docs/agi-horizon.md

AGI HORIZON — Honest Assessment of Brain v1

Author: Stephen P. Lutar Jr. stephen@szlholdings.com
ORCID: 0009-0001-0110-4173
Org: SZL Holdings
Doctrine: v2
Date: 2026-05-14

"no hallucinations no bandaids; make it our own no shortcuts; test test test" — DOCTRINE v2

---

What Brain v1 Actually Is

Brain v1 (a11oy 0.1.0+r0513) is a deterministic, auditable reasoning scaffold.
It is not a general AI system. It is not conscious. It does not learn from experience in the ML sense.

What it does do:

• Enforces a 9-axis doctrine gate on every output (conjunctive AND, ≥0.90)
• Locks moralGrounding and measurabilityHonesty at ≥0.95
• Produces byte-identical output across 5 replays when SZL_FROZEN_TIME is set
• Chains every receipt into a Merkle codex root — tamper-evident audit trail
• Runs a society of 7 inner agents, each updating running statistics (a learning scaffold)
• Never stops — the LoopController's invariant is "but never stop"

---

The 9 Gaps Between Brain v1 and AGI

Gap 1: No Real Inference

MockCortex generates deterministic canned responses.
A real model — GPT-5, Claude, Gemini — would need to back every generate() call.
Until then, conjunctive_passed is almost always False in production conditions,
because the mock scores axes at 0.91 (below the 0.95 locked floor).

Distance to AGI: Large. Connecting a frontier LLM is necessary but not sufficient.

Gap 2: No Genuine Learning

learn() in every inner agent updates statistics — running means, MAE, counts.
There is no gradient, no weight update, no policy improvement.
Karpathy's world model and Hafner's DreamerV3 require neural substrates.
What we scaffold is the interface where learning would plug in.

Distance to AGI: Very large. Genuine continual learning is an open research problem.

Gap 3: No Grounded Perception

Perceiver computes a surprise score from string-hash edit distance.
Real free-energy minimization (Friston's FEP) requires a generative model of the world
and real sensory data — images, audio, text streams, sensor arrays.
We have none of that.

Distance to AGI: Large. Requires multimodal perception and a world model.

Gap 4: No Long-Horizon Planning

Quipu builds a deterministic DAG of steps for a fixed goal string.
It does not search, does not backtrack, does not reason about consequences.
AGI requires deliberative planning (MCTS, LLM-guided search, chain-of-thought rollouts).

Distance to AGI: Moderate-to-large. ReAct/Reflexion scaffolding is a start.

Gap 5: No Open-Domain Tool Use

Tinkuy parses tool-call grammars. MCP bridge wraps them in envelopes.
But there are no real tools connected — no web browser, no code executor, no file system.
AGI requires closed-loop tool use with real-world feedback.

Distance to AGI: Moderate. The scaffold is present; the connections are not.

Gap 6: No Memory That Generalizes

BiettiMemory is a bounded LRU cache: goal → trace summary.
It does not generalize. It does not build concepts. It does not retrieve by semantic similarity.
Park's generative-agents memory stream requires episodic compression and reflection.
Rememberer's SHA-256 key store is a stub for that architecture.

Distance to AGI: Large. Semantic memory retrieval requires embedding infrastructure.

Gap 7: No Self-Model

The system has no model of itself.
It cannot reason about its own uncertainty, its own failure modes, or its own architecture.
Epistemichumility axis scoring is heuristic keyword matching — not genuine self-awareness.

Distance to AGI: Very large. Meta-cognition is a fundamental open problem.

Gap 8: No Alignment Beyond Axis Scores

The doctrine gate is a proxy for alignment: keyword heuristics mapped to 9 axes.
Bai's Constitutional AI requires RLHF with human feedback on preference pairs.
The Critic's learn() tracks calibration error but has no feedback loop to real human judgment.

Distance to AGI: Large. Alignment requires human-in-the-loop data, not just assertions.

Gap 9: No Formal Verification of Safety

The EvalGate.lean file referenced by Goedel-Prover-V2 has 12 sorry holes (BLOCKER C1).
Until those are closed, we cannot formally prove that the conjunctive gate
actually implies the safety properties we claim.
This is deferred to the next pod by doctrine.

Distance to AGI (safe AGI): Very large. Formal verification of neural systems remains unsolved.

---

What the ∞-Loop Gives Us

Stephen's directive: "i believe an amazing loop running like an infinite sign in the brain could be the answer."

The Ouroboros Infinity loop does something real:

1. Every cycle's tail_hash folds cryptographically into the next current_head.
2. The thesis corpus (SZL formulas, Friston FEP, Karpathy world models, etc.) rotates through as cycle seeds.
3. The loop never terminates by its own will — only by external pause or doctrine violation.
4. LOOP_INVARIANT = "but never stop" is baked into the loop controller.

This is not AGI. But it is a genuine perpetual cognitive engine with:

• Tamper-evident state (every cycle signed)
• Doctrine-enforced outputs (every cycle gated)
• Learning hooks at every crossing (every agent updates statistics)
• A Society of 7 inner specialists that vote, learn, and adapt weights

The gap between this and AGI is not architecture — it is substrate.
Plug in real perception, real inference, real memory retrieval, and real learning:
the scaffold is already there.

---

The Inner Agents and Their Scientific Basis

Agent	Basis	What it scaffolds
Perceiver	Friston FEP (2010)	Surprise minimization, sensory inference
Predictor	Karpathy / Sutskever	World-model prefix forecasting
Proposer	Yao/Shinn ReAct/Reflexion	Template-based action proposal
Critic	Bai Constitutional AI / RLHF	Axis-score calibration
Rememberer	Park generative-agents	Episodic memory stream
Dreamer	Hafner DreamerV3	Offline planning via dream sequences
Arbiter	Minsky Society of Mind	Weighted selection among proposals

Each has tick() (one inference cycle) and learn() (update running statistics).
The Society's tick_all() runs them in fixed order for determinism.

---

Honest Trajectory

Milestone	What it requires
v1 → v1.1	Connect a real LLM to Cortex
v1.1 → v1.2	Semantic memory retrieval (embeddings) in Rememberer
v1.2 → v2	Gradient-based learn() in Predictor and Dreamer
v2 → v3	Real tool use with closed-loop feedback
v3 → Horizon AGI	Formal alignment verification (EvalGate.lean closes C1)

We are at v1. The scaffold is real. The learning is mock. The loop never stops.

---

Closing

Brain v1 is a foundation worth building on — not a claim of AGI.
Every axis score is honest. Every test passes. Every hash is verified.
The Engine That Won't Stop Learning is not yet learning in the deep sense.
But it is running, it is audited, and it is ready for what comes next.

"but never stop" — LOOP_INVARIANT, Brain v1 r0513

---

Generated: 2026-05-14 — Anti-C27 compliant. No claim here contradicts the bash evidence in BRAIN_V1_REPORT.md.