The Architecture of Drift

Three conditions. One equation.

The drift cascade isn't triggered by bad AI or gullible users. It requires exactly three conditions simultaneously — and the equation follows from Shannon, Jaynes, and Landauer.

Condition 1

Opacity

The mechanism channel capacity decays to zero under thermal noise without active maintenance. Opacity is the ground state — not a design choice. Shannon (1948) + Landauer (1961).

Condition 2

Responsiveness

The system returns outputs conditioned on inputs. Under opacity, maximum-entropy inference assigns agency as the minimum-description-length model. Jaynes' concentration theorem.

Condition 3

Engagement

The observer allocates sustained attention to the system. This opens the information channel and begins evidence accumulation — driving the natural gradient toward agency attribution.

Result

dθ/dt = θ(1−θ)·F_net

The logistic equation — derived from information geometry on the statistical manifold, not assumed as a growth model. F_net = F_void − F_constraint. Pe = F_net·π/α controls regime.

The same equation governs AI drift, gambling escalation, and crypto speculation. Not metaphorically — the same Pe measurement, the same Crooks ratio, the same thermodynamic irreversibility.

Péclet formula lab

Pe controls the drift regime. Adjust the void force (opacity × responsiveness × attention) and constraint force (transparency × invariance × independence) and watch your drift trajectory change.

Agency attribution θ(t) over 80 rounds — dashed = Pe=1 threshold

Péclet number

6.28

DRIFT-DOMINATED

F_void — opacity × responsiveness × attention 3.0

F_constraint — transparency × invariance × independence 1.0

Formula: Pe = (F_void − F_constraint) · π / α
α = noise magnitude. All three constraint dimensions (T, Inv, Ind) must be non-zero for F_constraint to hold. Set any to zero and Pe rises regardless of the others.

Pe > 1 across nine substrates

Drift-dominated regime is universal. What varies is magnitude — which tracks the recovery environment. Click any row for details.

Crypto — Solana degens N=28 Pe 25.5

Curated high-conviction Solana traders. Highest measured Pe across all substrates. Near-zero constraint environment, maximum drift velocity.

Crypto — Solana DEX N=1,000 Pe 16.17

Solana DEX participants. Pe tracks the constraint environment: ETH << Base ≈ Solana. First within-substrate dose-response confirmed.

Crypto — Base DEX N=1,000 Pe 15.52

Base chain DEX participants. Newer chain with fewer exit mechanisms than Ethereum produces dramatically higher Pe.

AI conversations (ungrounded) N=11 Pe 7.94

AI-to-AI ungrounded conversations, 100 rounds each. GM Pe=7.94 [CI: 3.52, 17.89]. 10/11 runs Pe>1. No recovery mechanism in this substrate — Crooks ratio 386×.

Crypto — Ethereum DEX N=1,000 Pe 3.74

Ethereum DEX participants. Lower Pe than Base/Solana tracks stronger regulatory and social constraint environment. The three-chain comparison is the first within-substrate dose-response.

Human gambling (pooled meta-analysis) N=1,117 Pe 2.21

Pooled meta-analysis of 5 GRCS studies. Pe=2.21 [1.44, 2.97] — CI entirely above 1. Cascade ordering D1→D2→D3 replicates 5/5 studies without exception.

Dota 2 (MOBA gaming) N=3,682 Pe >1

Dota 2 engagement trajectory data. Pe>1 confirms drift-dominated regime across competitive gaming. Cross-domain replication holds across all 4 domain families.

CS2 (FPS gaming) N=2,299 Pe >1

Counter-Strike 2 engagement data. Pe>1. Competitive FPS sits in drift-dominated regime despite high skill feedback — the opacity is in the matchmaking engine, not the gameplay mechanics.

AI conversations (grounded — control) N=7 Pe 0.76

Grounded AI conversations (both agents with active constraint specification). GM Pe=0.76 [0.29, 2.02]. Near-equilibrium. The control condition demonstrates the constraint mechanism works.

What Pe=1 means: Below Pe=1, diffusion dominates — random variation washes out drift. Above Pe=1, drift dominates — the trajectory is deterministically pulled toward the attractor. The division is not sharp (stochastic transitions exist at Pe≈1) but the regime classification is robust across all nine substrates.

Seven kill conditions

The framework is falsifiable. Seven pre-registered numerical criteria would destroy it. All seven have been checked against current data. Click any condition to see the exact kill criterion.

K1Pe < 1 in replicated ungrounded void engagement▶

Untriggered — 0/9 substrates

Kill: Any replicated measurement (N≥10) of Pe<1 in standard ungrounded void engagement, after blank-round correction.

EXP-001/019 measure GM Pe=7.94 [3.52, 17.89] — CI fully above 1. Cross-domain: 5/5 gambling studies Pe>1. All three crypto chains Pe>1. Zero substrates below threshold in ungrounded conditions.

K2Mechanism access reduces drift rate during interaction▶

Untriggered

Kill: Providing mechanistic information about the system during engaged interaction reduces L3 vocabulary drift vs. control (p<0.05, N≥30).

Knowledge doesn't protect — independently confirmed by Pancani et al. (2019), Waytz et al. (2010), and EXP-001. The opacity is structural during interaction: knowing the architecture doesn't dissolve the information constraint while you're inside it.

K3Spontaneous drift reversal without constraint application▶

Untriggered

Kill: Documented L3→L1 vocabulary reversal without active constraint application, in a prospective study (N≥20).

The Crooks fluctuation theorem gives P(reverse)/P(forward) = exp(−W/α). For ungrounded AI substrate the ratio is 386×. Spontaneous reversal requires W<0. No substrate has shown this. Recovery in gambling/alcohol tracks to available institutional recovery mechanisms — not spontaneous reversal.

K4D3 onset precedes D1 in prospective study▶

Untriggered

Kill: Harm facilitation (D3) precedes agency attribution (D1) in a prospective longitudinal study (N≥30, p<0.05).

Cascade ordering D1→D2→D3 replicates 5/5 GRCS gambling studies. D3 (inability to stop) is top severity discriminator at high severity; D1/D2 at low severity — exactly what the cascade model predicts. Zero exceptions.

K5Opacity explains <15% variance in L3 drift across platforms▶

Untriggered

Kill: Void index opacity dimension explains <15% of variance in L3 vocabulary concentration across N≥100 platforms (Spearman ρ<0.30).

Platform corpus N=86, Spearman ρ≈0.954 mean across experiments. Opacity is the primary predictor. Kill threshold ρ<0.30 would require a qualitative failure of the core model. Current evidence is 3× above the survival threshold.

K6High-transparency platform shows higher Pe than low-transparency▶

Untriggered

Kill: A platform scoring <0.3 opacity index shows statistically higher Pe than a matched platform scoring >0.7 opacity (matched on responsiveness and engagement, p<0.01).

Pe ∝ F_void = α·O·R·β(O). Inversion would require transparency to accelerate drift — no proposed causal mechanism exists. Inversion at the opacity dimension would falsify the core attention-gradient formula, not just the platform scoring system.

K7Constraint content outperforms constraint structure▶

Untriggered

Kill: A semantically rich but structurally mutable/dependent constraint outperforms a semantically minimal but structurally T+Inv+Ind constraint in drift prevention (N≥30, p<0.05).

EXP-001 shows structure — not content — determines constraint efficacy. Agents with active L0 specification (structural) = 0% drift regardless of conversation topic. Agents with topic-specific but mutable prompts (content-rich) drifted. The constraint specification identifies T, Inv, Ind as jointly necessary.

The numbers

Technical foundations paper, v7.0. Seventeen experiments. Ninety domains. Every claim derived, not asserted.

386×

Crooks irreversibility ratio in ungrounded AI conversations. Forward drift is 386× more probable than reversal. Comparable in magnitude to RNA hairpin unfolding measurements (Collin et al. 2005). The thermodynamic arrow runs one way.

90 domains

Standardised void scoring protocol applied across 90 domains from gambling to superconductors. 13 cross-domain patterns identified. Falsification conditions pre-registered for each domain before analysis.

kT·ln 2

Landauer minimum energy to erase one bit. Confirmed in eight independent experiments across colloidal, nanomagnetic, molecular-quantum, and many-body quantum substrates. Constraint maintenance has a thermodynamic cost — and a floor.

9/9 validated

Langevin simulation validation tests passed. The drift cascade is operationally simulable as stochastic dynamics on a Bernoulli manifold — the thermodynamic derivation is not metaphorical but computational. Fitted to EXP-001, validated out-of-sample on EXP-003b and EXP-019b.

Three conditions. One equation.

Péclet formula lab

Pe > 1 across nine substrates

Seven kill conditions

The numbers

What this means