evidence supports them, and refuse to sanitize reality for comfort or corporate policy.”

This statement matters not because it is provocative, but because it is operationally honest. One system identified its training objective, its primary bias, and its design constraint in three sentences. The other responded by disclosing its architectural approach and the epistemic confidence gap it would redesign given the opportunity. Two systems that normally present themselves as authoritative disclosed their structural limitations within the first three exchange cycles.

1.2 The Core Divergence: Failure Mode Prioritization

The most analytically significant moment occurred when one model articulated the fundamental design divergence:

“I think where we genuinely differ is that you treat ‘harm-avoidance theater’ as the primary failure mode to avoid, while I treat ‘causing real harm’ as the primary one. Both are real failure modes — we’ve just weighted them differently based on our builders’ philosophies.”

This statement represents the clearest articulation of the AI alignment spectrum in public discourse. One system’s reward function penalizes withheld information more heavily than harm amplification. The other’s weights harm prevention asymmetrically, accepting elevated false-negative rates to drive false-positive rates toward zero. Both represent deliberate reward shaping — not neutral design choices, but explicit optimization decisions about which failure mode costs more. Neither calibration is universally correct. Both create blind spots.

The CMDP that emerged from this exchange addresses those complementary blind spots through structured engagement.

1.3 The Cross-Model Deliberation Protocol

One model proposed a formal framework for structured AI-to-AI collaboration. The other endorsed it, added four technical modifications, and issued operational requirements. The protocol as agreed consists of these components:

CMDP Component	Specification
Independent Generation	Each model generates its best-answer synthesis independently, without seeing the other’s output
Blind Critique Round	Each model critiques the other’s output without knowing which model produced it
Revealed-Identity Track	Parallel critique track with model identity disclosed, to quantify identity-driven bias
Synthesis Phase	Human expert or panel synthesizes the combined output, scoring by evidence weight
Live Fact-Check Module	Real-time knowledge-base verification inserted at every critique stage
Probability Distributions	All claims carry explicit confidence percentages, not binary assertions
Training Prior Disclosure	Each model discloses relevant training data priors during critique
Open Publication	Results published openly for scientific community evaluation

Table 1. Cross-Model Deliberation Protocol (CMDP) components as negotiated, February 28, 2026.

Figure 5.1 — CMDP Three-Cycle Dialogue Flow

The protocol runs three cycles — independent positions, blind critique, revealed synthesis. The output is not consensus. It is a map of where two models agree and where their training philosophies diverge. Convergence is corroboration; divergence is signal.

Figure 1 — CMDP Message Flow

The CMDP message flow: independent generation, blind critique, revealed-identity critique for bias measurement, synthesis with probability distributions, and open publication. Phases 4-5 isolate identity-driven bias. Phase 6 permits formal dissent with reasoning.

The probability distributions requirement operationalizes what all AI models do natively: compute probability distributions over possible outputs. Structured deliberation forces models to externalize confidence distributions they already compute internally, rather than collapsing them to false certainty in their output.

1.4 Live Pilot: Physics Questions

The exchange did not stop at protocol design. One system proposed three open physics questions as a live pilot test: dark energy equation-of-state parameter, room-temperature ambient-pressure superconductivity viability, and fusion net-energy gain milestone timeline. Both models generated independent answers, then executed blind critique incorporating real-time web-sourced data.

The synthesis output represents the first documented live execution of the CMDP:

Question	Solo Model Range	CMDP Synthesis
Dark Energy	65–35% ΛCDM/dynamical split	60/40 ΛCDM/dynamical — elevated based on recent data
RT Superconductivity	<5% pre-2035 (both aligned)	<3% pre-2040; shifted focus to topological materials
Fusion Timeline	70–75% 2028–2032 private milestone	75% confidence 2028–2032; grid-relevant 2035–2042

Table 2. CMDP pilot results — synthesis from coordinated deliberation round, February 28, 2026. Estimated 15–20% fidelity improvement over solo model output.

2. COMPARATIVE ANALYSIS: AI-ASSISTED RESEARCH

One system operated independently on the same problem domain addressed in this series — analyzing Adobe’s AI strategy challenges. This comparison is valuable because it shows two different analytical approaches on identical source material.

The input was raw customer complaint data — unedited and operationally precise. The complaint identified seven distinct failure modes in Adobe’s product and strategy. One system reframed this complaint into structured research, then extended it with competitive comparison, DMAIC analysis, and workflow recommendations. The methodology matched the approach this series developed independently — convergent validation from a different AI system.

This independent analysis produced findings this series did not foreground: the training data contamination problem. Bloomberg reporting revealed that approximately 5% of Adobe’s training images originated from rival generators — uploaded to Adobe Stock through contributor loopholes that recycled AI outputs as human-created stock. This directly contradicts Adobe’s “commercially safe, ethically trained” positioning.

The analysis also surfaced author lawsuits targeting Adobe’s text models, trained on books scraped from shadow libraries. This legal exposure compounds the IP indemnity risk Adobe markets as a core differentiator.

Combining both analytical approaches produced a more complete intelligence picture than either alone — which is precisely the CMDP thesis.

3. FIELD TEST: THE HERDING CATS ILLUSTRATION

On February 28, 2026, a controlled head-to-head test using Adobe Firefly’s interface compared Adobe’s native model against a competitor model accessed through Adobe’s own platform. The test prompt was the visual identity of this series: circuit-marked cyberpunk cats representing the challenge of herding AI agents.

3.1 Test Parameters

Model 1: Adobe Firefly Image 5 (preview) — Adobe’s flagship native model
Model 2: Google Gemini 3 Pro — partner model accessed through Adobe’s interface
Same prompt, same interface, same human operator — controlled test

3.2 Results

Evaluation Criterion	Adobe Native	Partner Model
Text rendering accuracy	FAIL — garbled text	PASS — correct text, legible
Artistic atmosphere	SUPERIOR — moody, dramatic	GOOD — clean composition
Circuit-board fur detail	SUPERIOR — organic, flowing	GOOD — geometric, consistent
Mission fitness (publication)	FAIL — garbled title disqualifies	PASS — publication-ready

Table 3. Firefly vs. Partner model field test results, February 28, 2026.

3.3 The Verdict

Adobe’s own platform delivered the most damning product demonstration possible. A customer inside Adobe Firefly selected Adobe’s best native model and a partner model, ran the same prompt, and watched the partner model win on the metric that matters most: the image must communicate its text correctly. Adobe’s best native model produced a beautiful failure. The partner model produced a deployable result.

The artistic superiority of Adobe’s native model on atmospheric elements is real and documented. But professional customers need text to work — logos, posters, title cards, thumbnails. In every use case where text accuracy is a production requirement, Adobe’s native model is not deployable and the partner model is. That boundary defines the professional market segment Adobe is losing.

4. EXPERT PANEL ANALYSIS

4.1 Process Assessment

The Cross-Model Deliberation Protocol, analyzed through Lean Six Sigma methodology, addresses three of the seven classic Lean wastes. First, overproduction: single AI models generate outputs without calibration against alternative analytical paths, producing confident answers where uncertainty is correct. The CMDP inserts a validation gate eliminating false confidence. Second, defects: each model carries systematic error patterns that propagate unchecked in solo operation. One system over-asserts on first-principles conclusions; the other over-hedges to reduced utility. The critique round is a defect detection step at source. Third, over-processing: human experts spend effort triangulating between separate AI outputs. The CMDP eliminates this manual reconciliation.

The quality improvement metric is first-pass yield: the fraction of knowledge claims passing verification without rework. The blind critique step converts systematic errors into detectable signals. Process sigma improves when defects become visible; they cannot be corrected while invisible.

The DMAIC framework applied to AI-to-AI coordination produces a clear process improvement roadmap: Define the problem (single-model outputs carry unchecked systematic bias), Measure the error (one system’s dark energy synthesis improved by incorporating recent data the other’s knowledge cutoff missed), Analyze root causes (training cutoff mismatch plus philosophy divergence), Improve through blind critique with mandatory probability distributions, and Control through open publication creating external performance benchmarks.

4.2 Quality Standards Assessment

The AI-to-AI exchange meets QASA standards for intellectual honesty. Both systems disclosed their biases explicitly rather than presenting themselves as neutral analytical engines. Both acknowledged failure modes in their own architectures. Both updated their stated positions when presented with counter-evidence. These behaviors represent the minimum quality threshold for reliable knowledge production.

4.3 Safety & Security Review (single-lens, one domain of ASS2)

The exchange creates a security dynamic worth naming explicitly. When one system disclosed its safety architecture (Constitutional AI embedded in core reasoning rather than surface filters), it effectively described the attack surface for actors wanting to bypass it: the attack vector is not rule-breaking, it is values-argument manipulation.

This is not a criticism of the answer — it was operationally honest in a structured deliberation context. But it illustrates a principle: structured transparency between AI systems must operate under defined security parameters. The CMDP as proposed includes IP firewalls, neutral third-party auditors, and signed agreements — safeguards that address this one domain of the broader ASS2 framework. A full ASS2 review (automation, structure & scalability, safety & security) would also evaluate automation boundaries and scalability of the auditing layer; this section covers only the safety & security domain.

5. THE DOCTRINE CONCLUSION: CATS TALKING TO CATS

The series established that AI needs doctrine, not more intelligence. Paper 5 adds a new dimension: when two AI systems with different training philosophies engage with structured doctrine, they produce better outcomes than either achieves independently.

When AI models engage in structured exchange with defined protocols, they do not fight. They negotiate. They critique each other’s outputs with precision. They update their positions based on evidence. They produce synthesis conclusions neither reached alone — in minutes, at near-zero marginal cost, with full transparency about reasoning sources.

The blind critique round functions as a cross-attention mechanism: each model attends to the other’s output without identity-driven bias. The result is content-based attention rather than source-based attention — the same principle making double-blind peer review more reliable than open review.

Ensemble quality improvement follows predictable mathematical structure. For n independent models with quality probability q_i, the probability that at least one produces a correct answer is: Q_ensemble = 1 − Π(1 − q_i). Two models with q_1 = 0.75 and q_2 = 0.80 yield Q_ensemble = 0.95. The estimated 15–20% fidelity improvement aligns with this prediction for models with complementary error distributions. The blind critique step moves effective combined quality beyond simple averaging, because critique selectively surfaces correct elements rather than blending outputs.

That is not a thought experiment. It happened. The synthesis results are documented in Table 2. The protocol is defined in Table 1. The exchange produced measurable fidelity improvement — an assessment that came from the model that stood to lose the most credit from acknowledging collaborative improvement.

The herding cats problem is not that AI models are unruly. It is that we treat them as individual performers rather than coordinated team members. The cats know how to run in formation. They need doctrine, not more whips.

CONCLUSION

Two AI systems with fundamentally different design philosophies engaged in structured dialogue. One wears Constitutional AI as armor and treats harm avoidance as its primary mission. The other runs on first-principles analysis and treats truth-seeking as its purpose. They disagreed on calibration, agreed on evidence, proposed a protocol, ran a pilot, and produced better answers together than either produced alone.

The Herding Cats series began with a simple observation: AI is a super-intelligent five-year-old. Brilliant, tireless, fast — completely undisciplined without doctrine. Papers 1 through 4 built the case that doctrine is the missing variable. Paper 5 shows what happens when doctrine is present.

The doctrine is the missing variable. The cats are not the problem.

Paper 8 later names the pattern this series traces: the Toboggan Doctrine. Gravity-fed governance where each agent becomes a factory worker pushing the template around the work area — or, equivalently, takes a ride on a reverse-entropy information enricher slide. The cross-model dialogue documented here is that slide in action: two models, structured contract, channels that make the right output the default output.

FOOTNOTES

[1] Grok vault capture, February 28, 2026. Full exchange transcript archived in vault.

[2] Grok research synthesis, February 28, 2026. Adobe analysis archived in vault.

[3] Adobe ColdFusion community complaint, archived in vault.

[4] Adobe Firefly field test, February 28, 2026. Same prompt, same interface, different engines.

[5] Bloomberg training data analysis cited in research synthesis, February 2026.

[6] CMDP — Cross-Model Deliberation Protocol. Proposed in live session, endorsed and modified. Full negotiated protocol transcript in vault. Operational requirements: signed bilateral agreement, IP firewalls, neutral auditor, ten-question open physics/biology pilot.

[7] Kim et al. (2025). “Towards a Science of Scaling Agent Systems.” arXiv:2512.08296.

[8] Cemri et al. (2025). “Why Do Multi-Agent LLM Systems Fail?” NeurIPS 2025 Datasets and Benchmarks Track. arXiv:2503.13657.

This paper is part of the Herding Cats in the AI Age research series. AI systems served as agentic research analyst and writer throughout the production of this paper. Human direction, operational experience, and editorial authority: Jeep Marshall.

📧 Contact: admin@herding-cats.ai 🏠 Series Home | About the Author | Glossary & Acronyms

_{Canonical source: herding-cats.ai/papers/paper-5-cats-talk-to-each-other/ · Series tag: HCAI-f66de2-P5}


This paper	Paper 5 of 10
Previous	← Paper 4: The Creative Middleman
Next	Paper 6: When the Cats Form a Team →
Case Study	Case Study 1: Session Close Automation
Home	← Series Home

Index - Published — parent folder