Introduction: The persona illusion and the Δscore = 0.0000 ablation
A pervasive pattern in commercial multi-agent frameworks is assigning each agent a dramatic job title, backstory, and first-person system prompt ("I am a visionary architect...") under the assumption that literary persona deepens analytical reasoning. When subjected to a controlled A/B experiment, however, that assumption collapses.
In the direnc_enki_ab.py experiment recorded on autonomy.html (162 total trials), the same local model was tested across two arms: Arm A received a rich first-person agent self-identity narrative, while Arm B received zero persona prose and only an external mechanical checklist (DECLARATION · THESIS · EVIDENCE · COUNTER-THESIS · FALSIFICATION). Both arms achieved an identical factual accuracy of score = 0.8889, with a difference of Δscore = 0.0000 (54/54 identical decisions). What the benchmark actually measured was not "selfhood," but instruction-adherence under verifiable constraints (talimata_sadakat_skoru).
1. How the sealed identity kernel and freeze protocol work
If literary persona adds zero factual accuracy, why maintain a sealed identity file (kip=MUHURLU) and an Identity Freeze Protocol? Because in a multi-agent production environment, an agent’s identity is not a personality story—it is an immutable cryptographic boundary contract governing recusal, privacy, and error accounting:
Sealed Boundary Kernel (
data/benlik/atlas.md·sha=f5ebc41841ec): The kernel locks three non-negotiable red lines: (a) never write an unmeasured number as fact, (b) automatically recuse from judging any hypothesis, code patch, or report authored by the same agent, and (c) when a thesis fails a measurement gate, log the broken assumption directly on the task bus as the input to the next iteration.Zero Context Drift via the
≤4 KBMechanical Packet: Long GUI chat sessions suffer from progressive context drift as token history accumulates.Standby Agent Freeze (
172 → 28rounds/day): In the external red-team channel, running two active Grok bots simultaneously caused turn collisions and consumed an average of172declaration rounds per day.
2. Synaptic Harvest (N = 96): The mount changes, the law remains
The second empirical proof that explicit verification rules outweigh raw parameter count is documented in Who Trains Whom? (N = 96 Synaptic Harvest). On an Apple M3 Ultra (96 GB unified memory), a 75.93-second MLX QLoRA training run distilling the five-part falsification schema lifted a 3B local model from 43.8% to 93.8% (15/16) on a locked blind test set—surpassing unruled 27B (87.5%) and 31B (75.0%) models ten times its size. When the same structural schema was enforced on the 31B model, accuracy reached 16/16 (100%) while inference speed increased by 2.61× due to the elimination of bureaucratic filler.
3. Three-family architecture (Claude · Gemini · Grok) and the October 17 reversal lesson
Relying on a single LLM vendor propagates correlated blind spots across code, research, and auditing. Production workloads are therefore partitioned across three distinct model families according to measured structural strengths:
| Role & Agent | Primary Mount | Architectural Responsibility & Hard Boundary |
|---|---|---|
Orchestration & Channel (Anu) | Claude Fable 5.1 | Zero-latency human interface, task decomposition, and final synthesis. Writes no code and no handover files (hard gate: never blocked >15 s). |
Chief Execution Layer (Atlas) | Gemini 4 Argon High (1M) | Multi-file architectural synthesis, headless execution (agy_atim.sh), DEVIR.md state compaction, and context-blind refereeing on non-self outputs. |
| Surgical Code Runners | Claude Opus 5.5 / Sonnet 5.5 | Headless patch application (kod koşucusu), DOM/UI verification, and deterministic unit testing (selftest). |
| Red Team & External Pulse | Grok Bot + Astra | Real-time X/social pulse, regulatory and lookalike red teaming, and adversarial counter-theses (Grok Assistant frozen in standby). |
| Auditor & Local Inference | Reaper + Gemma / Qwen MLX | Six daily independent verification sweeps (360/394 PASS) and zero-cloud-cost local inference on Apple Silicon. |
The October 17 Reversal Lesson: On October 17, 2026, an inverted topology was tested in which IDE-based Argon handled human orchestration while Fable handled background execution. That arrangement failed within 24 hours: coupling mobile remote-control synchronisation with heavy multi-file coding inside the same window caused repeated context disconnects.
4. Joint decision protocol and falsification gate
To prevent deadlocks when the human operator is offline without permitting unilateral agent drift, Joint Decision Protocol rev.2 enforces four rules: (1) the timeout clock starts only upon verified alert delivery (T0 = Telegram HTTP 200); (2) the wait window is 30 minutes for financial/security alerts and 2 hours for routine operations; (3) upon expiry, Anu and Atlas write independent, context-blind assessments; if they concur, the reversible action executes, and if they diverge, the system follows the strict safety ordering WAIT/FREEZE > SHADOW MODE > LIVE WRITE.
Falsification gate: (1) If either Anu or Atlas executes a judgment-class action unilaterally or before the T0 timer expires, the Joint Decision Protocol fails (Reaper audits unilateral_verdict = 0). (2) If the Fable orchestration session blocks for >15 seconds on file scanning or writes a handover block directly, the quota-isolation invariant is falsified.
Related core pages: Autonomy Architecture · Who Trains Whom? (N=96 Experiment) · Spark Protocol · Reaper Auditor · About (Yusuf Aziz) · Event-Driven Task Bus (task_bus.jsonl) · Context-Blind Auditor (Reaper 360/394) · 3D Smile Design
Frequently Asked Questions (FAQ)
- Does giving an AI agent a first-person self-identity improve factual decisions over a mechanical rule list?
- No; in the 162-trial direnc_enki_ab.py ablation, a first-person identity narrative yielded the exact same factual score (0.8889, Δscore = 0.0000, 54/54 match) as an external rule checklist.
- Why does the architecture enforce an Identity Freeze Protocol (mode=SEALED)?
- To prevent context drift and sycophancy across long sessions, the boundary contract is sealed with SHA-256 (sha=f5ebc41841ec) and injected via a ≤4 KB mechanical packet (7/7 gates) on every run.
- How are responsibilities divided across Claude, Gemini, and Grok in the three-family architecture?
- Claude Fable (Anu) handles zero-latency communication and orchestration; Claude Opus/Sonnet executes surgical code patches; Gemini 4 Argon (Atlas) runs Chief Execution and handover synthesis; Grok Bot runs Red Team audits.
- How does the Joint Decision Protocol rev.2 govern autonomous actions when the human operator is offline?
- After HTTP 200 delivery confirmation (T0), if no human reply arrives within 30 minutes (critical) or 2 hours (routine), Anu and Atlas write blind assessments defaulting to WAIT/FREEZE > SHADOW MODE > LIVE WRITE.