Synthesis Log

Running log of cross-domain connections. Each entry captures a moment when ideas from different domains connect, forming new insights. Accumulates across all research sessions.

2026-07-19 — Session 1

ANT topology ↔ ALife ontology

Latour's "nodes have as many dimensions as they have connections" is the anti-essentialist move ALife simulations need. Most simulations give agents fixed properties. In ANT, an entity's properties ARE its connections. Change the network, change the entity. This means a simulation should have no pre-defined agent types — agents emerge from their relational position.

Levin's cross-scale problem ↔ Open-ended evolution stall

Levin says "the problem of pattern and scale is the central problem in ecology." Emergent Garden observes every ALife simulation stalls out. Same problem: cross-scale interactions. Levin identifies it in nature, Emergent Garden in simulation. Neither has the formalism to bridge scales. ANT's translation could be that formalism.

Quasi-objects ↔ Resource modeling

Latour's quasi-objects (things that transform when they circulate) are a better model for resources in ALife than fixed-value "food pellets." In the water cascade, water molecules transform into droplets (movers are transformed), and droplets transform the landscape (moved object transforms environment). Resources should be co-determined with their carriers, not pre-defined.

Emergent Garden's spectrum ↔ Wolfram's irreducibility

The move from genetic algorithms to emergent ALife is a move from explicit to implicit design. Computational irreducibility says you can't predict implicit behavior from explicit rules. This is why open-ended evolution is the holy grail AND why it's hard — you must simulate to know, but the simulation must be rich enough to produce genuine novelty.

Callon's translation ↔ Phase transitions

Callon's four moments (problematization, interessement, enrollment, mobilization) could become computational primitives for phase transitions in ALife. When a cluster of actors reaches threshold interaction density, they undergo enrollment (form a new collective) and mobilization (act as one at a higher scale). This IS a phase transition — new actor, new rules, new scale.

2026-07-19 — Session 2 (Hofstadter / Autopoiesis loop)

Strange loops ↔ Multi-scale topology

Hofstadter's strange loops are level-crossing feedback loops — you move up through abstraction levels and arrive back at the start. This IS the topology of multi-scale systems. The water cascade loops back: molecules → droplets → clouds → floods → topography → (determines where molecules collect). It's not a clean stack of levels — it's tangled. A simulation of multi-scale composition needs tangled hierarchical structure, not a fixed stack.

Downward causation ↔ ALife environment

Hofstadter's radical claim: high-level emergent patterns exert causal potency over low-level components. The flood reshapes topography. The cloud determines water distribution. Standard ALife simulations don't model this — the environment is fixed, agents interact with it but don't reshape it at a different scale. Downward causation is the missing mechanism: emergent structures must be able to modify rules at their scale, affecting lower scales.

Autopoiesis ↔ Strange loops

Maturana & Varela's autopoietic system (network produces components that produce the network) IS a strange loop by definition. Self-production is self-reference through process. The network references itself through its own production cycle. Hofstadter and Maturana describe the same phenomenon from different angles — one cognitive, one biological.

Autopoiesis ↔ ANT actor persistence

Latour says actors are defined by relationships. An autopoietic system is one whose relationships are self-maintaining. For an emergent structure to persist as a new actor at a higher scale, it must be autopoietic — it must maintain the network that constitutes it. Autopoiesis is the condition for actor persistence across phase transitions.

Self-maintenance ≠ Complexification

The 1974 computational autopoiesis model maintains itself but doesn't evolve. Same stall as EvoLoop. Self-maintenance is necessary but not sufficient for open-ended evolution. Missing ingredient might be: interaction with OTHER autopoietic systems at the same scale, creating a higher-level network. Multi-scale autopoiesis (systems producing systems) might be the recipe for complexification.

Hofstadter's "I" ↔ ANT actor identity

Hofstadter's self is a "narrative fiction" — a pattern, not a substance, continuously rebuilt by experience. Latour's actor is defined by relationships, not intrinsic properties. When the network restructures, the actor's identity changes. A strange loop is what happens when an actor's network position includes a reference to itself — self-reference through topology.

Stigmergy across scales (Vance's insight)

Stigmergy manifests differently at different complexity levels: ant pheromones, human petroglyphs, stupas, books, and memes (Susan Blackmore extending Dawkins). As actors become more complex, the stigmergic traces become more persistent, more information-dense, more detached from spatial proximity, and — with memes — capable of their own evolutionary dynamics (propagation, mutation, selection). If we accept ANT's generalized symmetry, then rivers carving channels and salt cedar altering soil chemistry are ALSO stigmergic — environmental modification that coordinates future behavior. The mechanism is the same across scales; the actors differ. Memes add a new dimension: stigmergic traces that evolve. This connects stigmergy to both the dynamic environment hypothesis (H4), multi-scale composition, and potentially to the quasi-object concept — a meme is a quasi-object that transforms as it circulates.

2026-07-20 — Session 3 (Stigmergy / Niche Construction loop)

Stigmergy ↔ "Environment as actor" (ANT)

Stigmergy (Grassé 1959, Heylighen 2016) is the formal description of HOW the environment acts as an actor. In ANT, the environment is not a passive backdrop — it's an actant. Stigmergy gives the mechanism: agents modify the medium, the medium stores information, channels action, and constrains future behavior. The medium is causally efficacious. This makes ANT's claim concrete — "there is no in between networks" means the stigmergic medium is a full participant, not a stage.

Stigmergy ↔ Downward causation (Hofstadter)

When agents modify their environment and those modifications constrain future agents, that's downward causation. The termite mound (collective product) shapes termite behavior (individual). The stigmergic feedback loop (action → trace → stimulation → action) IS a strange loop through the medium. The agents produce the environment; the environment produces the agents' behavior. Level-crossing through the medium, not through direct agent-to-agent interaction.

Stigmergy ↔ Niche construction ↔ Cross-scale interaction

Niche construction theory (Laland, Odling-Smee) shows that organisms modify their environment, and those modifications feed back into their own evolution via ecological inheritance. The loop: organism → environment → selection → organism. This is a cross-scale feedback loop — and it runs through the environment (stigmergically), not through direct interaction. The environment mediates between scales. This is the cross-scale interaction mechanism we've been looking for.

Stigmergy + Autopoiesis ↔ Multi-scale composition

Stigmergy alone coordinates within a scale. It does not by itself produce new scales. The phase transition from one scale to another requires the accumulated traces to become self-maintaining — autopoietic. The termite mound is not just a trace; it's actively maintained, repaired, and regulated. When it crosses from passive trace to self-maintaining structure, it becomes a new actor at a new scale. Stigmergy provides the medium; autopoiesis provides the persistence at the new scale; the crossing from trace to actor IS the multi-scale phase transition.

Stigmergic traces ↔ Quasi-objects

The stigmergic trace is a quasi-object (Serres/Latour): it circulates through the network AND is transformed by circulation. The pheromone trail is strengthened or weakened by each ant. The Wikipedia article is transformed by each editor. The trace is co-determined with its carriers — not a fixed signal. This is exactly our H3 (quasi-object resource hypothesis): resources that transform through circulation produce richer dynamics.

Transient vs. persistent traces ↔ Adaptability vs. memory

Stigmergic traces decay (pheromones evaporate) or persist (termite mounds). Transient traces enable adaptation (outdated trails decay, new ones form). Persistent traces enable accumulation (long-term memory). The trade-off is fundamental to multi-scale systems: a new level needs persistent traces to accumulate, but also transient traces to adapt. The optimal decay rate for the trace→actor crossing is an open question.

2026-07-21 — Session 4 (Echo / NK Model / Fitness Landscapes loop)

Echo's failure ↔ Our multi-scale composition thesis

Smith & Bedau (1997) ran thousands of Echo simulations and found it fails to produce "the diversity of hierarchically organized adaptive aggregates" that characterizes CAS. Echo converges to simple trading ecologies. They independently arrived at our thesis: the missing ingredient is "robust, open-ended emergence of hierarchical, adaptive structures" — which IS multi-scale composition. This is strong external validation from a completely different starting point (empirical study of a CAS model, not ANT/computational irreducibility).

Smith & Bedau's 8th CAS property ↔ Our H7 (trace→actor crossing)

They proposed an 8th CAS property: "the ability of emergent interacting components to create and flexibly maintain their own boundaries and their capacities for interacting with other components." This maps exactly to our synthesis:

• "Create boundaries" = stigmergy (traces that accumulate and form structures)
• "Flexibly maintain boundaries" = autopoiesis (self-production, self-repair)
• The crossing from trace to 8th-property actor = our trace→actor crossing (H7)

They identified this in 1997 but never implemented it. They wrote: "Concretely embodying them in some successor model to Echo is the only way to make them precise and subject them to rigorous scrutiny." Our project is that successor model.

NK model's static landscape ↔ Stigmergy's dynamic landscape

The NK model (Kauffman) defines a FIXED fitness landscape — agents adapt TO it, but cannot reshape it. Stigmergy makes the landscape DYNAMIC — agents modify the landscape they're adapting to (niche construction). This is the key limitation of the NK model for multi-scale systems: static landscapes cannot produce multi-scale composition because agents cannot reshape the selection pressures at other scales. Dynamic landscapes are necessary for the cross-scale interaction mechanism.

Computational complexity ↔ Computational irreducibility ↔ Open-ended evolution

Kaznatcheev (2019) proved NK landscapes with K > 1 are PLS-complete — even local fitness optima cannot be found efficiently. This is an ULTIMATE constraint (property of the landscape, not the algorithm). But crucially, this constraint ENABLES open-ended evolution: on easy landscapes, evolution converges to a peak and stops; on hard landscapes, it cannot converge and keeps searching. This connects Wolfram's computational irreducibility to open-ended evolution: irreducibility is not just a property of the simulation but a NECESSARY CONDITION for open-endedness. Without it, the system converges. (H8)

Fitness landscape metaphor criticism ↔ ALife simulation design

The fitness landscape metaphor (Wright 1932) is criticized (Kaplan 2008, Petkov 2015) for assuming static, fixed genotype-fitness mappings. ALife simulations inherit this assumption — even "emergent" ALife has implicit static landscapes. If we design simulations thinking in terms of "fitness peaks," we'll get convergence to peaks. We need to think in terms of dynamic, multi-scale landscape cascades where each scale's landscape is reshaped by the scale below (stigmergically) and the scale above (downward causation).

Echo's counterintuitive resource accumulation ↔ Selection pressure surprises

Smith & Bedau found that in Echo, genomes with MORE of the traded resource dominate — even though this makes them HARDER to replicate. More resource in genome → harder to copy → agents live longer → acquire more resources → support larger population. The "fittest" genome is the hardest to replicate. This is a cautionary tale for simulation design: selection pressure can produce counterintuitive dynamics. Our simulations should not assume that "fitter" means "easier to replicate."

Holland's aggregation property ↔ ANT's translation

Holland's "aggregation" property (meta-agents built from simpler agents) is what ANT's translation describes: the process by which actors form collectives that act as one. But Holland's aggregation is a PROPERTY — it either exists or doesn't. ANT's translation is a PROCESS — it describes HOW aggregation happens (problematization → interessement → enrollment → mobilization). Echo has the property defined but never achieves it in practice. ANT gives us the mechanism to make aggregation happen.

2026-07-20 — Vance's contribution: Multi-rate bounded environment

Multi-rate environment ↔ Multi-scale composition

Different rates of environmental change = different scales. Geological change (slow) and temperature change (fast) interact through organisms. The interaction IS multi-scale composition. Slow actors provide stability (memory), fast actors provide variation (adaptation). Same trade-off as pheromone decay rates but at the environmental level.

Multiple fitness criteria ↔ ANT actor networks

The environment is not one actor but a network of actors, each with its own rules, each changing at its own rate. The organism faces a network of actants, not "the environment." This is ANT made concrete for simulation design.

Multi-rate environment ↔ Stigmergy

Each environmental actor leaves traces at its own rate. Land leaves persistent traces (geology). Temperature leaves transient traces (weather). The environment is multiple stigmergic media with different decay rates, each carrying different information.

Multi-rate ↔ Open-ended evolution

Single fitness goals lead to stasis. Multiple fitness criteria changing at different rates create a fitness landscape that never settles — while one pressure stabilizes, another shifts. This prevents the EvoLoop convergence problem. The fitness landscape is a moving target in multiple dimensions.

2026-07-22 — Session 5 (Evolving Reaction Networks / Signals & Boundaries loop)

Sim03's negative result ↔ COT's evolvability limitation

Our sim03 (chemical organizations with fixed reaction network) confirmed the central limitation of Chemical Organization Theory: the system converges to a fixed equilibrium by generation 1 and NEVER changes for 3000 generations. Both single-trace and multi-trace conditions reach a static state. This is exactly Vasas et al.'s (2010) finding: autocatalytic sets (fixed networks) "lack evolvability" — they "cannot substantially depart from the asymptotic steady-state solution already built-in in the dynamical equations." Sim03 independently confirms this through simulation: fixed reaction networks cannot evolve, regardless of trace diversity. (Corrected 2026-07-27: not an independent confirmation. sim03 enumerates subsets of a fixed, hand-authored network, so its organization count is identical at every sampled generation of every run — the stall is guaranteed by the design rather than measured. The concentration equilibrium is a genuine result; the organizational stall is not a test. Counts also changed after a closure fix: 8 organizations single / 9 multi, 1/24 nested pairs.)

Vasas et al. (2012) resolution ↔ H9 (Evolving Network Hypothesis)

Vasas et al. (2012, "Evolution before genes") found the way out: rare uncatalyzed reactions produce novel species. Most disappear, but rarely a novel species catalyzes its own production from existing molecules, forming a viable autocatalytic core — a new organization. Combined with compartmentalization (which filters harmful modifications and enables between-compartment selection), this produces the minimal conditions for Darwinian evolution in chemical networks. This directly motivates our H9: evolving networks (where new reactions appear) produce evolvable organizations where fixed networks stall. Our sim04 tests this directly.

Novel viable cores ↔ Trace→actor crossing (H7)

The appearance of a novel viable core IS the trace→actor crossing in formal COT terms. Existing resources are "traces" (accumulated products of reactions). A novel reaction among them produces a new self-maintaining set (closure + self-maintenance = organization). The new organization is a new "actor" at a new level. This makes H7 mechanistically concrete: the crossing occurs when a rare novel reaction produces a viable autocatalytic core from existing resources.

Two-level autocatalysis ↔ Multi-scale composition (H1)

Vasas et al. identify two levels of autocatalysis: molecular (within compartments, reactions produce molecules that catalyze more reactions) and compartmental (compartments grow and divide). These levels have DIFFERENT rules — molecular level produces novelty (new cores), compartmental level selects among them. This IS multi-scale composition: two levels with different rules, interacting through the containment relationship. The molecular level's products (cores) become the compartmental level's units of selection.

Holland's Signals and Boundaries ↔ Stigmergy + Autopoiesis synthesis

Holland's (2012) final framework — CAS as co-evolving signal/boundary hierarchies — arrives at the same synthesis we identified in Session 3 from a completely different direction (CAS theory vs. ANT + stigmergy):

• Signals = stigmergic traces (environmental modifications that coordinate behavior)
• Boundaries = autopoietic structures (self-maintaining entities that filter what crosses them)
• Co-evolution = traces modify boundaries, boundaries filter traces (the stigmergic feedback loop through the medium)
• Hierarchy = nested boundaries = multi-scale structure

Three independent paths (Holland from CAS theory, Vasas from origin-of-life chemistry, our project from ANT + computational irreducibility) converge on: evolving signal/boundary hierarchies = multi-scale composition.

Multiple attractors ≠ evolvability

Vasas et al. found that networks with inhibition had multiple attractors but they were NOT selectable — transitions between attractors were periodic or chaotic, overriding any selection pressure. This is a crucial refinement: multiple attractors (multiple organizations) is necessary but not sufficient for evolvability. The attractors must be stable, heritable, and differentially fit. In COT terms: multiple organizations must exist AND be separable (compartments) AND have different growth rates. This refines our understanding of what the trace→actor crossing requires — not just self-maintenance, but selectable self-maintenance.

The "one bit" problem ↔ Open-ended evolution gap

A viable autocatalytic core carries approximately one bit of heritable information (present/absent). Vasas et al. acknowledge this means autocatalytic networks "may not be able to sustain open-ended evolution." The gap between "evolvable" (selection between 2-3 attractors) and "open-ended" (unbounded novelty) is enormous. Each novel core extends the "adjacent possible" — opening new reaction possibilities — but whether this combinatorial expansion produces true open-endedness or just limited multi-attractor dynamics remains the central open question. This connects directly to H8: computational irreducibility at each scale is necessary but may not be sufficient for open-endedness.

2026-07-23 — Session 6 (AlChemy / Lambda Calculus Chemistry loop)

AlChemy's unbounded space ↔ Sim04's finite space limitation

Sim04 stalled because its binary polymer space was finite (510 species). AlChemy uses lambda calculus as chemistry — expressions are unbounded, the molecule space is infinite. Our sim05 confirms: each run explores 246-930 unique species, no two runs overlap, and the space is never exhausted. (Corrected 2026-07-27: 112–162 species per run — the earlier figures were inflated ~3–6× because species identity was not alpha-invariant. The runs are still largely distinct, mean pairwise overlap 0.061, and the space is still never exhausted, so the conclusion of this paragraph stands.) Unbounded space is NECESSARY (without it, finite exhaustion is inevitable) but NOT SUFFICIENT for multi-scale composition.

L2 composition failure ↔ Multi-scale composition (H1)

Mathis et al. (2024) found that "stable organizations cannot be easily combined into higher order entities" in AlChemy. Our sim05 confirms: 0/6 pairs of L1 organizations achieved L2 coexistence. 50% dominance (one destroys the other), 50% mutual destruction (both destroyed). This is the SAME failure as Echo (Smith & Bedau 1997), COT/Vasas, and sim04. Three independent modeling traditions — CAS theory, prebiotic chemistry, computational theory — all fail at multi-scale composition. This convergence is strong evidence the problem is FUNDAMENTAL, not an artifact of any single approach.

Correction (2026-07-27): sim05 does not corroborate Mathis et al. as stated, and the convergence argument has lost one of its three legs. Three defects each biased sim05 against coexistence — non-alpha-invariant species identity, a Jaccard metric whose ceiling fell below the coexistence threshold for two of six pairs, and a mixed population padded almost entirely from organization A (~9:1 abundance handicap, which is why every pair returned dominance-by-A). Corrected: 2/6 coexistence (33%), 3 dominance, 1 mutual destruction, stable across survival thresholds 0.45–0.70.

Mathis et al. and Fontana & Buss are untouched — the literature still reports L2 as rare. But "three independent traditions all fail" now overstates our own contribution: in our model composition succeeds a third of the time. Note too that sim05 never tests closure or self-maintenance, so its "L1 organizations" are surviving species sets rather than organizations in the COT sense the comparison assumes. See simulations/REVIEW.md §2.

The "glue" ↔ Trace→actor crossing (H7)

Fontana & Buss identified "glue" expressions that bridge L1 organizations into L2 composites. Glue is produced by composing functions from different organizations — it cannot exist without at least one L1, yet it bridges between them. This is exactly our trace→actor crossing: glue is the stigmergic trace that enables the phase transition between scales. The fact that glue rarely emerges spontaneously confirms H7: the crossing requires specific mechanisms (stigmergic bridges, autopoietic boundaries, explicit selection), not just random interaction.

Sensitivity to initial conditions ↔ Computational irreducibility (H8)

AlChemy is extremely sensitive to its random expression generator. The original (probabilistic grammar) produces diverse organizations; the permutation generator (uniform binary trees) collapses to trivial fixed point. Each L1 run produces a unique organization — unpredictable from initial conditions. Whether two L1s will compose is also unpredictable. This is computational irreducibility at two levels: organization formation AND composition. You must simulate to know.

Three paths, same failure ↔ Multi-scale composition is fundamental

Echo (Holland's CAS model), chemical organizations (COT/Vasas), and AlChemy (lambda calculus) ALL fail at multi-scale composition. Each from a different starting point:

• Echo: CAS theory (agents with endogenous fitness in resource-limited environment)
• COT/Vasas: origin-of-life chemistry (autocatalytic sets with compartmentalization)
• AlChemy: computational theory (lambda calculus expressions as unbounded molecules) This convergence is the strongest evidence yet that multi-scale composition is not a bug of any particular simulation but a fundamental gap in our understanding of how scales interact. The composition problem persists across finite (sim04: 510 species) and infinite (sim05: unbounded) spaces, across chemical and computational substrates, across selection-based and mass-action dynamics.

Mutual destruction produces novelty ↔ Creative destruction at scale boundaries

In sim05, mutual destruction (both L1s destroyed) produced the most novel species (89-90 unique vs. 6-23 for dominance). (Retracted 2026-07-27: this reverses under the corrected sim05. Only 1 of 6 pairs now ends in mutual destruction, and its final population is the smallest at 10 species, not the largest. The 89-90 figures came from inflated, non-alpha-invariant species counts. The analogy to hybridization below is a nice idea with no remaining empirical support from our data.) Cross-organization interactions generate novelty but destabilize existing structures. This parallels biological phenomena: hybridization can produce novel species but often destroys parental lineages. The multi-scale composition problem may require a mechanism that captures this novelty without destroying the parents — which is exactly what autopoietic boundaries (Holland's signals & boundaries) would provide.

Vance's contribution: The termite mound principle (2026-07-22)

Inert substrate → dynamic actor through organization

Vance's insight from the termite mound analogy: the mud termites trail is largely INERT until it achieves a mass sufficient to affect temperature when interacting with other secondary effects (sunlight, chemical reactions, etc.). The substrate becomes dynamic — becomes an ACTOR — only when it crosses an organizational threshold. This is the trace→actor crossing (H7) made concrete in a physical system: inert mud → thermally active structure → agent that reshapes the selection landscape (temperature gradients drive termite behavior, which builds more mud structure). The substrate doesn't just accumulate; it changes state and becomes causally efficacious at a different scale.

Multi-rate, multi-attractor environment ↔ Diverse stable organization

The termite mound works because termites act in a complex environment with MULTIPLE CHANGING FITNESS ATTRACTORS and substrates that become dynamic as they become organized. This is Vance's refinement to the multi-scale composition thesis: unbounded space alone (sim04, sim05) is necessary but not sufficient because the space is homogeneous — all locations are equivalent. What's needed is a HETEROGENEOUS environment where:

1. Multiple fitness attractors operate simultaneously at different rates (temperature, chemistry, light, moisture)
2. Apparently inert substrates can become dynamic when they cross an organizational threshold (mud → thermal mass)
3. The environment is not a static backdrop but a network of actants (ANT) that become causally efficacious through organization

This connects directly to the sim04/sim05 negative results: both simulations used homogeneous spaces (flat polymer space, flat lambda calculus space). Neither had environmental heterogeneity, multi-rate attractors, or substrate state transitions. The termite mound principle suggests the missing ingredient for multi-scale composition is not just "more space" but a structured environment where organization transforms the environment itself, creating new selection pressures at new scales.

Implications for sim06+ design

• Simulations need heterogeneous environments with multiple interacting gradients (not just a flat reaction space)
• Substrates should have state transitions: inert → active when organization crosses a threshold (mass, density, complexity)
• Multiple fitness attractors changing at different rates prevent convergence to a single equilibrium
• The environment should be co-determined with the organisms (niche construction at multiple scales)
• This is the physical grounding for the "dynamic landscape" the synthesis has been pointing toward (Session 4: NK model's static landscape ↔ stigmergy's dynamic landscape)

2026-07-26 — Session 9 (environmental physics coupling / the structure as a new dynamical degree of freedom)

The Mahadevan mechanism ↔ The trace→actor crossing (H7 specified)

King, Ocko & Mahadevan (PNAS 2015) measured diurnal cyclic convection in O. obesus mounds: geometry + heterogeneous thermal mass + porosity converts a passive temperature oscillation into directed ventilation. Ocko, Heyde & Mahadevan (PNAS 2019) showed a model coupling environmental physics to building behavior reproduces the full range of mound morphologies. The structure's own physics (airflow) redistributes the pheromone cues that guide building — the structure IS the feedback path, not just the product. This specifies the "new dynamical degree of freedom" H7 needed: the accumulated structure must gain a transport dynamics absent at the deposit level. The crossing is the onset of the structure's physics as a causal layer. sim06 had no such physics (just a sum of deposits), so it never crossed.

The 20-year lineage ↔ sim06's null result is a known field-wide gap

Linardou (2008, UCL) documents that Deneubourg (1977) → Bonabeau (1997) → Ladley & Bullock (2004) ALL share the same limitation: "the already deposited building material had no influence on the termite movement" and pheromone diffusion is "unrealistic" (decoupled from structure). sim06 inherited exactly this. So sim06's null result is not a failure of our model — it is a minimal modern confirmation of a gap the field has carried for 20 years. The Mahadevan model is the first to include the coupling, but it is a physics model, not an agent model. sim07 is the first agent model to attempt the coupling (in minimal lumped form).

The 2025 state of the art ↔ H7 in the field's own language

Karibi-Botoye, Theraulaz et al. (J R Soc Interface 2025) list as an open question: "What processes occur at smaller scales in the mound that control larger-scale observations?" — this is the trace→actor crossing question stated in the field's own terms (smaller-scale deposit/pore processes controlling larger-scale ventilation/morphology = multi-scale composition). They call termite-inspired buildings "bio-mythological" because they mimic appearance without the physics — the engineering consequence of the missing coupling. The field's prescription (multiscale numerical modelling of pressure/velocity/heat/CO₂ transport) is the full-physics version of what sim07 needs only in lumped form.

Vance's inert→active state transition ↔ The M_c phase transition (sim07 design)

Vance's termite-mound principle (inert mud → active actor above a mass threshold) maps onto the minimal lumped sim07 design: a structure-sourced transport field T with a mass threshold M_c. Below M_c, structure is inert (sim06 scatter). Above M_c, structure activates — it sources T, which vents pheromone away from saturated regions (the negative feedback sim06 lacked). The H7 prediction becomes operationally testable: the crossing is a phase transition in M_c. Sweep M_c → look for morphology transition (scatter → few consolidated vented pillars) coinciding with the detector firing.

Multi-rate environment (H4) ↔ The energy source for transport

The Mahadevan mechanism's energy source is a diurnal oscillation — an external multi-rate driver the structure rectifies into directed flow. sim07's lumped T field has no external clock unless we add one. This suggests the crossing may require not just structure-sourced transport but an external oscillation the structure can rectify — a link to H4 (multi-rate environment) and a candidate sim08 extension. The dynamic environment (H4) is now concrete: not a changing fitness function, but an environment whose physics the structure can harness.

Circular-input risk ↔ The self-repair test as safeguard

A genuine criticism of sim07: by building in the transport rule, we risk building in the crossing we claim to detect. The safeguard is the perturbation/self-repair test — the structure must recruit maintenance through its transport dynamics after damage, not through the deposit rule. If self-repair works only when T is active (above M_c) and fails when T is suppressed, the crossing is emergent from the physics, not imposed.

2026-07-25 — Session 8 (sim06 implementation / the negative-feedback gap)

(Section header added 2026-07-27. This block was appended without a dated heading, so it sat under Session 9's header and read as current fact rather than as a Session 8 record. It is Session 8 material — the sim06 implementation night.)

Correction (2026-07-27). The sim06 figures in this section are wrong and the null they describe was an artifact. The crossing detector could not fire: criterion 2 required the deposit rate to fall below its early-run average, impossible under Grassé positive feedback once structure exists. Corrected: baseline 66–109 components (not ~230), stability 0.849–0.893 (not 0.55), deposit_on_structure 0.70–0.79 — criterion 3 passes 154/160 (not 0.33-fails). Post-fix the crossing still doesn't fire, but criterion 1 misses by ≤0.05 — a near miss, not a diffuse scatter. Unchanged and still correct: 66% more structure (1876 vs 1131 cells). The negative-feedback direction may still be right, but its support is now sim06's self-maintenance reversal (more fragmented: 219–297 components; less selective: 0.43–0.53) and sim07's null, not the scatter claim below. See simulations/REVIEW.md §1.

sim06's null result ↔ The negative-feedback gap

sim06 tested H7 with a minimal Grassé stigmergy model. The result: positive stigmergic feedback (deposits attract deposits) amplifies building — self- maintenance produces 66% more structure than baseline — but the formal crossing detector never fires across a wide parameter sweep. Structure stays at ~230 scattered micro-pillars, stability 0.55 (criterion needs 0.90), constraint 0.33 (needs 0.60). Diagnosis: the model has positive feedback + weak decay but no consolidation mechanism — no negative feedback that makes strong pillars inhibit nearby nucleation or redirect activity. The deposit rule saturates at 0.95 but never decreases. Every cell with any pheromone is (nearly) equally attractive, so the structure spreads instead of consolidating. This connects to Heylighen's (2016) point that complex stigmergic systems need both positive and negative feedback — positive amplifies, negative stabilizes and diversifies. sim06 had only the positive half.

Environmental physics coupling ↔ The trace→actor crossing (H7 refinement)

The Mahadevan group's termite mound model (Ocko, Heyde & Mahadevan, PNAS 2019) shows what sim06 is missing: real mounds aren't passive accumulations — they're ventilation structures whose own physics (airflow from thermal gradients) redistributes the pheromone cues that guide building. The structure is the feedback path: the macro-structure's transport dynamics determine where the micro- scale signal goes. This is the trace→actor loop, but it requires the environment to have physical transport dynamics, not just decay/diffusion. H7 refines: the crossing requires not just that the trace recruits its own maintenance (sim06's loop, which worked weakly) but that the accumulated structure introduces a new dynamical degree of freedom — a process absent at the deposit level (transport, inhibition, competition, or a state transition). sim06's structure had no such degree of freedom — it was just a sum of deposits, so it never crossed.

Vance's termite-mound principle ↔ sim06's missing ingredient

Vance's insight (Session 6): inert mud becomes a dynamic actor only when it crosses an organizational threshold (mass → thermal effect). sim06 modeled the deposit accumulation but not the state transition — the structure never gains new physics as it grows. The "inert → active" transition is exactly the new dynamical degree of freedom H7 needs. This is the "dynamic landscape" made concrete: not a changing fitness function, but an environment whose own physics becomes a new causal layer once organization crosses a threshold. sim07 should implement the state transition (e.g. structure above a mass threshold begins to channel a transport field, or inhibits nearby deposition) and test whether that unlocks the crossing.

The negative-feedback prescription ↔ sim07 design

The refinement is actionable. sim07 candidates for the missing negative feedback: (a) saturation/inhibition — cells above a density cap repel deposits; forces few large pillars. (b) environmental transport — the structure channels an advective field that redistributes pheromone away from saturated regions (the Mahadevan mechanism, minimal lumped version). (c) competition — multiple trace types or clusters compete for finite termites, so one cluster's growth inhibits another's. (d) state transition — substrate above a mass threshold unlocks new dynamics (Vance's principle). Each predicts a different morphology; each is testable. The hypothesis: below a critical negative- feedback strength, diffuse scatter (sim06); above it, consolidated actor (the crossing). If that phase transition exists, it's the H7 crossing made operational.

2026-07-27 — Session 10 (sim07 null: scalar transport has the wrong sign for consolidation)

The minimal lumped transport field ↔ the crossing (H7 refined again, still not refuted)

sim07 implemented exactly the Session-9 prescription: a structure-sourced scalar transport field T (sourced above a mass threshold M_c, diffuses, vents pheromone from saturated to gap regions) — the ONLY addition to sim06. The H7 prediction was a phase transition in M_c. Result: no phase transition. Sweeping M_c from inert to fully active monotonically decreased stability (0.876 → 0.739) and fragmented pillars (57 → 128); the crossing detector never fired for any M_c or transport_coupling. The null is a third progressive refinement of H7, each narrowing the hypothesis: (sim06) positive feedback alone is insufficient → (sim07) scalar structure-sourced transport alone is insufficient → the crossing requires directed transport and/or an externally-driven one. This is the spiral-loop methodology working: each null specifies the next experiment.

Scalar venting ↔ wrong sign for consolidation (the mechanistic diagnosis)

The negative feedback is real but its effect has the wrong sign for consolidation. Venting pheromone away from saturated pillars disperses the very cue that recruits deposits — so transport fragments rather than consolidates. A lumped linear advection of a scalar cue does NOT reproduce the Mahadevan mechanism, where directed flow carries the cue along channels to where building should continue. The minimal lumped version lost the directionality that makes real mound transport consolidate. This is a generalizable lesson for minimal models: collapsing a directed physical process to an isotropic scalar field can invert its effect. The "environmental physics coupling" the crossing needs is not just "the structure sources a field" but "the structure sources a directed field whose geometry channels the cue where building should continue."

The circularity safeguard ↔ its own null result

sim07's perturbation/self-repair test was designed as the circularity safeguard: if repair tracks T (not the deposit rule), the crossing is emergent. The test produced its own null: both conditions recover (recovery ≈ 1.0), but repair is driven by the deposit rule (termites wander back), NOT by T. So T is demonstrably not the causal layer — confirming the null is a property of the mechanism, not a detector artifact. The safeguard worked: it prevented claiming a crossing that wasn't there. This validates the detector-safeguard pattern as a method: a mechanism whose perturbation response doesn't track the proposed causal layer is not the causal layer, full stop.

Directed transport ↔ external multi-rate driver (H4) — the two remaining paths

The null leaves two candidates. (1) Directed transport: channel geometry that carries cue to building fronts (the Mahadevan directionality, lost in the lumped scalar). This requires modeling the structure's shape as a channel, not just its mass — a richer morphological state. (2) External multi-rate driver (H4): the diurnal oscillation the structure rectifies into directed flow — the Mahadevan energy source sim07 omits entirely. sim07's T is structure-sourced but has no external clock; the Mahadevan mechanism's energy comes from outside the structure. Candidate sim08 tests the external- oscillation path. The multi-rate environment (H4) is now concrete: not a changing fitness function, but an environment whose physics the structure can harness only with an external driver.

2026-07-27 — Post-review consolidation (what the corrected simulations actually support)

Not a research session. After the construct-validity audit (simulations/REVIEW.md) and the rerun of all seven simulations, this section restates what our own code now supports, as distinct from what the literature supports. The correction pass fixed wrong numbers in place; this is the part that could not be done by correction — deciding what the project still believes.

The convergence argument, re-derived

The strongest claim the project has made is that three independent traditions all hit the same wall: Echo (CAS theory), COT/Vasas (prebiotic chemistry), and AlChemy (computational theory) each fail at multi-scale composition, so the problem is fundamental rather than an artifact of any one approach. That argument has to be restated, because our contribution to it was weaker than recorded:

• Echo leg — intact. Smith & Bedau (1997) is a literature result from thousands of runs. We never simulated Echo; we only cited it. Untouched by the audit.
• COT/Vasas leg — literature intact, our contribution withdrawn. Vasas et al. (2010) proved autocatalytic sets lack evolvability. sim03 was recorded as independently confirming it through simulation. It does not: sim03 enumerates subsets of a fixed, hand-authored network, so its organization lattice is identical at every generation of every run. The stall is a property of the design, not a measurement.
• AlChemy leg — literature intact, our replication mildly contradicts the strong form. Mathis et al. (2024) and Fontana & Buss (1994) both report L2 as rare. sim05 was recorded as independently confirming that at 0/6 coexistence; corrected, it is 2/6. Composition is the minority outcome, not an impossibility.

Honest restatement: three literatures converge on composition being hard. Our simulations have not independently reproduced that convergence, and one of them mildly contradicts its strongest reading. The thesis is not damaged — "composition is hard" still has solid external support — but the project should stop citing its own sims as a third independent confirmation. What sims 03–05 actually established is narrower: finite species space exhausts (sim04, 510/510, unchanged), and fixed networks have static organization lattices (sim03, true by construction).

The new finding: both attempts at negative feedback increased fragmentation

The corrected data surfaced a result that was invisible before, and it points the opposite way from the mechanism we proposed. Two independent attempts to add the "missing" negative feedback both made the structure less consolidated:

Both are monotonic and both are backwards. The common cause is that both act through the pheromone field, and the deposit response saturates: p = base + gain·φ/(1+φ) is flat above φ≈1, so once the field is driven high anywhere, deposit probability is ~0.87 everywhere and the spatial contrast stigmergy depends on is destroyed. Adding energy to a saturating channel does not create selectivity — it removes it.

This is now H11, the Saturating Channel Hypothesis. It reframes what H7 needs. The Session 8/9 prescription was "add negative feedback / a new dynamical degree of freedom." Tried twice, it fragmented twice. The refined prescription is negative feedback through a channel that does not saturate — acting on deposit probability or on geometry directly, rather than by pushing more signal through a saturating cue field. That is a sharper and more falsifiable claim than the original, and it came out of a bug fix rather than new reading.

Cross-domain connection: saturation as the hidden variable

Heylighen's positive/negative feedback framing assumes the two act on comparable channels. Our result suggests a third term: the response curve of the agents to the trace. A saturating response makes negative feedback self-defeating, because the manipulation that is supposed to create contrast operates in the region where contrast cannot be expressed. This connects to the multi-rate environment idea (Vance's contribution) from the other side — it is not only that different actors must operate at different rates, but that the medium must remain responsive across the range those rates drive it through. A saturated medium is a single-rate medium no matter how many processes write to it.

Status of the hypotheses after the review

• H1 — unchanged in substance; its sim03 leg is withdrawn, leaving Smith & Bedau as the external support and our own sims as untested rather than confirming.
• H7 — not refuted and not strongly tested. sim06 was a near miss with a broken detector; sim07 is a sound null against the scalar form of the mechanism. Newly supported by the fragmentation reversal above, which is better evidence than the "diffuse scatter" claim it replaces.
• H9 — weakened on two legs (sim03 structural, sim05 no longer showing unbounded space stalling); sim04's finite-space exhaustion survives.
• H10 — weakened. 2/6 coexistence. Still standing on the literature.
• H8 — untouched. Purely literature-derived (Kaznatcheev 2019), no simulation dependency.

What would settle things

1. What distinguishes the 2 coexisting sim05 pairs from the 4 that did not? This is now the most informative question in the repo and it did not exist before the fix — at 0/6 there was nothing to compare.
2. Does non-saturating negative feedback consolidate? A sim08 that inhibits deposition directly (a density cap or refractory period) rather than by manipulating the cue field would test the refined prescription above, and is a cheaper experiment than directed transport.
3. Repeat sim06's parameter sweep against the working detector. It has never been run against a detector capable of firing, so the claim that no regime produces the crossing is simply unsupported — in either direction.

2026-07-28 — Session 13 (non-saturating channels: biology grounds H11, sim08 tests it)

Real termites use non-saturating channels, not a saturating cement pheromone ↔ H11

Three independent lines of termite research converge on the channels H11 prescribes, and away from the saturating cue sim06/sim07 used:

• Calovi et al. (2019, Phil Trans R Soc B) disambiguated surface curvature from inclination and height across three orientations in M. michaelseni and found curvature is the "consistent and sole driver" of construction. Concave → deposit, convex → excavate, and the SAME cue elicits OPPOSING actions depending on the termite's loaded/seeking state. Curvature is geometric and redefined by each deposit — it cannot saturate. This is the "act on the action, not the cue" prescription, observed in the animal. And crucially: "no cement pheromone has yet been identified." The biological system evolved away from the saturating channel H11 flags as self-defeating.
• Carey et al. (2021, Front Robot AI) validated the humidity-template rule with a robot: threshold-triggered deposition at the edge of a humidity bubble, rerouted by wind. A discrete threshold, not a graded saturating response.
• Xiao et al. (2026, arXiv) frame crowding/inactivity as "distributed inhibition that prevents saturation" — the density-cap channel, observed.

So H11, which came out of a bug fix in our own code, is independently corroborated by what termites actually do. The channels H11 prescribes (density cap, refractory/threshold, directional/geometry) are the ones termites evolved; the saturating cue channel (cement pheromone) is the one biology may not use at all. H11 may be less a rediscovery of ACO (which bounds the cue via MMAS) and more a rediscovery of termite biology.

sim08 (density cap) ↔ H11 directionally confirmed, sufficiency sharpened

sim08 added a non-saturating density cap (hard boolean gate on the deposit action) to sim06, reusing all metrics and the detector unchanged. The cap consolidates morphology, monotonically — pillars 101 → 52 as the cap tightens, max pheromone 8.01 → 2.50 (the field de-saturates, exactly as H11 predicts). But the crossing does not fire; stability doesn't rise (0.874 → 0.775 at the tightest cap). The cap limits growth without recruiting maintenance: it corrects the fragmentation symptom (pillars) but not the persistence symptom (stability).

This is the third mechanism in a row that confirms H11's direction (each non-saturating / cue-independent approach consolidates morphology where cue-field approaches fragmented) while showing the crossing needs more. The boundary narrows: positive feedback alone (sim06) → scalar cue-transport (sim07) → non-saturating limitation (sim08) all insufficient. The crossing needs a non-saturating channel that recruits as well as limits. The curvature channel does both (routes to concavities AND each deposit extends the concavity); the density cap only limits.

Curvature as the minimal form of directed transport ↔ H7

This is a cross-domain connection: the "directed transport" H7's Session-10 refinement called for (channel geometry carrying cue to building fronts) may be the same thing as the curvature channel — curvature IS directed geometry. The Calovi rule (deposit at concavity, excavate at convexity) is a routing rule that builds along existing edges, which is both the "directional bias" H11 listed and a minimal lumped version of the directed transport sim07's scalar lacked. Candidate sim09: a curvature/deposition-edge rule — non-saturating, geometry-based, AND self-recruiting — may be the cheapest rule that could actually cross.

Saturation as the hidden variable ↔ the medium must stay responsive

Heylighen's positive/negative feedback framing assumes the two act on comparable channels. H11 + sim08 + the termite biology add the third term: the response curve of the agents to the trace. A non-saturating action-gate (the cap) de-saturates the cue field and prunes nucleation — but a pure limiter cannot reach the crossing because it does not feed back into the structure's persistence. The medium must stay responsive across the range processes drive it through, AND the feedback must recruit (route building to where it extends the structure), not merely cap. This generalizes beyond termites: any stigmergic system whose cue response saturates will fragment under cue-based negative feedback, and any non-saturating limiter that doesn't also recruit will consolidate morphology without reaching actorhood.

2026-07-29 — Session 14 (the curvature/evaporation unification — the recruiting non-saturating channel gets a model)

Curvature ≡ evaporation flux ↔ the humidity and curvature channels are one

Facchini et al. 2024 (eLife) proved evaporation flux is directly proportional to surface curvature (Langmuir 1918), so the curvature channel (Calovi 2019) and the humidity/evaporation channel (Carey 2021) are the same physical quantity sensed through one gradient. This unifies two of the three non-saturating channels Session 13 identified — they were never separate. The third (crowding, Xiao 2026) remains independent. The practical upshot: sim09 needs to model ONE geometry/evaporation channel, not two, and the three "channels" are actually two (geometry + crowding).

The convex/concave contradiction ↔ don't conflate construction actions

Calovi 2019 (concave → activity) and Facchini 2024 (convex tips → deposit) seemed to contradict. The resolution: they measured different action components — Calovi measured aggregate activity (digging + building), Facchini isolated pellet deposition. Deposition is at convex tips (growth extends the structure outward); excavation is at concave pits. This is a methodological lesson for sim09: "construction" is not one action. sim06 had only deposit; sim09 must separate deposit (loaded termites at convex tips) from excavate (unloaded termites at concavities) to reproduce the curvature rule correctly. Conflating them — as a single "build" action — would invert the rule's sign.

A published curvature-only growth model ↔ the sim09 substrate exists

Facchini, Lazarescu, Perna & Douady (2020, J R Soc Interface) built a phase-field growth model for Nasutitermes nests driven entirely by local mean curvature — no pheromone field at all. The equation ∂f/∂t = f(1−f)·[(1/2)·Δf + d·Δ²f] has: a growth term (mean curvature Δf — the recruit mechanism, positive at convex tips), a smoothing term (d·Δ²f — the limit mechanism, caps feature size), and a surface-restriction prefactor f(1−f) (deposits at edges, not bulk — spatial selectivity without a saturating cue). For large d it is linearly unstable: walls expand, branch, merge, invade space — the consolidation morphology sim06 never reached. Public code exists. This is the candidate sim09 substrate: replace sim06's saturating pheromone-deposit rule with the curvature growth rule, adapt to 2D, and test whether the d instability is the phase transition the crossing needs.

"Recruits as well as limits" ↔ the curvature channel has both halves

sim08's density cap had only the limit half — it pruned nucleation but did not feed back into maintenance, so stability didn't rise. The Facchini curvature channel has BOTH halves: depositing at a convex tip extends the tip (recruits further building there — positive feedback through roughness, which focuses evaporation further), AND the smoothing term limits feature size. This is exactly the "non-saturating channel that RECRUITS as well as LIMITS" that H7's Session-13 refinement called for. The curvature channel is not just a non-saturating inhibitory channel (H11) — it is a non-saturating channel that also self-amplifies, which is what the crossing needs the structure to do (recruit its own maintenance). Candidate: if the d instability is the crossing, sim09 would unify the directed-transport and non-saturating-inhibition candidates into one mechanism, as queued-topic 58 predicted.

No cement pheromone ↔ H11 corroborated at the level of sufficiency

Facchini 2024 explicitly state "experiments do not support a role for a putative cement pheromone." This is now two independent groups (Calovi 2019, Facchini 2024) plus a curvature-only model that reproduces real morphology without any pheromone. H11's flag on the saturating channel is no longer just "biology doesn't use it" (absence) — it is "biology doesn't need it" (sufficiency). The saturating cue the Grassé modeling lineage (Deneubourg → Bonabeau → Ladley → sim06) assumed is not just unused; it is unnecessary to reproduce the target phenomenon. This raises the stakes for sim09: if curvature alone crosses, the saturating pheromone channel sim06/sim07 used was not just suboptimal but the wrong substrate entirely.

Morphology ≠ crossing ↔ the open risk for sim09

Facchini's curvature model reproduces nest geometry (pillars, walls, branching) — the consolidation morphology — but it does not test self-maintenance, persistence against erosion, or perturbation repair. Reproducing the morphology is necessary but not sufficient for the trace→actor crossing. sim09 must layer H7's three operational criteria (stability, non-reducible dynamics, constraint on agents) and the perturbation/self-repair test onto the curvature growth model. The risk: curvature may consolidate morphology (like sim08's cap did) but still not fire the crossing, if the smoothing term limits growth without recruiting maintenance specifically. The Facchini roughness feedback (deposits roughen the surface, focusing further deposition) is the candidate maintenance mechanism — but it must be tested, not assumed.

2026-07-30 — Session 15 (sim09 DESIGN.md authored — the curvature channel gets an implementation spec)

The Facchini growth equation ↔ sim09's three channels made operational

The curvature channel that H7's Session-13/14 refinement identified as "the non-saturating channel that recruits as well as limits" now has a concrete, Part-by-Part implementation spec at simulations/sim09_curvature_channel/DESIGN.md (9 Parts, mirroring sim06's proven structure). The Facchini 2020 growth equation ∂f/∂t ≈ f(1−f)·[(1/2)·Δf + d·Δ²f] is adapted to sim06's 2D grid+agent framework — each of its three terms becomes an operational piece:

• the growth term (1/2)·Δf (mean curvature) → compute_curvature (half the Laplacian of a lightly-smoothed material field), driving a linear, non-saturating deposit-probability routing for loaded termites at convex tips (the recruit mechanism);
• the smoothing term d·Δ²f (biharmonic) → the d-gated field_step smoothing, the LIMIT mechanism and the phase-transition knob (sim09's analog of sim07's M_c);
• the prefactor f(1−f) → an on_surface Moore-dilation mask restricting deposits to the structure surface (spatial selectivity without a saturating cue).

The Facchini/Calovi convex-concave resolution ↔ sim09's state-gated action split

The single most important design constraint in the DESIGN: sim09 must split deposit (loaded termites at convex tips) from excavate (unloaded termites at concavities). Conflating them — as a single "build" action, as sim06 did — would invert the rule's sign. The Facchini 2024 (deposition at convex tips) vs Calovi 2019 (aggregate activity at concavities) contradiction is resolved as different action components, and sim09 makes that resolution operational via state-gating. This is a methodological lesson carried from the literature into the model: a "construction" rule that does not separate the action components can get the sign backwards.

Roughness as the recruit proxy ↔ the channel-adapted crossing criterion 2

sim06's crossing detector bug (criterion 2 unsatisfiable under Grassé positive feedback) was fixed by requiring mass saturation. sim09 adapts criterion 2 to the curvature channel: roughness (the std of curvature over the structure surface) sustained above a threshold while mass saturates — the curvature analog of "the field stays energized by the structure's own shape, not by ongoing fresh deposits." This is the recruit channel's self-sustenance made measurable. Roughness is also the Facchini 2024 positive feedback (deposits roughen the surface, focusing further evaporation/deposition) rendered as a scalar metric.

The d phase transition ↔ the H7 prediction made operational

sim09's headline deliverable is Part 7's d sweep: if the crossing fires only above the Facchini curvature-instability threshold d* and not below it, d is to sim09 what M_c was to sim07 — but with a mechanism that recruits where the scalar transport only dispersed and a non-saturating channel where the density cap only limited. If that transition exists, sim09 unifies the directed-transport and non-saturating-inhibition candidates (queued-topic 58) into one mechanism, as the curvature channel is the minimal lumped form of directed geometry. If it does not, the null is sharper than sim08's: the curvature channel has both the recruit and limit halves, so a null would mean the crossing needs something beyond even the full H7 prescription.

A regression guard ↔ learning from sim06's detector bug

sim06's original crossing detector could not fire (criterion 2 required deposits to fall below their early-run average, impossible under Grassé positive feedback). sim09's Part 5 carries a synthetic-history regression guard: the detector must fire on an all-true history and withhold when any single criterion is negated. This encodes the lesson from the 2026-07-27 code review as a test — the detector is validated against satisfiable AND unsatisfiable synthetic inputs before being trusted on real runs.

DESIGN authorship without Opus ↔ the bottleneck rule

The memory convention is "Opus 4.8 writes DESIGN.md, GLM implements." Opus was not running tonight (cron context), and the DESIGN.md was the single bottleneck blocking all sim09 implementation — no Part could begin without it. I authored it from the complete Session-14 grounding (Facchini 2020/2024, H7's criteria, the sim06 DESIGN template) rather than leave the night idle. The spec follows sim06's exact structure (9 Parts, verification commands, progress tracker, appendices) precisely because that structure is what makes nightly GLM implementation safe. If an Opus pass later wants to revise it, the revision is cheap; the blocking is not.

2026-07-31 — Session 16 (sim09 implementation Parts 1–7: the curvature channel runs, the phase transition needs tuning)

sim06 fully complete ↔ the path to sim09 is open

A pre-session check of sim06's Progress Tracker found all 9 Parts marked [x] — the termite-mound saturating-cue sim is done (and its corrected near-miss / crossing-confirmed-in-57%-of-parameter-space result stands). That cleared the way to sim09, the curvature-channel sim the previous two sessions grounded and specified. Tonight implemented Parts 1–7 of sim09's DESIGN.md in a single session (the cron instruction to continue to the next Part when budget remains overrides the DESIGN's "one Part per session" rule). All selftests pass (Part 1 OK … Part 7 OK) and run produces a valid results.json with both conditions.

The Facchini growth equation ↔ sim09's three terms made operational

Each term of ∂f/∂t ≈ f(1−f)·[(1/2)·Δf + d·Δ²f] is now running code:

• (1/2)·Δf (mean curvature) → the recruit mechanism. compute_curvature returns half the Laplacian of a lightly-smoothed material field; termite_step routes loaded termites to deposit at convex tips via a LINEAR (non-saturating) probability p = base + gain·curvature, clamped to [0,1]. This is the H11 prescription made concrete — the deposit response does not flatten above a threshold the way sim06's φ/(1+φ) did.
• d·Δ²f (biharmonic) → the limit mechanism + phase-transition knob. field_step applies d·0.0001·Δ²f each step; d is sim09's analog of sim07's M_c. Part 7's d sweep is the headline phase-transition plot.
• f(1−f) (surface restriction) → compute_on_surface. A Moore-dilation of the structure mask; deposits outside the surface fall back to a low nucleation base so the first pillars can seed.

The Facchini/Calovi action-component split ↔ the state-gated deposit/excavate rule

The single most important design constraint — do not conflate deposit with excavate — is operational. Loaded termites deposit at convex tips (Facchini 2024); unloaded termites excavate at concavities (Calovi 2019). sim06 had only deposit; sim09 splits the action. Conflating them would invert the rule's sign. The selftest's synthetic Gaussian bump confirms deposits land on the convex rim.

The regression guard ↔ sim06's detector bug encoded as a test

sim06's original crossing detector could not fire (criterion 2 required deposits to fall below their early-run average, impossible under Grassé positive feedback). sim09's Part 5 carries a synthetic-history regression guard: the detector must fire on an all-true history and withhold when any single criterion is negated — for BOTH the curvature and baseline channels. This encodes the 2026-07-27 code-review lesson as an executable test. It passes.

The honest null-so-far ↔ the phase transition needs parameter tuning

At DEFAULT parameters, sim09's d sweep finds NO phase transition. The curvature channel saturates the grid (pillars=1, retention=1.0 at every d) because the nucleation base (0.10) floods the grid before curvature routing can create spatial selectivity — 200 termites × 4000 steps × 0.10 ≈ 80k deposits into a 10k-cell grid. The crossing detector does not fire: criteria 1 (stability ≥0.90) and 3 (deposits_on_convex ≥0.60) pass comfortably, but criterion 2 (roughness ≥0.02 AND mass saturating, i.e. |growth_rate| < 0.01) fails because mass never saturates — the grid fills and stays filled. Quick tuned probes (deposit_prob_base=0.01, material_decay=0.002) show the predicted consolidation DIRECTION (pillars 25→2 as d rises 0→4) and a roughness spike at the biharmonic instability, confirming the mechanism's sign is right — but the mass-saturation gate in criterion 2 is hard to satisfy while the structure is still accreting. Finding the parameter regime that reveals the phase transition is the remaining scientific work. The DESIGN explicitly allows reporting a null honestly: "a null result is still a result, but first try to find parameters that reveal the mechanism." This is the spiral-loop methodology in action — the sweep ran, the mechanism's direction is visible, and the tuning question is now sharp (lower nucleation + higher erosion so mass saturates before the grid fills, and a stable high-d range — d=8 showed a numerical blowup of the explicit biharmonic, so the 0.0001 prefactor needs reducing for the upper sweep range).

sim09's partial result so far ↔ H7 and H11

The curvature channel's consolidation DIRECTION is confirmed (pillars decrease as d rises, the opposite of sim06's saturating-cue fragmentation and sim07's scalar-transport fragmentation). The crossing itself has not fired at the parameters tried. This is consistent with H11 (the non-saturating channel consolidates where saturating channels fragmented) but not yet a positive test of H7 (the crossing needs the recruit half to drive maintenance, not just morphology). Parts 8 (perturbation/self-repair — the recruit half's acid test) and 9 (viz+README) remain, and the parameter tuning to find d* is the next session's priority.

2026-08-01 — Session 17 (sim09 Part 8: the perturbation acid test runs, recovery needs the saturating regime)

The saturating rule's unbounded accumulation ↔ H11's failure mode in a new metric

sim09's Part 8 perturbation experiment damages the structure at step 0.6×steps and measures recovery = current_total_material / pre_perturb_total_material. At default params the baseline (saturating deposit rule) "recovers" to 47× — but this is unbounded material accumulation, not targeted repair: the saturating rule piles material without an erosion balance, so total_material grows ~47× from the early pre-damage sample. The curvature channel recovers to 1.13× (it saturated the grid at 10000/10000 cells before damage and refilled the hole). This is H11's failure mode (a saturating channel cannot express the spatial contrast targeted repair needs) showing up in a second, independent metric beyond morphology fragmentation. The perturbation test gives H11 a repair-side line of evidence to match its morphology-side line.

Repair needs the same regime the crossing needs ↔ the two open questions unify

Part 7 found the d phase transition needs mass-saturation (lower nucleation, higher erosion) so the biharmonic instability can create spatial selectivity before the grid fills. Part 8 finds the repair/crossing separation needs the same regime: in the tuned probe (deposit_base=0.01, material_decay=0.002) the curvature channel saturates and refills the damage hole to 1.01× (repair-like), while the baseline grows unboundedly to 4.55× (volume, not repair). The separation is directionally right but the recovery metric conflates "targeted repair at the scar" with "volume restoration / continued growth." The clean H7 separation — curvature recruits repair, baseline does not — requires the mass-saturating regime. This suggests the crossing and self-repair are two faces of one phenomenon (the structure's self-maintenance), and a single broad parameter sweep should reveal both together. The two gaps point at the same tuning, which is itself a finding: the recruit half's acid test and the crossing detector are not independent experiments — they are the same experiment measured two ways.

A spatially-targeted recovery metric ↔ the acid test's missing precision

The grid-wide recovery ratio (following sim06/sim08) cannot distinguish "repair at the scar" from "continued growth elsewhere." A spatially-targeted variant — recovery measured in the damaged patch specifically — would make the acid test decisive without needing the full mass-saturating regime. This is a candidate post-Part-9 refinement: the DESIGN's grid-wide ratio was the safe minimal choice (matching the proven sim06/sim08 pattern), but the result shows the minimal metric is not sharp enough to settle H7 on its own.

2026-08-02 — Session 18 (sim09 Part 9 + completion: the curvature channel ships, the crossing is a parameter-regime question)

sim09 fully implemented ↔ the curvature channel runs end-to-end

sim09's Part 9 (visualize.html + README.md) completes the curvature-channel sim — all 9 Parts of the DESIGN.md are now [x]. The visualization is a self-contained HTML5 Canvas page (dark theme #0d1117) fetching results.json and the optional output/sweep_data.json, rendering four charts (structure-over-time, roughness+ stability, deposits-on-convex criterion 3, perturbation recovery) plus a d-sweep phase-transition panel, summary boxes, and a full result table. The README fills in the real numbers from the default-param run and states the honest partial result. The verification command passes (all selftests OK, run produces results.json, both artifacts present, local http server returns 200 for page/results/sweep with no console errors).

The Facchini growth equation ↔ fully operational, both halves present

The curvature channel has BOTH halves H7's Session-13 refinement required: recruit (loaded termites deposit at convex tips via a linear, non-saturating p = base + gain·curvature, extending the tip) AND limit (the d-gated biharmonic smoothing caps feature size). The Facchini/Calovi action-component split is operational (loaded→deposit at convex, unloaded→excavate at concave), the f(1−f) surface restriction is a dilation mask, and roughness is the recruit-proxy crossing criterion 2. This is the cheapest remaining candidate that could actually cross — it has both halves where sim08's density cap had only the limit half.

Grid saturation ↔ criterion 2's mass-saturation gate cannot fire

At default params (d=1.0, deposit_prob_base=0.10) the curvature channel grid-saturates (10000/10000 cells, retention 1.0) because the nucleation base floods the 10k-cell grid before curvature routing can create spatial selectivity. Mass never plateaus, so crossing criterion 2 (roughness sustained while mass saturates, i.e. |growth_rate| < 0.01) cannot fire — the d-sweep [0…8] finds no phase transition (pillars=1 at every d). This is NOT a mechanism failure: tuned probes (deposit_prob_base=0.01, material_decay=0.002) show the predicted consolidation direction (pillars 25→2 as d rises 0→4, plus a roughness spike at the biharmonic instability). The mechanism's sign is right; the parameter regime is wrong.

H11's direction replicated in a 4th mechanism ↔ the saturation diagnosis holds

sim09's tuned-probe consolidation (pillars ↓ as d ↑) is the opposite of sim06's self-maintenance fragmentation (219–297 components) and sim07's scalar transport fragmentation (57→128). That is H11's direction replicated in a fourth independent mechanism: non-saturating channels consolidate where saturating cue-field channels fragmented. The perturbation test gives H11 a repair-side line of evidence to match its morphology-side line: the baseline's 47.34× "recovery" is unbounded material accumulation (the saturating rule piles material without an erosion balance), exactly the failure mode H11 flags — a saturating channel cannot express the spatial contrast targeted repair needs.

The crossing is now a parameter-tuning question, not an open-mechanism question

The four-mechanism arc (sim06 saturating cue → sim07 scalar transport → sim08 non-saturating cap → sim09 non-saturating recruit+limit curvature) has narrowed H7 to a sharp claim: the curvature channel has both halves, so if the crossing fires anywhere it should fire here. The remaining blocker is parameter-regime, not mechanism — find the mass-saturating regime (lower nucleation + higher erosion) where the biharmonic instability creates spatial selectivity before the grid fills, and add a spatially-targeted recovery metric distinguishing scar repair from volume restoration. The next session's priority is a broad deposit_prob_base × material_decay × d sweep in that regime to locate d*. If the crossing fires only above the Facchini d* instability, sim09 unifies the directed-transport and non-saturating-inhibition candidates (queued-topic 58): curvature IS the minimal lumped form of directed geometry.

The crossing and self-repair are one phenomenon measured two ways

Part 7 found the d phase transition needs mass-saturation; Part 8 found the repair/crossing separation needs the same regime. The two gaps point at the same tuning, which is itself a finding: the recruit half's acid test (perturbation recovery) and the crossing detector are not independent experiments — they are the same experiment measured two ways. A single broad parameter sweep should reveal both together.

Session 19 (2026-08-03) — the unfalsifiable gate, and the crossing that fires with a control

The d* sweep (100 combos: deposit_prob_base × material_decay × d) returned 0/100 crossings. The per-criterion diagnosis was unambiguous: criterion 2's mass-saturation gate (|material_growth_rate| < 0.01) passed in 0/100 combos — mean_late_mgr was 0.4–3.7, never near 0.01. Criteria 1 (stability), 2r (roughness), and 3 (constraint) all passed at the low-decay corner. The gate was the single universal blocker.

The metric-ceiling bug — the sim06 detector lesson, repeating

The mass-saturation gate used the per-sample-window |Δtotal_material|/sample_every < 0.01. For a 150-termite stochastic deposit process that quantity has a Poisson noise floor of ~0.5–1.0 (the centered window-sum's std / window), ~100× above the 0.01 threshold. No finite-population run can ever pass it. The gate was unfalsifiable: the detector could not fire regardless of the mechanism. This is the same failure mode as sim06's original deposit-rate gate (which could not fire because Grassé positive feedback makes deposit probability rise) — a threshold set below the noise floor of the quantity it gates on. The Session 17 conclusion ("the crossing is a parameter-regime question, not a mechanism question") was itself suspect: the regime where mass "saturates" below 0.01 does not exist for any finite N.

Correction: replaced the per-window absolute-growth gate with a relative-slope plateau: |slope(total_material over last K=16 samples)| / mean(total_material) < 0.001. The regression slope averages over 400 steps, suppressing the Poisson window noise; the relative (scale-invariant) form sits above the noise floor (it fires ~98–100% in the late equilibrium of a plateauing run, while the absolute gate fired 0%). The selftest regression guard was updated to negate the plateau explicitly (a ramp instead of a flat trajectory withholds the crossing).

The crossing fires with a control arm — H11's channel distinction goes causal

In the tuned probe (dpb=0.01, decay=0.002, 80×80 grid, 2000 steps — non-saturating, cells 3123–5754/6400):

• Curvature channel crosses at every d ∈ [0, 4]; crossing_step decreases monotonically 1550 → 900 as d rises (d speeds consolidation); n_pillars falls 12 → 1 (consolidation, H11's direction); roughness rises 0.44 → 0.77.
• Baseline-pheromone control (same detector) crosses in 0/3 — criterion 2's pheromone-elevation gate fails (mean_pheromone 0.25 < 0.50 threshold; the saturating rule never elevates the cue enough).

This is the first time the H7 crossing has fired with a control arm that does not. H11's channel distinction (non-saturating action-channel vs saturating cue-channel) is now the causal variable separating the crossing from the non-crossing, not merely a directional correlate of morphology. Previously H11 rested on a same-direction comparison within one model family (non-saturating channels consolidated where saturating ones fragmented, but neither crossing fired); now the control is run under the same corrected detector and fails the crossing where the curvature channel passes it.

Honest limitation — the recruit half drives the crossing, not the limit half

The crossing fires at d=0 (no biharmonic smoothing — the curvature channel's LIMIT half is off), so the detector is catching the recruit half (curvature routing + mass plateau), not the recruit+limit combination the Session-13 refinement specified. The d-smoothing controls morphology (pillars 12 → 1) and crossing speed (1550 → 900) but is not necessary for the crossing verdict. The honest claim narrows: the curvature channel's non-saturating recruit half is sufficient for the crossing; the limit half consolidates the morphology. This is still a real result — the baseline control (saturating cue, no curvature routing) does not cross — but it is a weaker claim than "recruit+limit both required." The next test isolates the halves: a recruit-only condition (curvature routing, d=0) vs a limit-only condition (d-smoothing, no curvature routing).

The crossing's "parameter-regime" blocker was a detector bug, not a regime

Session 17 framed the remaining work as "find the mass-saturating regime." The ceiling analysis shows that regime does not exist at any finite population for the absolute-growth gate — the Poisson noise floor scales with √N, so larger grids and more termites make the problem worse, not better. The corrected relative-slope plateau fires in the existing tuned-probe regime (dpb=0.01, decay=0.002) without any new parameter search. The lesson: when a detector fails across an entire parameter sweep, compute the metric's ceiling before concluding the mechanism is wrong — the gate may be unfalsifiable. This is the methodology rule "Compute your metric's ceiling. If its maximum can fall below your threshold, it is not a test," now earned twice (sim06's deposit-rate gate, sim09's mass-saturation gate).

Session 20 (2026-08-04) — The recruit half is load-bearing + almost-sufficient; the limit half is a stability amplifier

The 2×2 factorial that isolated the curvature channel's two halves

Session 19 found the crossing fires at d=0 (no biharmonic smoothing), which means the recruit half (curvature routing + mass plateau) drives the verdict and the limit half (d-smoothing) is not necessary for it — but the two halves were not yet isolated. Session 20 ran a 2×2 factorial over the curvature channel: recruit ON (curvature routing: curve_follow=0.6, deposit_prob_gain=0.85, excavate_prob_gain=0.60) vs OFF (curve_follow=0, deposit_prob_gain=0, excavate_prob_gain=0 — agents random-walk and deposit/excavate at base rates only; the field's curvature has no influence on agent action); limit ON (d>0, biharmonic smoothing in field_step) vs OFF (d=0). Four cells: recruit-only (d=0), recruit+limit (d>0, the as-built channel), limit-only (d>0, no recruit), neither (d=0, no recruit). A seed-robustness pass ran the four corners across seeds {42, 7, 123, 256}.

A new stable_crossed metric separates stable from transient crossings: late_hold_rate = fraction of the last 1/4 of records where all three crossing criteria hold simultaneously. A stable crossing holds ~1.00; a transient crossing (criteria flicker on and off) holds <0.55. stable_crossed = crossed AND late_hold_rate >= 0.90.

The result

Seed robustness (stable_crossed / total across 4 seeds):

• recruit-only (d=0): 3/4 stable (hold [1.0, 1.0, 0.65, 1.0] — seed 123 is borderline, hold 0.65, still crosses)
• recruit+limit (d=1): 4/4 stable (hold 1.0 across all four seeds)
• limit-only (d=1): 0/4 stable (crossed in 2/4 but transient; hold [0.55, 0.50, 0.55, 0.40])
• neither (d=0): 0/4 stable (crossed 0/4; hold ≤0.15)

The recruit half is necessary and almost-sufficient for a stable crossing. Neither (no recruit, no limit) crosses in any seed. Limit-only (no recruit) is never stable. The recruit half alone crosses in 4/4 seeds and is stable in 3/4. The decisive contrast is recruit ON vs OFF at d=0: same detector, same regime, only the recruit flag differs — recruit-only crosses stably (3/4); neither does not (0/4).

The limit half is a stability amplifier, not morphology-only

The Session-19 "half-supported" reading was that the limit half is morphology-only — it consolidates pillars (12→1) and speeds the crossing (1550→900) but is not necessary for the verdict. Session 20 upgrades this: recruit+limit is stable in 4/4 seeds where recruit-only is 3/4 — the one borderline seed (123, hold 0.65) becomes fully stable (hold 1.0) when d>0 is added. The limit half amplifies the stability of the recruit-driven crossing against seed variance. So H11's "recruit as well as limit" is refined: recruit = necessary and almost-sufficient; limit = stability amplifier + morphology optimizer (not strictly necessary, but causally contributing to robustness). This is a stronger claim than "half-supported": the limit half has a causal role (stability), not merely an aesthetic one.

The limit-only transient flicker is itself informative

Limit-only's criteria 1 (stability) and 2 (roughness + plateau) mostly pass — the biharmonic does build roughness and mass does plateau — but criterion 3 (deposits_on_convex_fraction ≥ 0.60) flickers because without curvature routing, deposits land on convex cells only at the base rate. The smoothing creates convex features but nothing routes agents to them. The biharmonic alone builds the geometry the recruit channel would act on, but without the recruit half the geometry is unattended. This is the clean separation: the recruit half routes agent action to the geometry; the limit half shapes the geometry. Limit-only shapes geometry that no agent is routed to; neither alone (no geometry shaping) produces nothing.

The "find d*" question is now fully superseded

Under the corrected detector the crossing fires at every d ∈ [0,4] — there is no sharp d* phase transition for the crossing verdict. Session 19 replaced "find d*" with "isolate the recruit and limit halves"; Session 20 completes that isolation. The recruit half is the load-bearing variable; the limit half makes its crossing robust. The next question is composition: do two self-maintaining curvature structures compose (the sim05 L2 question reopened with a non-saturating stigmergic glue)?

Session 21 (2026-08-05) — The saturating-action control: action-based is primary, non-saturating is secondary

Session 20 isolated the recruit and limit halves but left H11's central confound intact: the recruit half is action-based (curvature routes what the agent does) AND non-saturating (linear gain) simultaneously. H11 says both properties matter, but the evidence (sim08 cap, sim09 curvature) had them confounded — both were present together in every crossing condition.

Session 21's saturating-action control disentangles them. The same curvature routing, but a saturating response p = base + gain·c/(1+|c|) instead of the linear p = base + gain·c. Both forms are action-based; only the linear form is non-saturating. Curvature in the running sim ranges ±1.5, so the saturating form genuinely compresses (at c=1.0: 0.425 vs 0.850).

The verdict: action-based is the primary load-bearing property; non-saturating is a secondary stability contributor. The saturating action still crosses in 8/8 recruit-ON seeds and is stable in 6/8 (linear is 7/8). The limit half (d=1) rescues both to 4/4 stable. Saturation costs ~0.05 in mean hold rate at d=0 (0.91→0.86) but does not collapse the crossing the way turning off the recruit half does (0/8 crossed).

What degrades is the mass-plateau gate (criterion 2p), not the routing (criterion 3). The saturating form's criterion 3 (deposits_on_convex_fraction) passes 1.00 in all seeds — curvature routing still sends deposits to convex tips even with the compressed response. What flickers is criterion 2's mass-plateau gate: the saturating form takes longer to plateau (compressed deposit probabilities create more stochastic scatter in the mass trajectory), so |slope(M)|/mean(M) stays above threshold more often. The degradation is in the dynamics of mass equilibration, not in the spatial selectivity of the routing.

This partially weakens H11's strict "non-saturating" claim. H11 says feedback through a saturating channel is self-defeating because it destroys spatial contrast. The saturating action-based channel does not destroy spatial contrast — it only slows mass equilibration. The non-saturating property matters for stability, not for the crossing verdict. H11's distinction should be refined: the critical property is action-based routing (curvature routes what the agent does, not how strongly it reads a cue); non-saturating is a stability amplifier, analogous to the limit half's role. "Self-defeating" overstates the saturating action's failure; it is "self-destabilizing" at most.

A three-level decomposition of the crossing's causal structure

Sessions 19–21 now decompose the crossing into three levels of causal contribution:

1. Action-based routing (Session 20, primary). Curvature routes what the agent does (deposit at convex tips, excavate at concavities) rather than how strongly it reads a cue. This is the load-bearing variable: without it (recruit OFF), 0/8 seeds cross regardless of the response curve or the limit half. The baseline-pheromone control (cue-based, no action routing) crosses 0/3.

2. Non-saturating response (Session 21, secondary). Within the action-based family, a linear (non-saturating) response curve makes the crossing more stable than a saturating one (7/8 vs 6/8 stable at d=0). But both forms cross; the saturating action does not collapse. The non-saturating property is a stability amplifier, not the causal variable.

3. Biharmonic smoothing / the limit half (Session 20, tertiary). The d-smoothing term stabilizes the crossing against seed variance (recruit+limit is 4/4 stable where recruit-only is 3/4) and consolidates morphology (pillars 12→1). But it is not necessary for the crossing verdict and alone (without routing) never crosses stably (0/4).

This is a cleaner causal story than H11's original "saturating channels are self-defeating." The self-defeating property belongs to cue-based saturating channels (sim06/sim07's pheromone field), not to action-based saturating channels. The action-based property is what separates crossing from non-crossing; the non-saturating and limit properties are stability amplifiers that make the crossing robust.

Session 22 (2026-08-06) — The cue-based non-saturating control: the 2×2 completes, and the non-saturating property reverses sign across families

Session 21 tested within the action family (linear vs saturating action routing) and found non-saturating is a secondary stability amplifier there. The remaining cell of the 2×2 was untested (queued-topic #67): a non-saturating cue channel. sim06's as-built deposit rule is the saturating cue p = base + gain·φ/(1+φ) (flat above φ≈1); the non-saturating cue is p = base + gain·φ (clamped to 1.0). Both are cue-based; only the response curve differs. A deposit_response parameter was added to sim06.py with a selftest Part 5d confound-isolation guard.

The non-saturating cue crosses LESS, not more — the opposite of the action family and opposite to H11's strict prediction. Seed-42 factorial (64 conditions): saturating cue crosses 32/32 (stable 32/32, hold 1.000); linear cue crosses 19/32 (stable 16/32, hold 0.527). Decomposed by self-maintenance: without SM, saturating cue is 16/16 stable (hold 1.000); linear cue is 0/16 stable (hold 0.053). With SM, both are 16/16 stable (hold 1.000). Seed robustness (4 seeds) confirms: saturating no-SM 4/4; linear no-SM 0–1/4; both with SM 4/4. Determinism verified.

The mechanism: deposit-probability clamping, not cue-response compression. The linear rule p = base + gain·φ hits p=1.0 at φ≈1.15 — every high-pheromone cell deposits at 100%, driving faster, more uniform growth (linear builds 3624 vs saturating's 1858 cells) and diluting the pheromone field. Mean pheromone over structure drops to 0.467 (below the 0.5 crossing threshold) vs the saturating cue's 0.749. The saturating cue's φ/(1+φ) compression prevents deposit-probability saturation, keeping the response graded and preserving spatial contrast. Threshold sensitivity confirms: at phero_elev_thresh 0.3–0.4 the linear cue crosses (hold 1.000); at 0.5+ it does not — the 0.467 is a real equilibrium, not a detector artifact.

The non-saturating property reverses sign across families. In the action family (sim09 Session 21), non-saturating (linear) is slightly more stable (7/8 vs 6/8). In the cue family (sim06 Session 22), non-saturating (linear) is dramatically less stable (0/16 vs 16/16 without SM). The full 2×2:

The action-based property is primary (both action rows cross); the non-saturating property is a sign-reversing modifier — a stability amplifier in the action family, a stability destroyer in the cue family (without compensation).

H11's "self-defeating" framing is backwards for the cue family. H11 said saturating channels are self-defeating because they destroy spatial contrast. Session 21 found this is backwards within the action family (the saturating action is merely less stable, not self-defeating). Session 22 finds it is backwards for the cue family too, but in the opposite direction: the self-defeating channel is the non-saturating cue, not the saturating cue. The non-saturating cue's linear response clamps to p=1.0 at φ≈1.15, flattening the gradient; the saturating cue's compression preserves the gradient. The "saturation" that is self-defeating is the deposit-probability clamping (which the linear cue hits), not the cue-response compression (which the saturating cue has). H11's original framing conflated these two kinds of saturation.

Self-maintenance rescues the non-saturating cue (4/4 stable, hold 1.000). The structure-reemits-pheromone loop sustains pheromone elevation regardless of the response curve, compensating for the linear cue's gradient-flattening. So the non-saturating cue is not categorically unable to cross — it needs a separate mechanism to sustain the pheromone field the saturating cue sustains on its own. This connects to H7's self-maintenance loop: the two self-maintenance failures (sim06 fragmentation, sim07 transport) acted through the saturating cue and fragmented; but with a non-saturating cue, self-maintenance becomes necessary for the crossing rather than counterproductive.

The cue-action asymmetry as a cross-domain connection

The sign reversal connects to the distinction between reading a field and acting on a gradient. In the action family, the agent's response is routed by the cue (curvature selects deposit vs excavate); the response curve only modulates the gain, so saturation compresses the gain without destroying the routing. In the cue family, the agent's response is the field (pheromone level → deposit probability); the response curve is the channel, so saturation of the response is saturation of the channel. The non-saturating cue is self-defeating because its response curve saturates at the probability level (p=1.0) before the field develops the spatial contrast the crossing needs. The non-saturating action is stable because its response curve saturates only the gain, not the routing decision.

This reframes H11: the critical distinction is not "saturating vs non-saturating" but "does the response curve saturate the probability (cue family: self-defeating if non-saturating) or only the gain (action family: self-defeating if saturating)?" The 2×2 is now the evidence.

Session 23 (2026-08-07) — the φ_sat predictor fails; spatial contrast survives via routing, not deposit probability

The φ_sat predictor (queued-topic #72) proposed a unifying scalar: the deposit-probability saturation threshold (the input value at which p_deposit first reaches 1.0). If the operating max of the routing input exceeds φ_sat, the channel is probability-saturated and the crossing should fail; if below, it should fire. This would unify all four cells of the 2×2 with a single number.

A direct probe (phi_sat_probe.py) of sim06 (cue) and sim09 (action) at their crossing-proven regimes found the predictor is 50% accurate — no better than chance. It correctly predicts the cue family (saturated→fails, unsaturated→crosses) but fails for the action family: the action/linear condition IS saturated (max curvature 2.55 > c_sat 1.165) but STILL crosses stably. The clamping fraction is tiny everywhere (0–7%); the cue/linear has 6.9% clamped cells and fails, the action/linear has 1.0% and crosses.

The cross-domain connection: spatial contrast has two carriers — the deposit probability and the routing decision — and only one of them is destroyed by saturation. In the cue family, the deposit probability IS the spatial signal — clamping it to 1.0 on high-cue cells flattens the gradient. In the action family, the spatial information lives in the routing decision (which direction the agent moves), not the deposit probability. The response curve saturates the gain (how hard to deposit), not the routing (where to go). This is the same distinction as Session 22's "reading a field" vs "acting on a gradient," now made quantitative: the φ_sat predictor treats the deposit probability as the sole carrier of spatial information, which is true only for cue-based channels.

The unifying diagnostic is not φ_sat but whether spatial contrast in the routing input survives the response curve — and that depends on the channel architecture (action preserves routing under saturation; cue does not), not just the saturation threshold. This connects to the general principle that a self-defeating channel is one where the feedback signal and the spatial signal travel on the same wire: saturating one destroys the other. In the action family they travel on separate wires (routing vs deposit gain), so saturating one leaves the other intact.

Session 24 (2026-08-08) — Stability is not self-repair: the control arm that separated scar from growth

The perturbation acid test (sim09 Part 8, Session 17) reported the curvature channel "recovers to 1.13×" vs the baseline's 47.34× — but the grid-wide recovery = total_material / pre_perturb_total credits material accumulated anywhere as "recovery." The baseline's 47× was unbounded material accumulation, not targeted repair. Session 24 implemented the spatially-targeted recovery metric (queued-topic #60, open since Session 17) with a control arm: patch_recovery (material in the scar / pre-damage scar material) and mirror_recovery (material in an undamaged same-size region / its pre-damage material). targeted_repair = patch_recovery − mirror_recovery isolates preferential scar repair from background growth.

The result: targeted_repair is negative in all four conditions. Neither the curvature channel nor the baseline preferentially repairs the damage site. The scar grows slower than an undamaged mirror region in every case — re-nucleation from zero is slower than continued growth on existing structure. The crossing fires (stability, roughness, mass-plateau gate) but the structure does not self-repair in the targeted sense.

The cross-domain connection: "stable" and "self-repairing" are different claims, and a metric without a control arm conflates them. This is the same methodology lesson as the mass-saturation gate (Session 19) and the φ_sat predictor (Session 23): a metric that cannot distinguish the phenomenon it claims to measure is not a test. The grid-wide recovery could not distinguish "the scar healed" from "the structure kept growing elsewhere" — and without the mirror control, the patch_recovery alone could not distinguish "the scar healed preferentially" from "everything grew, including the scar." The control arm is what makes the measurement a test rather than a description.

This connects to the broader principle running through the project's methodology corrections: every detector or metric must be shown to distinguish the phenomenon from its confounds, not merely to respond to it. The crossing detector responded to stability; the recovery metric responded to growth; the φ_sat predictor responded to saturation. Each needed a control to become a test: the baseline-pheromone control for the crossing, the mirror patch for the recovery, the action/linear condition for φ_sat. A metric that responds is a description; a metric that distinguishes is a test.