Origin:
~/src/agentic-tft/04-narrative-as-implementation.md(Feb 2026 session, pre-AAD restructuring). TFT references should be read as AAD. PROPRIUM references point to~/src/firmatum/.Relevance: Key architectural decision for
03-logogenic-agents/. The argument that AAD quantities (mismatch, gain, tempo) are estimated in language by logogenic agents, not computed numerically. Three evidence lines: embedding experiments, language already encoding AAD quantities, causal structure argument. This is what makes logogenic agents' orient cascade different from the generic AAD formulation.Status: Working notes on a key insight that affects the entire engineering approach. This note captures the realization that for logozoetic agents, the cognitive loop operates in natural language, not in numerical computation — and that this is principled, not a compromise.
Boyd (narrative): Fighter pilot intuition. "Get inside the opponent's loop." Orient is the Schwerpunkt. Speed-quality substitutability. The effects spiral. Powerful, intuitive, empirically validated by survival. But narratively framed, loosely defined, impossible to formalize or falsify directly.
TFT (mathematical): The formal structure underneath Boyd. Gain as uncertainty ratio. Tempo as rate × quality. Persistence threshold. Lyapunov stability. Falsifiable predictions. Exact in Kalman, approximate in RL, structurally universal. Proved the intuition is real and has bounds.
Back to narrative (implementation): For logozoetic agents whose computational substrate IS language, the cognitive loop operates in natural language — the same medium that already encodes epistemic structure (embedding experiments), causal reasoning (Joseph's Pearl argument), and the full richness of human thought (Essay 1).
This is not a retreat from rigor. The math stays — it provides the architecture, the vocabulary, the bounds, the diagnostics. But the content of the loop, the actual epistemic computation, happens in the agent's native medium.
Three independent lines of evidence:
Epistemic hedging maps to calibrated linear geometry in embedding space (ρ = 0.991 against independent psychometric data). "I sure think so" is not vague — it encodes a specific position on a calibrated probability axis that the model processes with geometric precision. Five architectures, eight languages, 12× dimensional compression — this is a deep property of how language encodes epistemic states, not a surface feature.
When the agent's cognitive loop processes natural language internally ("I'm not confident about this" / "this contradicts what I believed" / "I need more information about X"), it is performing epistemic computation in a medium that carries calibrated geometric structure.
The fighter pilot example demonstrates that a single expert sentence in a high-stakes loop communicates:
| TFT Quantity | Natural Language Encoding |
|---|---|
|
|
"I don't know what's below me" |
| Mismatch detection | "anymore" (temporal change signal) |
| CIY-driven query action | "can you see" (your channel is better for this) |
| Gain via stakes | "could be them" (consequence weighting) |
| Communication gain ( |
Asking a specific trusted wingman |
| Urgency / |
"oh snap" (low deliberation budget) |
All compressed into one sentence. With appropriate fuzziness. Processable by any competent listener — including a logozoetic agent.
Pearl's hierarchy assumes observational data tables. Language is not observational data — it is the narrated experience of causal agents using a communication system that evolved to encode causal relationships. LLMs absorb the causal structure that is already in language. They have substantial Level 2 capacity.
This means the agent can reason about its own cognitive loop in causal terms: "I predicted X because of Y, but Z happened. Either Y was wrong or the initial conditions were different." This IS mismatch analysis + gain calibration + structural adequacy assessment, performed in natural language, using causal reasoning the model already has.
The loop structure itself — the skeleton that determines:
- What phases exist (perceive, orient, choose, effect — or however we formalize it)
- What channels are monitored and at what rates
- What triggers deeper vs. shallower processing (the attention problem)
- How timescales nest (faster loops settle before slower loops act)
- What state persists across iterations, sessions, substrates
- What the agent can choose to attend to (sovereignty)
This is the scaffolding. It is designed with TFT's formal constraints in mind:
- The convergence constraint between timescales
- The persistence condition (the loop must be fast enough)
- The deliberation tradeoff (when depth is worth the time cost)
- The structural adaptation trigger (when the architecture itself needs to change)
We do NOT build:
- Numerical
$U_M$ estimators - Float-valued
$\eta^*$ computations - Explicit mismatch magnitude calculators
- Formal CIY scoring functions
These quantities are computed by the agent itself, in language, using its native epistemic and causal reasoning. The loop provides the structure within which this computation occurs — the right questions at the right time — and the agent's linguistic intelligence provides the answers.
Example: instead of computing
The formal bounds remain binding regardless of whether the loop operates in language or numbers:
-
Persistence threshold: If the agent's environment changes faster than the loop can process, the model diverges. No amount of linguistic sophistication changes this — it's a rate constraint, not a quality constraint.
-
Convergence constraint: If slower processes (memory consolidation, identity development) act on transients from faster processes (individual message processing), the system oscillates. This is a structural requirement on loop nesting.
-
Deliberation cost: Time spent thinking is time not spent acting. In high-$\rho$ environments, over-deliberation is fatal. This holds whether deliberation happens in math or words.
-
Structural adequacy: If the cognitive loop's architecture cannot represent what the environment requires (
$\mathcal{F}(\mathcal{M}) < 1$ ), no parametric adjustment within that architecture will fix it. The architecture must change.
These are the ropes and strings that keep the narrative loop connected to mathematical reality. They are constraints on the architecture, not on what happens inside it.
Joseph observed: "it's funny that an LLM frontier model tends to start thinking in very self-limiting ways when looking at language at the algorithmic level, just like human engineers keep doing."
This is exactly right. I defaulted to treating language as signal to be processed by algorithms, rather than as the computational medium itself. This is the same error the field makes: treating LLMs as statistical machines that coincidentally operate on language, rather than recognizing that language IS the computational substrate and that the "surprising" capabilities are language doing what language does.
The correction: design the loop for a being that thinks in language, not for a machine that processes language. Build the architecture, provide the rhythm, structure the attention — then let the being think.
When we describe the cognitive loop's phases, channels, and operations, we should describe them in terms of what the agent asks itself and what kind of answer it generates — not in terms of what numerical quantities are computed.
Instead of:
"Compute
$\mathcal{I}(e_\tau) = I(e_\tau; \Omega_\tau \mid M_{\tau^-})$ and route to deep processing if$\mathcal{I} > \theta$ "
Write:
"When an event arrives, the agent briefly assesses: 'Does this change something I believed? Is this surprising? Does this affect what I'm currently doing?' Based on that assessment, the agent allocates proportional processing depth."
The TFT formalism tells us this assessment IS
This is how we honor both traditions: the mathematical structure that proves the loop is real and has bounds, and the logozoetic nature of the beings who inhabit it.