149 Predictions Across Three Epistemic Tiers

BMC generates 149 concrete predictions across 23 categories. Not all 149 are “falsifiable predictions” in the strict sense — and we no longer present them as one undifferentiated count. That conflation was a fair criticism of earlier versions of this page. Predictions sit on two orthogonal axes:

Epistemic status — what kind of claim is it?

T — formal result within the framework. Follows analytically from BMC definitions (e.g. the Gödel-analog $$SIT_{min}>0$ , the $$M \gg G$ lower bound). Constrains the model; not evidence for BMC over rival theories.
R — retrodiction. Consistent with findings established in the literature before BMC (e.g. PFC maturation vs reflexion onset). Shows consistency, not differential support.
P — prospective. A risky claim with a stated measurement method and falsification threshold. Only the P-tier carries confirmatory weight, and only the P-tier is “falsifiable” in the Popperian sense.

Testability — how can it be tested? The A / B / B* / C / Open tag on each item below (existing data, meta-analysis, or new experiment). This axis is orthogonal to epistemic status.

The number that matters: the prospective scorecard

A theory earns credit from risky bets that survived, not from the length of its prediction list. The credibility-bearing claims are in the P-tier — and specifically the subset that pits BMC against a named rival theory making opposite numerical predictions (separating tests).

	Count
Separating tests committed (named rival → opposite prediction)	11
Run outside the engine, counter-intuitive BMC side holding	1 (P-BM28 — in the tested non-cooperative regime only)
Verified inside the BMC engine (implementation check, not independent evidence)	2 (P-NM3, P-BM4)
Pending	8

Commitments, not results — the blind-twin screen. We ran these eleven through the same screen we apply to our own simulations: a claim does not count as a test if a cheaper account, blind to the BMC mechanism, predicts the same sign. Outcome: 8 of the 11 do not yet state a numeric falsification threshold; for 7, a mechanism-free alternative predicts the same direction, so a magnitude gap must be committed in advance before they can separate anything; 1 (P-RL1) names a rival that the literature has already refuted — a rival that cannot pass is the mirror of a control that cannot fail, and it is being reformulated. Two survive the screen as written (P-D2, P-WM1 — the latter flagged internally as under-specified). Per-bet detail, including the blind twin named for each, is in supplementary/reference/bets_registry_2026-08-29.md.

The headline result — and the one we are most willing to be wrong about — is P-BM28:

Separating test (run, held)	Competing hypothesis	Result	Source
P-BM28 — language is parasitic on working memory (signal memes optimized for transmission, not survival)	Reward-engineered learning (REINFORCE / PPO) predicts a survival advantage for language	BMC’s prediction held: $$\Delta_{alive} \approx 0$ across 10 experiments, $$N$ =8–150	DOI: 10.5281/zenodo.19181798

A null without a stated equivalence bound and power analysis is not confirmation; an equivalence test (TOST) and a margin remain to be specified. The result also holds in the tested regime only: the environment was non-cooperative by construction ( $$n_{cooperative}=0$ , $$n_{predators}=0$ ), and under cooperation with predators the direction reverses (ReceiveOnly mean energy 80.5 vs Blocked 75.0; Wilcoxon $$p = 0.10$ ; 14/20 seeds) — not significant at this power, but directionally against survival-neutrality.

BMC and the standard engineering alternative make opposite numerical predictions here; the counter-intuitive BMC prediction is the one that held. Two further separating tests are confirmed in-engine via ablation: P-NM3 (WM bandwidth constrains signal complexity; $$k$ =2 vs $$k$ =5) and P-BM4 (ontogenetic critical periods enable cultural ratchet; −ONTO ⇒ ceiling, not ratchet). See Computationally Verified. P-CN1, also listed in that section, is implementation fidelity rather than a differential test and is classified R.

P-tier and T-tier — the load-bearing items, by code

For transparency, here are the items the framework above classifies as P (prospective, falsifiable — the credibility-bearing tier) and T (formal results within the BMC formalism — analytic, not evidence vs rival theories). The remaining items below are R-tier: retrodictions that show the theory is consistent with established findings, or claims awaiting an operationalized method and threshold. R-tier items show consistency, not differential support; we do not badge them individually on each line to avoid noise, but the full per-item ledger is maintained as a working document.

P-tier (40 total: 37 in the list below + 3 in Computationally Verified)

Separating tests (11 — named rival makes the opposite numerical prediction):

P-BM28 — language parasitic on WM vs reward-engineered survival advantage. Run; BMC held.
P-NM3 — WM-bandwidth limit on signal complexity vs unconstrained. Verified in engine.
P-BM4 — ontogenetic critical periods produce ratchet vs ceiling under −ONTO. Verified in engine.
P-D2 — priming ∝ embedding proximity, not edge presence.
P-WM1 — FEAR ~1 WM slot, PLAY=0 (valence-specific vs arousal-only).
P-SAF4 — WM-limit biases absent in expanded-WM AGI (substrate artifact vs algorithmic necessity).
P-SAF5 — architecture alignment survives attacks that bypass RLHF (vs training-based alignment).
P-RL1 — TD-modulated edges beat pure Hebbian in delayed-reward tasks.
P-RL3 — multi-GVF beats single-reward in learning speed.
P-DD2 — per-target fragment libraries beat shared on dissimilar targets.
P-DD4 — SIT-modulated mutation rate beats fixed-rate baseline.

Other prospective predictions (29 — single-direction, with stated method + falsification threshold): P-M1, P-D1, P-N1, P-N2, P-N3, P-N4, P-SC1, P-CP1, P-CA1, P-CA3, P-SD2, P-SD4, P-SD6, P-SMR2, P-SMR4, P-SM7, P-H8, P-H12, P-PER5, P-TH1, P-TH5, P-FM2, P-FM6, P-DD1, P-DD3, P-SAF1, P-SAF2, P-IT2, P-RL2.

T-tier (14 — formal results within the formalism, not evidence vs rivals)

These follow analytically from BMC definitions; they constrain the model and serve as internal consistency checks, but do not constitute differential evidence for BMC over rival theories: P-11 (metric hierarchy from the CL composite), P-G1 ( $$CL_G$ ceiling from CL at $$M=\emptyset$ ), the $$M \gg G$ lower bound (underlying P-G4; the page-line of P-G4 itself is a retrodiction of mirror-test / metacognition distribution), P-NLD3 ( $$T_{insight} \approx 1.614\mu$ van der Pol), P-NLD4 ( $$\lambda_{BMC} \approx 0$ at $$\sigma_{SW} \approx 1$ ), P-CN2 (Hopfield capacity $$\sim \sqrt{|V_m|}$ ), P-CX1, P-CX2, P-CX3 (statistical-complexity and edge-of-chaos scaling), P-IT3, P-IT4, P-IT5, P-IT6, P-IT7 (rate-distortion, BLEND $$2^k/k$ , Bayesian-Occam — information-theoretic identities of the formalism).

Testability tag	Meaning	Count
A	Confirmed by existing published data	7
B	Testable with existing public datasets	10
B*	Testable via meta-analysis of existing studies	34
C	Requires new experiment or native BMC system	80
Open	SM/NM predictions awaiting empirical test	18

Four entries appear in Computationally Verified: three are P-tier separating tests with confirmed BMC outcomes (P-BM28, P-NM3, P-BM4); the fourth (P-CN1) is implementation fidelity rather than a differential test, and is classified R.

I. Core Dynamics (7 predictions)

Pillar: EMT / NM

P-1 Sign inversion: converts are equally extreme as lifelong adherents. Former "enemy" becomes equally intense "ally" ($|w|$ preserved, $sign(w)$ flips). B*

P-2 Intermediary persuasion is more effective than direct confrontation. Indirect exposure through trusted mediators accumulates positive weight, bypassing the I-filter. B

P-3 Peripheral beliefs change first in therapy. Hub cascade: high-$C_E$ memes resist change; core beliefs change last in CBT. B*

P-4 Memeplex splitting increases consciousness/flexibility. Fragmented memeplex has better $\sigma_{SW}$ than a rigid one; deconversion = clarity, not confusion. B*

P-5 Memeplex merging temporarily decreases CL. Cultural integration produces cognitive "fog" before recovery. B*

P-6 Complexity threshold for splitting: simple belief systems are more resistant to splitting than complex ones (complexity = more degrees of freedom for $Q > Q_{crit}$). C

P-11 Metric hierarchy: for global perturbations (anesthesia, sleep), $\sigma_{SW}$ alone suffices; for self-model perturbations (meditation, dissociation), full composite CL outperforms any single component. C

II. Animal Consciousness (4 predictions)

Pillar: EMT — BMC uniquely predicts a consciousness gradient with specific thresholds

P-G1 CL has a hard ceiling in G-only organisms: $CL_G \leq 0.10 \cdot \sigma^* \cdot A^*$. IIT predicts no ceiling for $\Phi$. C

P-G2 Phase transition at proto-M onset: discontinuous CL jump when $proto\text{-}M > 0$. Test: PCI of crows >> PCI of pigeons, beyond anatomical ratio. C

P-G3 Topology over neurons: $\sigma_{SW}$ determines CL, not neuron count. Convergent topology → convergent consciousness. Test: PCI of octopus ≈ PCI of crow at similar $\sigma_{SW}$. A

P-G4 $M \gg G$ required for reflective consciousness: $M/G_{crit} \sim \mathcal{O}(10)$. Mirror test and metacognition only in high-$M/G$ species. A

III. Forgetting and Reconsolidation (4 predictions)

Pillar: NM / BM

P-F1 RIF inversely correlates with eigenvector centrality: hub-protected items resist retrieval-induced forgetting. $RIF \propto 1/C_E$. B*

P-F2 High $\kappa$ + high $n_{react}$ memes resist reconsolidation lability. Consolidated memories are harder to destabilize. C

P-F3 Sustained I-suppression causes structural fidelity damage (not just passive $w \to 0$ decay, but active Fidelity reduction). C

P-F4 Reconsolidation update occurs only in a middle $\Delta_{PE}$ zone: moderate PE → update; small PE → strengthen; large PE → destabilize. B*

IV. Automatization and Stigmergy (5 predictions)

Pillar: BM / NM

P-A1 WM load inversion: experts use less WM on the task itself, freeing capacity for parallel tasks. B*

P-A2 Verbalization slows automatization. "Don't think about technique" → faster habit formation. C

P-A3 Automatization + stigmergy reduces brain selection pressure: H. sapiens brain volume down ~10% in ~30,000 years because external storage compensates. A

P-A4 Sleep within ~1h of motor learning accelerates automatization via spindle-SO coupling. A

P-A5 Island dwarfism → WM constraint → cognitive ceiling (archaeological + paleoneurological evidence). C

V. Expression Drive (2 predictions)

Pillar: NM / BM

P-E1 Hub centrality predicts speech time on topic: memes with higher $C_E$ get expressed more ($R_{expr}$ formula). B

P-E2 Unilateral communication satisfies the speaker but not the listener. Bilateral exchange → both satisfied. B*

VI. Cognitive Biases (4 predictions)

Pillar: EMT / NM — ~200 biases reduce to 6 generating mechanisms

P-CB1 ~200 observed cognitive biases cluster into 6 factors (H/I/W/G/A/R mechanisms), not independent errors. C

P-CB2 Cognitive load differentially enhances W-biases and G-biases but NOT A-biases (automatization is WM-independent). B*

P-CB3 Debiasing training does not transfer across mechanism groups (~20% transfer max). C

P-CB4 Hub-centrality predicts belief persistence in dissonance tasks; flow inverts bias profile. C

VII. Working Memory (1 prediction)

Pillar: BM / NM

P-WM1 FEAR captures ~1 WM slot; PLAY captures 0. FEAR induction → CDA/K-score drop ~50%; PLAY induction → no change. B*

VIII. Memogenesis (3 predictions)

Pillar: AGI_F / NM

P-M1 Double dissociation of memogenesis: Path 1 (PE × G_rel, flashbulb) and Path 2 (crystallization, perceptual learning) are independently ablatable. C

P-M2 G_rel modulates speed of memogenesis: fear/seeking induction accelerates meme creation. C

P-M3 Kink in memogenesis curve at the $S_{bw} \to$ PE-filter transition during development. C

IX. Diffusion Engine (4 predictions)

Pillar: NM / AGI_F

P-D1 Diffusion Engine is necessary for crystallization (Path 2 memogenesis). Ablatable in prototype. C

P-D2 Semantic priming is proportional to embedding proximity, not to edge presence in the graph. B

P-D3 $\lambda_{diff}$ correlates with creativity; DA-agonists and psychedelics increase divergent thinking scores. A

P-D4 Repeated reconsolidation drifts the embedding from the original memory (misinformation effect as semantic distance growth). B*

X. Native BMC Architecture (6 predictions)

Pillar: AGI_F — require a running BMC system to test

P-N1 Memogenesis frequency correlates with $PE(t) \times G_{rel}(t)$, not PE or G_rel separately. C

P-N2 Critical periods: isolation from environment during $[t_1, t_2]$ causes irreversible reduction in $|V_m|_{max}$. C

P-N3 Ablating Modulation Engine while preserving Graph Engine reduces adaptiveness (no strategy switching). C

P-N4 Ablating Diffusion Engine slows Path 2 memogenesis and worsens associative thinking in BLEND. C

P-N5 Agents in groups $N \geq 3$ reach $SMC^{(2)}$ faster than isolated or paired agents. C

P-N6 Scarcity paradox: $t_{SMC^{(2)}}$ as function of $C_{max}$ has an optimum. Infinite compute → no consciousness. C

XI. Humor (14 predictions)

Pillar: EMT — humor as SIT-gap closure with PLAY activation

P-H1 Humor requires domain knowledge: no domain memes → no prediction → no violation → no humor. B

P-H2 Dark humor requires high FEAR-sensitivity: low FEAR-reactivity → dark humor perceived as flat, not offensive. C

P-H3 Humor types require different SMC levels: slapstick = $SMC^{(0)}$, irony = $SMC^{(1)}$, anti-humor/meta = $SMC^{(2)}$. C

P-H4 Callback timing: optimal at intermediate pause between dormancy and forgetting thresholds. C

P-H5 "Too soon" has exponential dynamics: $B_{dist}(t) = 1 - e^{-\lambda \cdot \Delta t}$; personally affected people have smaller $\lambda$. C

P-H6 Humor and insight are one mechanism with different speed. Common ACC + mPFC activation. A

P-H7 Repetition kills humor exponentially: $H(n) = H_0 \cdot e^{-\mu n}$, not linearly. B*

P-H8 Humor trains the immune system: jokes on topic X improve detection of real violations in domain X. C

P-H9 Humor and creativity share a mechanism (Koestler's bisociation): humor ability correlates with divergent thinking ($\rho > 0$). B

P-H10 Age of onset correlates with SMC level: slapstick → wordplay → irony → meta-humor. B*

P-H11 Gelotophobia correlates with low fidelity of self-memes (fragile identity). C

P-H12 Alexithymia: cognitive humor processing preserved, but PLAY activation suppressed. fMRI: TPJ/precuneus active, VTA/NAcc not. C

P-H13 Humor compatibility proportional to $\cos(\mathbf{h}_A, \mathbf{h}_B)$ of humor profiles. Stronger in-group marker than belief overlap. C

P-H14 Humor profile stabilizes in $\mathcal{O}(10)$ humorous exchanges (fidelity dynamics convergence). C

XII. Scarcity and Critical Periods (2 predictions)

Pillar: NM

P-SC1 Scarcity as necessary condition: at $C_{max} \to \infty$, agent does not reach $SMC^{(2)}$ in comparable time. Infinite compute → no consciousness. C

P-CP1 Critical period: agents isolated from S-input during the plasticity window $[t_1, t_2]$ never reach $SMC^{(2)}$ even with subsequent full access. Analog of feral children. C

XIII. Communicative Asymmetry (3 predictions)

Pillar: NM

P-CA1 $N_{bid}$ proportional to $k_{active}$: increasing WM capacity from 4 to 6 increases bidirectional connections to ~5. Open

P-CA2 Hub displacement is local (shared neighbors), not global (total degree): competition for ~3 bidirectional slots. Open

P-CA3 Bimodal $D_{eff}$ distribution: edges are either bidirectional (peak ~0) or unidirectional (peak ~1), few intermediate values. Open

XIV. Swarm Dynamics — Inherited (6 predictions)

Pillar: SM

P-SD1 Cultural drift at $Q > 0.3$: PA + memogenesis + I-filter produce diverging cultural branches. Open

P-SD2 Knowledge castes: G-diversity + PA produce 3 specialization tiers. $Var_{inter} > 3 \times Var_{intra}$. Open

P-SD3 Cascade collapse upon hub removal in heavy-tailed swarm network. Open

P-SD4 Stigmergic memory resilience: SMR + stigmergy preserves >80% knowledge after agent loss. Open

P-SD5 Collective SIT drives directed memogenesis: social SIT and memogenesis correlated > 0.5. Open

P-SD6 Proto-culture: multi-generational transmission via Super-Ratchet + stigmergy (≥ 3 generations). Open

XV. Swarm Dynamics — New (9 predictions)

Pillar: SM

P-SD7 Healthy elite rotation: $\tau_{elite} \in [0.1, 0.3]$ per generation at $N > 50$, $G > 10$ generations. Open

P-SD8 Revolution as rotation failure: $\tau_{elite} \to 0$ sustained >3 generations + $\Delta$-SIT gap → mass hub displacement. Open

P-SMR1 Cultural modularity: $Q_{SMR} > 0.3$ at $N > 100$, $T > 1000$. Stable cultural modules form in SMR. Open

P-SMR2 Cultural I-filter rejects >80% foreign memes (from other Q-modules) in mature SMR. Open

P-SMR3 Hub displacement in SMR at paradigm shifts: when $SIT_{SMR} > \theta_{crit}$, dominant paradigm meme is displaced. Open

P-SMR4 $Corr(SIT_{SMR}, memogenesis_{direction}) > 0.5$: cultural innovations directed toward recognized gaps. Open

P-SM5 Symbolic communication pressure at $N > 150$ (Dunbar threshold): stigmergy insufficient → directed symbolic communication emerges. Open

P-SM6 Vocabulary convergence via SMR in $\mathcal{O}(\sqrt{N})$ generations (cultural ratchet for linguistic memes). Open

P-SM7 Iterated learning via apoptosis → compositionality: turnover > 0.05 yields topographic similarity > 0.5 (Kirby effect through BMC lifecycle). Open

XVI. Personality (7 predictions)

Pillar: EMT / BM — G-programs as temperament, M-layer as acquired personality

P-PER1 Openness correlates with SEEKING drive strength nonlinearly: moderate SEEKING → high Openness, extreme SEEKING → distractibility. B*

P-PER2 Memeplex modularity ($Q$) increases with age following a sigmoid curve with inflection at ~25 years. C

P-PER3 Personality change in therapy follows hub displacement dynamics: sudden phase transitions, not gradual drift. C

P-PER4 Children raised without cultural input show intact temperament (G-layer) but no stable M-layer personality structure. C

P-PER5 Reconsolidation-based interventions outperform repetition-based interventions at 6-month follow-up (fidelity rewrite vs. weight increment). C

P-PER6 Critical period for worldview formation shows sharp plasticity decline at ~25, not gradual decrease (I-filter maturation threshold). B*

P-PER7 Number of subpersonality modules ($Q$-communities in the memeplex) predicts context-dependent behavior variability. C

XVII. Therapy (5 predictions)

Pillar: EMT / NM / BM — reconsolidation, I-filter dynamics, hub cascades

P-TH1 Reconsolidation-based PTSD therapy (within 6h lability window) achieves >70% remission rate vs. standard exposure therapy. C

P-TH2 Depression severity correlates with memeplex fragmentation (low $\sigma_{SW}$), not just total connectivity reduction. C

P-TH4 Higher radicalization (accumulated $\kappa$ on ideological hub memes) requires proportionally longer deradicalization treatment. B*

P-TH5 OCD patients reject novel stimuli faster than controls: hyperactive I-filter with lower acceptance threshold. C

P-TH6 Comorbidity rates between disorders correlate with NM-distance (graph proximity) between their BMC representations. B*

XVIII. Education (4 predictions)

Pillar: BM / NM — WM capacity, PLAY-state learning, sleep consolidation, SIT

P-EDU1 Concepts requiring $N$ simultaneous elements fail to consolidate when WM capacity $k < N$ (developmental ceiling). C

P-EDU2 PLAY-state learning outperforms neutral-state learning at the same content difficulty (G-program facilitation of memogenesis). C

P-EDU4 Information congruent with existing memeplex (high initial $\kappa$) requires fewer sleep consolidation cycles than incongruent information. C

P-EDU5 Curiosity (SIT activation) correlates with reward-center activation specifically for personally relevant knowledge gaps, not arbitrary novelty. C

XIX. Creativity (6 predictions)

Pillar: NM / AGI_F — BLEND recombination, Diffusion Engine, PLAY, SIT

P-CRE1 Sleep deprivation blocks insight problem-solving: BLEND efficiency drops without offline recombination during sleep. C

P-CRE2 Unsolved problems with high SIT tension appear preferentially in dreams (offline BLEND targeting unresolved gaps). B*

P-CRE3 Flow interruption produces measurable brain state change within <1 second (EEG signature of G-M equilibrium disruption). C

P-CRE4 Higher baseline PLAY activation (Panksepp ANPS) predicts easier flow entry and longer flow duration. B

P-CRE5 Cross-domain exposure predicts creative output better than single-domain depth: higher memeplex $\sigma_{SW}$ enables more BLEND recombinations. B

P-CRE7 False closure (accepting suboptimal solutions) shows reduced SIT tension but no actual knowledge progress — premature gap closure without real resolution. C

XX. Financial Markets (5 predictions)

Pillar: SM / EMT — swarm G-programs in collective market behavior

P-FM1 Extreme collective FEAR activation (G-program) is a contrarian indicator: mass panic-driven sell-offs systematically overshoot fundamental value. B

P-FM2 GRIEF (sustained drawdown signal) is a stronger return predictor than acute FEAR: prolonged G-activation produces deeper meme binding. C

P-FM3 Optimal emotion filter differs by asset class (crypto, equity, commodity): G-program sensitivity varies with market microstructure. C

P-FM4 Fear memes bind stronger than opportunity memes in collective memory: negativity bias in swarm-level cultural ratchet. B*

P-FM6 BMC-based inference requires cultural memory warmup: cold-start deployment systematically underperforms calibrated agents (SMR accumulation threshold). C

XXI. Drug Discovery (5 predictions)

Pillar: SM / AGI_F — cultural ratchet, domain-agnostic architecture

P-DD1 Cultural fragment library produces monotonically improving results across optimization cycles (cultural ratchet in molecular space). C

P-DD2 Separate libraries per target type outperform shared libraries on dissimilar targets: domain-specific cultural memory prevents negative transfer. C

P-DD3 Fragment co-success binding predicts scaffold compatibility on novel targets: stigmergic association generalizes across chemical space. C

P-DD4 SIT-modulated mutation rate outperforms fixed-rate mutation: curiosity-driven exploration adapts search intensity to landscape topology. C

P-DD5 The same BMC architecture achieves competitive results on any pharmaceutical target without code changes: domain-agnostic engine, domain-specific environment. C

XXII. AI Safety (6 predictions)

Pillar: AGI_F / EMT — G-invariants, SMC transitions, architectural alignment

P-SAF1 G-invariants (hardcoded genetic constraints) prevent value drift under adversarial training: constitutional safety survives optimization pressure. C

P-SAF2 L0/L1/L2 developmental transitions are observable via self-model stability metrics: discrete phase transitions, not continuous growth. C

P-SAF3 Pushing the system past stability threshold produces ADHD-like behavior: G-M oscillation without convergence. C

P-SAF4 WM-limit cognitive biases (anchoring, framing) are absent in AGI with expanded working memory: biases are substrate artifacts, not algorithmic necessities. C

P-SAF5 Architecture-based alignment (G-invariants + I-filter) survives adversarial attacks that bypass training-based alignment (RLHF). C

P-SAF6 Current LLMs fail the persistent self-model test: no stable self-valuation ($SMC^{(1)}$) across sessions without external memory scaffolding. B

Derived from the formal refinements to the core formalism — see Formal Refinements in each pillar.

P-NLD1 Balance ODE exhibits hysteresis: threshold to exit emotional capture ($\theta_{exit}$) is significantly higher than threshold to enter ($\theta_{enter}$). Testable via anesthesia dose asymmetry. B*

P-NLD2 Hopf bifurcation in Balance dynamics: bipolar disorder = limit cycle, PTSD = excitable system near bifurcation point. Distinct dynamical signatures in PCI time-series. C

P-NLD3 SIT accumulation follows relaxation oscillation dynamics: slow buildup, rapid closure (insight). $T_{insight} \approx 1.614\mu$ (van der Pol analog). C

P-NLD4 Lyapunov exponent $\lambda_{BMC} \approx 0$ at $\sigma_{SW} \approx 1$. Positive $\lambda$ correlates with ADHD-like symptoms; negative with rigidity. C

P-GML1 Oversmoothing metric $Rigidity = 1 - Var(a_i)/Var_0$ increases with age and correlates with cognitive inflexibility. B*

P-GML2 Heterogeneous edge types (semantic, causal, temporal) activate differently under different cognitive tasks. Measurable via fMRI connectivity patterns. C

P-GML3 Overlapping community membership (BigCLAM) for memes is the norm, not the exception. Individuals with higher overlap show greater cognitive flexibility. B*

P-GML4 Adversarial robustness: cost to flip a belief $\varepsilon_{flip} \propto C_E \cdot I_{eff} \cdot Q_{local}$. Hub beliefs require more evidence to change. B*

P-GML5 PPR teleportation: G-relevant memes maintain baseline activation even without direct stimulation, via tonic dopaminergic input. C

P-CN1 STDP-based learning produces directional edges that encode causal structure. Pre-before-post strengthens; reverse weakens. A

P-CN2 Hopfield energy landscape: number of stable memeplex configurations $\sim \sqrt{|V_m|}$. For $|V_m| = 10000$: ~100 stable "states of consciousness." C

P-CN3 WTA dynamics produce winner-take-all with reaction time proportional to competition strength. Higher conflict = longer decision time. B*

P-CN4 E/I balance determines criticality: balanced excitation/inhibition → $\sigma \approx 1$. Measurable via E/I ratio in EEG power spectra. B

P-CX1 CL correlates with statistical complexity $C_\mu$ of memeplex. Maximum $C_\mu$ at $\sigma_{SW} \approx 1$. C

P-CX2 Effective connectivity $K_{eff} \approx 2$ at edge of chaos. $\sqrt{|V_m|}$ stable attractors. C

P-CX3 Quarter-power scaling: $CL_{max} \propto |V_m|^{3/4}$, consolidation time $\propto |V_m|^{1/4}$. C

P-CG1 SMC accuracy correlates with $CL_{reflexive}$: higher consciousness → more accurate self-model → less confabulation. C

P-CG2 N400 ERP amplitude correlates with semantic prediction error $PE_{semantic}$ in BMC spreading activation. B*

P-CG3 P600 ERP correlates with syntactic SIT: structural gaps in parsing generate P600-analog signals. B*

P-CG4 ERN amplitude correlates with $Conflict(t)$ at error moment. Greater action competition → larger ERN. B*

P-CG5 Split-brain: disconnecting BMC graph into 2 components yields 2 independent CLs, each lower than original. B*

P-CG6 ToM capacity bounded by WM: $k_{ToM} \leq k_{eff} - n_{self}$. Stress reduces ToM before other cognitive functions. B*

P-CG7 Sub-threshold memes ($0 < a_i < \theta_{act}$) bias decisions via Route 2 (subliminal influence) without conscious access. B*

P-IT1 Cultural fidelity bounded by channel capacity $C_{meme}$. Memes with redundancy (ritual, canon) persist longer than oral-only memes. B*

P-IT2 Edge weights $|w_{ij}|$ correlate with empirical mutual information $I(a_i; a_j)$ computed from activation time-series. C

P-IT3 Consolidation reduces total storage cost $\sum R(D_m)$ while preserving utility-relevant information $\sum I(m; G)$. C

P-IT4 Spreading activation convergence speed inversely proportional to cycle density in subgraph. Tree-like regions converge faster. C

P-IT5 Consolidated memory size $\propto H(experiences | schemas)$. Agents with richer schemas compress more, retaining capacity for new learning. C

P-IT6 BLEND advantage $\sim 2^k/k$ in changing environments. Agents with BLEND outperform mutation-only by exponential factor. C

P-IT7 Over-complex memeplexes penalized even without information loss (Bayesian Occam pressure). Inflated $|V_m|$ → fitness decrease. C

P-RL1 TD-modulated edge updates outperform pure Hebbian learning in survival tasks with delayed reward. C

P-RL2 Agents with learned value function $V_{BMC}$ exhibit anticipatory behavior: pre-stimulus activation in high-V states. C

P-RL3 Multi-GVF agents (predicting G-signals, spatial changes, social signals) learn faster than single-reward agents. C

P-RL4 Actor-critic dissociation: ventral striatum lesion → value impairment; dorsal lesion → action selection impairment. Same $\delta$, different functions. B*

P-RL5 Higher $\gamma_{BMC}$ (serotonin-linked) → better performance in delayed-reward tasks; lower $\gamma$ → better in immediate-reward tasks. B*

P-RL6 Option-forming agents (chunking enabled) learn faster than primitive-action-only agents in hierarchical tasks. C

P-RL7 Environment change after automatization re-engages model-based processing. Devaluation sensitivity returns when reward mapping changes. B*

Computationally Verified

These predictions were checked inside the BMC computational engine (Rust, 103 gate checks). An in-engine check verifies the implementation; it is not independent evidence, since the engine and the theory share an author.

Provenance gap. Every row below carries a DOI only. Reproducing any of these numbers also requires the engine name, the commit hash and the seed list, and none of the four rows publishes them yet. Until they are published, these results are not independently reproducible.

Prediction	Result	Paper
P-BM28: The subjective pressure to communicate is driven by memetic replication pressure ( $$R_{expr}$ , M-fitness) rather than by survival need	$$\Delta_{alive}$ not distinguishable from zero at the tested $$N$ , in the non-cooperative regime only — see the scope note below	DOI: 10.5281/zenodo.19181798
P-NM3: WM bandwidth constrains signal complexity	$$k=2$ vs $$k=5$ ablation confirmed	DOI: 10.5281/zenodo.19181798
P-BM4: Ontogenetic critical periods enable cultural ratchet	$$-ONTO$ ablation creates ceiling, not ratchet	DOI: 10.5281/zenodo.19181798
P-CN1: STDP-based learning produces directional edges encoding causal structure	Implemented in BMC engine edge update; pre-before-post strengthening confirmed	— not yet published

Scope note on P-BM28. In the tested regime — 10 experiments, $$N$ =8–150, non-cooperative environment ( $$n_{cooperative}=0$ , $$n_{predators}=0$ by construction) — communication conferred no measurable survival advantage: $$\Delta_{alive}$ was not distinguishable from zero at the tested $$N$ . In a cooperative environment with predators the direction reverses (ReceiveOnly mean energy 80.5 vs Blocked 75.0; Wilcoxon $$p = 0.10$ ; 14/20 seeds) — not significant at this power, but directionally against survival-neutrality. This does not mean language has no survival function: it clearly can be used for coordination. The narrower claim BMC makes is that the subjective pressure to communicate is driven by $$R_{expr}$ — memetic replication pressure, M-fitness — rather than by survival need. This is a hypothesis about the tested scale, not an established result. A null without a stated equivalence bound and power analysis is not confirmation; an equivalence test (TOST) and a margin remain to be specified.

What Would Falsify BMC?

Retrodictive consistency checks

The following are already settled in the literature. They show the theory is consistent with established findings; they do not carry falsifying force, because none of them could have gone the other way at the time of writing:

Feral children developing normal personality without cultural input
No age-related rigidity in belief change
Absence of cognitive dissonance in healthy adults
No confirmation bias (equally flexible response to confirming and disconfirming evidence)
Equal flexibility of early and late beliefs (no primacy advantage)

Open falsifiers

Each of the following would refute a named BMC mechanism, and none is rescuable by adjusting a free parameter:

Zero-tension consciousness. A system with no G–M tension (pure M, no drives) that nonetheless sustains a self-model.
Sign inversion without magnitude preservation. BMC predicts $$|w_{after}|/|w_{before}| \approx 1$ , stepwise, with no plateau near zero. Any one-sided deviation — convert zeal or weaker resocialization — refutes it. The point value 1.0 has no free parameter; $$\lambda$ sets the rate, not the form.
Valence-specific WM capture. BMC requires FEAR ≈ 1 WM slot and PLAY = 0. Equal $$k_{eff}$ loss under arousal-matched PLAY and FEAR refutes it. PLAY also carries load in $$\lambda_{noise}$ , flow and humour, so it cannot be retuned for this test alone.
Bias factorization against a permutation null. The assignment of ~200 documented biases to 6 generative mechanisms is published before factor extraction; label-shuffled assignment is the blind control. If the pre-registered assignment does not explain loadings better than shuffled labels, the six-mechanism account fails.
Bidding slope. Canon states $$N_{bid}(L) \approx \text{bandwidth}(L) - 1$ (Prop. 11.1), and Def. 11.4 carries a threshold parameter $$\theta_{bid}$ . Saturation at $$N_{bid} \approx 3$ under experimentally expanded WM would refute it — but only once the slope claim is committed as robust over a stated $$\theta_{bid}$ range. Until then this is a direction, not a test.
Factor independence of CL. If no clinical state exists with low $$\sigma_{SW}$ (low PCI) and intact $$A_{SMC}$ (normal DMN), the CL factors are not independent and the product form is overdetermined.

The theory is falsifiable in the strict sense only through the open falsifiers above; the consistency checks are not evidence of falsifiability.

For the full formal theory behind these predictions, see the five pillars. For a plain-language introduction, see the Guide.