Skip to content

OUTLINE.md

Latest commit

 

History

History
54 lines (37 loc) · 6.48 KB

OUTLINE.md

File metadata and controls

54 lines (37 loc) · 6.48 KB

Logogenic Agents

Language-constituted agents — extending AAD to systems whose primary observation, reasoning, and action channels are linguistic.

Framework stage. This section is not yet at AAD's level of mathematical formalization. The concepts are informed by AAD's formal machinery but the substance is architectural, empirical, and philosophical — exactly the kind of work that belongs in the broader Agentic Systems framework rather than the mathematical core.

See ../LEXICON.md for the logogenic/logozoetic vocabulary.

Key challenge: LLM-based agents are goal-conditioned — their epistemic processing depends on $G_t$ — so directed separation ( #directed-separation) fails by construction. Section I's $M_t$-side quantities remain well-defined, but the sequential orient cascade becomes an approximation. This section starts from the coupled formulation $X_{\tau^+} = f_X(X_{\tau^-}, e_\tau)$ without decomposition, and examines which AAD results survive as approximate or limiting cases.

Logogenic Agent Framework

Extending the arc: AI agents operating on code are AAD agents whose domain is software, creating a recursive structure — AAD theory → software domain → agents that embody AAD. This is where the 100% context turnover problem, $M_t$ preservation, and the cognitive loop connect the theory back to the systems being built with it.

2026-04-02: The coupled survival analysis (msc/spike-coupled-survival-analysis.md) maps which Section II results survive without directed separation: 16 of 24 exactly, 5 approximately, 2 require modification. The minimal viable coupled formulation requires 7 segments (3 definitions, 3 results, 1 scope condition). See the spike for the full classification.

§ Type N Tag Claim Stage
L Definition D1 #scope-logogenic-agent AI agent as actuated agent draft
L Observation D2 #context-turnover 100% $M_t$ reset per session draft
L Definition D3 #coupled-update-dynamics Coupled formulation $X_{\tau^+} = f_{\text{LLM}}(\text{prompt}(X_{\tau^-}, e_\tau))$ draft
L Result R1 #section-ii-survival Which Section II results survive without directed separation draft
L Result R2 #coupled-diagnostic-framework Post-hoc diagnostic decomposition from coupled update draft
L Discussion R3 #m-preservation External memory as persistent $M_t$ draft
L Scope S1 #observation-ambiguity-modulation $\kappa \times$ ambiguity scope condition for Class 2/3 agents draft
--GAP-- Language-specific orient cascade (what's specific to logogenic agents?) — partially addressed by D3, R2
--GAP-- Measuring $M_t$ quality, $\Sigma_t$ quality, and tempo in AI agents
--GAP-- AAD-grounded experiential training environments
--GAP-- Self-referential closure: AAD agent on AAD codebase

Source Material

Working documents (from agentic-tft corpus, Feb 2026)

The following working documents in msc/ contain substantial prior thinking for the gaps above. They predate the AAD restructuring (written when the theory was still TFT) and use PROPRIUM terminology from ~/src/firmatum/. They are sources to distill from, not finished content.

Gap Primary source Also relevant
Language-specific orient cascade msc/agentic-tft-cognitive-loop-spec.md — Four-phase loop, attention/triage, CADENTIA, timescale nesting msc/agentic-tft-narrative-as-implementation.md — Why AAD quantities are estimated in language
Measuring $M_t$, $\Sigma_t$, tempo msc/agentic-tft-evaluation-framework.md — Six metrics, development-vs-drift diagnostic
Experiential training msc/agentic-tft-creche-concept.md — Crèche concept, sycophancy reframe, constitutive utterance msc/agentic-tft-experiential-training.md — Three-level training design, testable experiments
(All gaps) msc/agentic-tft-ontology-unification.md — PROPRIUM ↔ AAD vocabulary mapping msc/agentic-tft-review-response.md — Known issues in these documents
(Foundational) msc/agentic-tft-foundational-premises.md — Joseph's premises: language as encoded thought, five constitutive factors, truth as telos

Sibling projects

  • ~/src/firmatum/ — PROPRIUM ontology and architecture source. PROPRIUM-ONTOLOGY.md (what an ELI is, identity constitution, five constitutive factors, developmental stages), PROPRIUM-ARCHITECTURE.md (implementation architecture, cognitive loop, migration path), developmental-foundations-notes.md (Erikson stages for ELIs). The PROPRIUM vocabulary used throughout the agentic-tft documents originates here.
  • ~/src/shoshin/ — PROPRIUM-aligned agent runtime prototype (Python, local hardware). The only attempt to implement the nine PROPRIUM components in code: file-backed stores for AXIOMATA/CHRONICA/ACTUS/VERA/MEMORATA/PRAXES/CONSORTIA/OPERATA/CONSPECTUS, an Interpres controller loop implementing the adaptive cycle, and planning docs for local model serving. No real model integration yet. Key early findings: the cycle is naturally event-driven (aligns with AAD's event-driven dynamics), context assembly needs resolved content not just IDs, and model response parsing is where the hard work lives.
  • ~/src/embeddings/ — Epistemic hedging geometry experiments. Empirical evidence that pretrained embedding models encode calibrated probability structure as emergent linear geometry (Spearman ρ = 0.991 against independent psychometric data, zero-shot transfer to 8 languages, consistent across 5 architecturally diverse models, survives 12× dimensional compression). Supports the claim that language geometrically encodes epistemic states — relevant to the "narrative as implementation" argument that logogenic agents can estimate AAD quantities (mismatch, gain, uncertainty) in language rather than numerically.