Conceptual Map — How the Current Project Connects¶

Status: Reader map — navigation, not theory.

Scope: A compact map after the foundations audit. Every load-bearing term below resolves to a home file, an instrument, or an open problem.

The Shared Root¶

Given a trace, declared evidence, and a model family, which candidate process models remain — and what observation or intervention could distinguish them?

Forward and inverse costs depend on representation, evidence, target equivalence, and algorithm. The inverse-reconstruction benchmark measures selected finite regimes; the Foundations Reconstruction states the formal limits.

From that root, the repository asks two different questions.

The Two Spines¶

Spine	Question	Route	Current status
Epistemic Loop	Can a system construct and revise a useful world-model?	Trace → Candidate Model → Construction → World-Coupling → Intervention → Revision → Self-Binding	selected phases measured in toy systems; composition into one loop hypothesized; experience out of scope
Viability Arc	Can a system remain viable and correctable while optimizing?	Emergence → Optimization → Constraint Architecture → Survivability	conditional model results + synthetic ODE/ABM evidence; sufficiency and external applicability open

The Epistemic Loop and Canonical Path v2.1 are navigation folds, not additional theories. They touch at world-coupling: better contact with reality improves capability, but does not supply values or safety constraints.

The Layers¶

Layer	Central move	Main homes	Instrument / boundary
Forward process	declared rules unfold into traces and global organization	Emergence Manifesto, simulation models	selected toy processes; no claim that all intelligence reduces to one formalism
Inverse reconstruction	traces constrain a class of typed generators or candidate process models	Foundations Reconstruction, Generator Question	benchmark v0–v1.13
Construction and coupling	candidates are built, run, or turned into distinguishing queries and refereed by a world	Construction vs. Deduction, Witness Principle, World Models & VLA	family-search curves; coverage–distinction lemma and exact witness frontier; runtime/provenance constraints
Intervention and revision	distinguishing queries collapse the class; execution updates the model	Measurement as Weak Intervention	watching < perturbing < preparing; closed-loop revision
Situated agent boundary	capability is evaluated over a declared process–body–interface–environment–memory coupling, constraints, history, tasks, lenses, and tests	The Agent Is Not Where the Model Ends, Competence, Constraint, and Verification, Embodiment and Goal Decomposition	situated-stack and constraint-release benchmarks; no universal boundary, scalar intelligence law, or invariant-factorization theorem
Identity boundary	persistence is tested under perturbation and commitment	Invariance & Identity, Chord vs. Arpeggio	Exp5–7; real models and optimized mimics open
Constraint architecture	capability loads concentration, coherence, and substrate constraints	TEO, Viable Corridor	demonstrated in two synthetic models; real systems open
Consciousness boundary	functional self-binding is a research question at the loop's reflexive edge	Global Availability, Coherence Work	organizational hypotheses only; no inference to experience or present-system consciousness

What the Repository Has Actually Measured¶

Known-family inversion can be cheap. Full observability and clean data permit exact or low-error recovery.
Coverage limits this observation process. A selected rule-90 orbit leaves an eight-member consistent-model class in the elementary-CA family.
Intervention can reduce the class. In the selected CA setup, passive watching plateaus while a prepared query identifies the rule in one step.
A distinguishing query can itself be constructed. In the exact width-eight ECA baseline, the best cost-four preparation leaves one rule while an equal-cost unstructured preparation leaves 5.71 rules on average.
Occam is world-dependent. It helps on simple worlds, is chance on uniform worlds, and is systematically wrong on complex worlds until coverage is nearly complete.
Unmarked uncertainty is exploitable. The optimizer's-curse wedge grows with class size and disappears when the planner represents the class honestly.
Optimization is not curiosity. In the closed loop, argmax planning does not collapse uncertainty faster than random exploration.
Binding has a functional signature in the current toys. Exp5–7 separate joint satisfaction at the commitment boundary from rotating or blended constraint use; IP and Δ-Kohärenz each have measured blind spots.
Controller identity does not fix selected capability in the situated-stack toy. One policy scores from 0 to 1 as sensors, actuators, body, environment, and goal interface vary; a coordinated mirror preserves physical behavior, and equal aggregate scores can hide different traces.
Constraint release and lens reinterpretation separate in one exact finite system. With the transition process and policy fixed, releasing one blocked edge changes the physical trace and competence score; changing only the lens leaves the trace fixed while changing the reported score.
Capability loading and single-axis insufficiency occur in-model. They have been reproduced in the TEO ODE and a structurally different agent-ecology ABM, not yet in real agent ecologies.
Coupling can expose family misspecification. In v1.8, data from a coupled process can leave the declared single-rule family empty; supplying the coupled family restores a fit.
Useful support is not automatically retained. v1.10 establishes a bounded, budget-matched redundancy benefit. v1.11 lets the population and realized links turn over: the evolved network helps in paired acute ablation, but paid contribution is selected downward in all 16 seeds and abundance is lower. Functional collective benefit and evolutionary stability separate.
Making contribution visible sorts the network without retaining the trait. In v1.12, four arms under one accounting — blind, partner choice, conditional reciprocity, assortment, with partner information explicitly paid for — leave the sign of support selection unreversed in a majority of seeds. Partner choice does something else instead: it excludes non-contributors from the network almost entirely (1.2% of linked agents against 22.0% blind) while they persist in the population (27.1%).
Relocating the cost nearly removes the penalty and falsified the preregistered prediction. v1.13 moves the cost of support from the donor to the local group. Selection against support falls from −0.1056 to −0.0075 (44% of seeds positive, up from 0%), and seeded cheaters drop to 2.6% against 21.6% — below their starting frequency, where the preregistration expected pooling to ease free-riding. Population size pays for it, and the result is not yet a win: a group-mean levy makes within-group variation nearly cost-neutral, so drift rather than advantage may be carrying the retained trait.

These results are receipts, not universal laws. Their scope is maintained in What This Project Does NOT Claim.

The Current Frontier¶

Learned witness construction — can a system approach the exact query frontier across unseen candidate subsets and access geometries without enumerating the query space again?
Learned searchers vs. exact family search — can LLMs or program synthesizers beat the enumeration floor under matched evidence and query budgets?
From useful support to evolutionarily stable support — v1.12 answered the first form of this question negatively: neither partner choice, conditional reciprocity, nor assortment reverses selection against contribution under matched costs. What remains is whether the retention v1.13 bought by relocating the cost is selected for or merely no longer selected against, which needs a within-group cost gradient that keeps the relocation while restoring a measurable differential — and whether network exclusion is a stable outcome or a reservoir that returns when the network weakens.
Real-model identity probes — do the Exp5–7 distinctions survive outside handcrafted toy agents?
Real-agent constraint architectures — do P7/P8 survive when budgets and capability are varied in actual agent ecologies?
Self-model causality and trace governance — do hidden-state perturbations distinguish self-report from metacognitive control, and do provenance/decay/reset policies improve recovery after memory regime shifts?
Constraint attribution and replay equivalence — can physical exposure be separated from observer reinterpretation beyond finite toys, and which distinctions should empirical replay preserve?
External review and calibration — does the Viable Corridor survive scrutiny beyond its own equations and synthetic models?

The maintained problem statements live in Open Problems. New concepts do not enter this map until they have a home, status, and instrument or explicit open problem in the Concept Registry.

Recommended Route¶

The repository is a research notebook: theory proposes, code constrains, fiction stress-tests consequences, and meta-governance keeps the claim surface below the evidence.