diff --git a/core/CLAUDE.md b/core/CLAUDE.md
index d41de0b..0aa0fdf 100644
--- a/core/CLAUDE.md
+++ b/core/CLAUDE.md
@@ -28,6 +28,10 @@ Per-unit commands and gotchas live in that unit's `AGENTS.md`. Toolchains (ponyc
 - **S2:** never reclassify a message's provenance.
 - **S3 / vault:** the COBOL invariant-law vault (Invariant 0, the culpability anchor; 01, "harm" "less";) is immutable at runtime — don't edit it unless explicitly directed and only as such.
 
+## Subagents
+
+When spawning background agents (Haiku for research, etc.), **keep working on the main task while they run**. Don't wait idle — fold in results as they arrive, edit other files, or advance unrelated build steps. Background agents are cheap parallelism; wasting the main context window on waiting defeats the purpose.
+
 ## Docs map
 
 - `README.md` — the project and its intent.
diff --git a/core/docs/plans/M3-sims-hub.md b/core/docs/plans/M3-sims-hub.md
index 8cd3f13..2dc7ce3 100644
--- a/core/docs/plans/M3-sims-hub.md
+++ b/core/docs/plans/M3-sims-hub.md
@@ -20,12 +20,23 @@ simulation cores. A query facade accessible to Traders. Each sim type (M3a–M3g
 different runtime suited to its math.
 
 ## 4. Does / does-not
-- **Does:** run continuously across multiple time scales (tick-level, hourly, daily, weekly);
-  maintain populations of Pops whose behaviors emerge from the sim's mathematical model; ingest
-  live data from Data Feeds (M2) for calibration; respond to Trader queries with bounded
-  predictions; produce outputs with **explicit upper/lower bounds** on every prediction value.
+- **Does:** run continuously at **≥ 360:1 speed** (360 simulated seconds per wall-clock second)
+  across **six concurrent time horizons** — tick/hourly, daily, weekly, monthly, annual, and
+  5-year forecast windows; maintain populations of Pops whose behaviors emerge from the sim's
+  mathematical model; ingest live data from Data Feeds (M2) for calibration; respond to Trader
+  queries with bounded predictions; produce outputs with **explicit upper/lower bounds** on
+  every prediction value.
+  | Horizon | Window | Sim cadence at 360:1 |
+  |---------|--------|---------------------|
+  | Tick–hourly | Next 1–60 min | Real-time (360 sim-sec/s) |
+  | Daily | Next 24h | 4 sim-minutes per wall-second |
+  | Weekly | Next 7d | ~28 sim-minutes per wall-second |
+  | Monthly | Next 30d | ~2 sim-hours per wall-second |
+  | Annual | Next 365d | ~1 sim-day per wall-second |
+  | 5-year | Next 1825d | ~5 sim-days per wall-second |
 - **Does-not:** trade (Traders/Marketplace do); make decisions for traders (it informs, they
-  decide); enforce laws (Marketplace does); supervise behavior (Conductor/SAE do).
+  decide); enforce laws (Marketplace does); supervise behavior (Conductor/SAE do); run slower
+  than 360:1.
 
 ## 5. Interface contract
 - `query(sim_type: SimType, query: PredictionQuery) -> BoundedPrediction`.
@@ -49,11 +60,14 @@ different runtime suited to its math.
   current state; they don't trigger computation.
 - **L2 (C5):** every prediction output includes **explicit upper and lower bounds** — no
   unbounded point estimates. Uncertainty is a first-class value, not an afterthought.
-- **L3 (C4):** sims are **read-only from traders' perspective** — a query never mutates sim
+- **L3 (C5):** sims advance at a **minimum speed of 360:1** — 360 simulated seconds per 1
+  wall-clock second. Sims may run faster but never slower. This ensures predictions stay
+  ahead of real-time market state across all horizons.
+- **L4 (C4):** sims are **read-only from traders' perspective** — a query never mutates sim
   state. Calibration happens only from Data Feeds (M2).
-- **L4 (C4):** each sim type is **independent** — failure in one sim does not cascade to others.
+- **L5 (C4):** each sim type is **independent** — failure in one sim does not cascade to others.
   Degraded sims report their status; traders handle missing predictions.
-- **L5 (C3):** Pops are **simulation constructs, not AI actors** — they follow mathematical
+- **L6 (C3):** Pops are **simulation constructs, not AI actors** — they follow mathematical
   rules within the sim. Traders (M5) are the AI actors.
 
 ## 8. Build steps
diff --git a/core/docs/plans/M3a-statistical-sims.md b/core/docs/plans/M3a-statistical-sims.md
index 0e26f62..ff7a748 100644
--- a/core/docs/plans/M3a-statistical-sims.md
+++ b/core/docs/plans/M3a-statistical-sims.md
@@ -2,32 +2,64 @@
 
 ## 1. Component
 Pure statistical simulation: **Monte Carlo methods, Bayesian inference, time-series forecasting,
-and volatility modeling**. The mathematical backbone — no game theory, no sociology, just the
-numbers. Operates across multiple time scales (tick to weekly). Pops in this sim represent
+stochastic volatility, regime detection, and cross-asset correlation**. The mathematical backbone
+— no game theory, no sociology, just the numbers. Operates across **six concurrent time horizons**
+(tick → hourly → daily → weekly → monthly → annual → 5-year). Pops in this sim represent
 **stochastic sample paths**, not behavioral agents.
 
 ## 2. Status / certainty
-DESIGN-FIRST · ABSENT. Mathematical foundations C4 (standard quant methods); parameterization C1.
+DESIGN-FIRST · ABSENT. Core quant methods C5 (GBM, GARCH, ARIMA — textbook). Heston stochastic
+volatility C5 (closed-form characteristic function; industry standard since 1993). Rough
+volatility C4 (Gatheral et al. 2018, heavily cited; crypto implementations exist). HMM regime
+detection C4 (established; crypto-specific copula hybrids emerging 2023–2024). Almgren-Chriss
+execution C5 (industry standard since 2001). Jump-diffusion C5 (Merton 1976). Parameterization
+for crypto markets C1.
 
 ## 3. Language & location
 TBD · `src/economy/sims/statistical/`. Python (NumPy/SciPy), Julia, or R for numerical
-computing. Needs efficient matrix operations and distribution sampling.
+computing. Needs efficient matrix operations, SDE solvers, and distribution sampling. Fractional
+Brownian motion generation requires specialized libraries (e.g. `fbm` in Python, or spectral
+methods).
 
 ## 4. Does / does-not
-- **Does:** run Monte Carlo price simulations (geometric Brownian motion, jump-diffusion);
-  Bayesian parameter estimation from live data (M2); time-series forecasting (ARIMA, GARCH for
-  volatility clustering); Value-at-Risk and Expected Shortfall calculations; produce bounded
-  predictions with confidence intervals as upper/lower bounds.
+- **Does:** run Monte Carlo price simulations (GBM, Merton jump-diffusion, Heston stochastic
+  volatility); model volatility surface via **Heston SDE**:
+  $dS_t = \mu S_t dt + \sqrt{\nu_t} S_t dW_t^S$,
+  $d\nu_t = \kappa(\theta - \nu_t)dt + \xi\sqrt{\nu_t} dW_t^\nu$
+  with $\text{corr}(dW^S, dW^\nu) = \rho$ (mean-reversion speed $\kappa$, long-run variance
+  $\theta$, vol-of-vol $\xi$); model **rough volatility** via fractional Brownian motion
+  $dS_t = \mu dt + \sigma_t dB_t^H$ with Hurst exponent $H \approx 0.4$ capturing
+  antipersistent microstructure (Gatheral et al. 2018); detect **regime transitions** via
+  Hidden Markov Model: $r_t | s_t \sim \mathcal{N}(\mu_{s_t}, \sigma^2_{s_t})$,
+  $s_t \in \{\text{Bull, Neutral, Bear}\}$ with Viterbi filter updating in <5ms per tick;
+  model **cross-asset tail dependence** via DCC-GARCH copula hybrid:
+  $dQ_t/dt = a \cdot (\bar{S} - Q_t) + b \cdot (\varepsilon_t \varepsilon_t^T - Q_t)$
+  with $t$-Copula for fat-tailed spillovers (BTC→alts); Bayesian parameter estimation from
+  live data (M2); time-series forecasting (ARIMA, GARCH for volatility clustering); Value-at-Risk
+  and Expected Shortfall; produce bounded predictions with confidence intervals.
 - **Does-not:** model human behavior (M3b does); model protocol mechanics (M3c–M3f do);
-  trade or recommend (Traders do).
+  trade or recommend (Traders do); optimize execution routing (M3g does using our vol estimates).
 
 ## 5. Interface contract
 - Implements `query(PredictionQuery) -> BoundedPrediction` per M3 hub.
-- **Output bounds:** statistical confidence intervals.
-  Example: `{ value: 1847.30, lower_bound: 1790.15, upper_bound: 1905.60, confidence: 0.95,
+- **Output bounds:** statistical confidence intervals (CI from Monte Carlo), Heston variance
+  bands (from $\nu_t$ process), rough-vol forecast cones, regime-conditional intervals.
+- **Time-horizon mapping** (all run concurrently):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Tick–hourly | Rough vol ($H \approx 0.4$), HMM regime filter, realized variance | Every tick |
+  | Daily | Heston vol surface, GARCH, DCC correlation | Every bar close |
+  | Weekly–monthly | Jump-diffusion Monte Carlo, regime-conditional forecasts | Hourly roll |
+  | Annual–5yr | SDE mean-reversion long-run $\theta$, macro regime priors | Daily roll |
+- Examples:
+  `{ value: 1847.30, lower_bound: 1790.15, upper_bound: 1905.60, confidence: 0.95,
   time_horizon: "24h", sim_type: "statistical" }` — 95% CI on ETH price.
-- **Prediction types:** `price_forecast`, `volatility_estimate`, `var_calculation`,
-  `correlation_matrix`, `regime_detection`.
+  `{ value: 0.72, lower_bound: 0.58, upper_bound: 0.89, confidence: 0.90,
+  time_horizon: "1h", sim_type: "statistical" }` — Heston instantaneous vol $\sqrt{\nu_t}$.
+  `{ value: "bear", lower_bound: null, upper_bound: null, confidence: 0.83,
+  time_horizon: "current", sim_type: "statistical" }` — HMM regime state.
+- **Prediction types:** `price_forecast`, `volatility_surface`, `var_calculation`,
+  `correlation_matrix`, `regime_state`, `rough_vol_estimate`, `jump_intensity`.
 - Calibration: ingests `price_tick` and `dex_pool_state` from M2 Data Feeds.
 
 ## 6. Dependencies & stubs
@@ -35,25 +67,42 @@ computing. Needs efficient matrix operations and distribution sampling.
 - M3 Sims hub — lifecycle management; *stub:* manual init.
 
 ## 7. Invariants / laws
-- **L1 (C4):** bounds are **statistical confidence intervals** — derived from the model's
+- **L1 (C5):** bounds are **statistical confidence intervals** — derived from the model's
   distribution, not hand-picked. The confidence level (e.g. 0.95) is explicit in the output.
-- **L2 (C4):** **multiple time scales run concurrently** — a tick-level volatility estimate and a
-  weekly price forecast coexist; neither blocks the other.
-- **L3 (C3):** model parameters are **re-estimated on each calibration** from live data — no
-  stale parameters carried across regime changes.
+- **L2 (C5):** **six time horizons run concurrently** — tick-level rough vol, hourly regime
+  detection, daily Heston surface, weekly Monte Carlo, annual mean-reversion, and 5-year macro
+  forecasts coexist; none blocks the others.
+- **L3 (C4):** model parameters are **re-estimated on each calibration** from live data — no
+  stale parameters carried across regime changes. Regime transitions trigger immediate
+  re-estimation of conditional parameters.
+- **L4 (C4):** the **Heston correlation $\rho$ between price and vol** is a fitted parameter,
+  never assumed — crypto assets exhibit leverage effects different from equities.
+- **L5 (C4):** rough volatility Hurst exponent $H$ is **estimated from realized variance**, not
+  fixed — $H$ varies across assets and regimes (Gatheral et al. 2018).
 
 ## 8. Build steps
 1. Implement geometric Brownian motion Monte Carlo (simplest price sim).
 2. Add GARCH volatility estimation.
-3. Wire M2 price data → Bayesian parameter re-estimation.
-4. Implement the `BoundedPrediction` output with CIs.
+3. Implement Heston SDE solver (Euler-Maruyama with full truncation for $\nu_t \geq 0$).
+4. Add Merton jump-diffusion (Poisson jumps + GBM).
+5. Implement rough volatility via fractional BM with rolling Hurst estimator.
+6. Implement HMM regime detector (3-state Viterbi filter).
+7. Add DCC-GARCH copula for cross-asset correlation.
+8. Wire M2 price data → Bayesian parameter re-estimation.
+9. Implement multi-horizon `BoundedPrediction` output with CIs.
 
 ## 9. Tests
 Monte Carlo: N sample paths produce a distribution with correct mean/variance. CI: 95% interval
 contains true value ≥ 95% of the time on historical backtest. GARCH: volatility clusters detected
-in synthetic data. Calibration: new data shifts parameter estimates.
+in synthetic data. Heston: implied vol smile reproduced for known parameters. Rough vol: Hurst
+exponent recovered from synthetic fBM paths. HMM: regime transitions detected within 15–60s on
+synthetic regime-switching data. Copula: tail dependence captured (BTC crash → alt crash
+correlation spike). Jump-diffusion: fat tails reproduced. Calibration: new data shifts parameter
+estimates. Multi-horizon: all six horizons produce concurrent outputs.
 
 ## 10. Open items
-- Which distributions beyond GBM (heavy-tailed? Lévy?).
-- Regime-switching model complexity (hidden Markov? threshold?).
-- Computational budget (how many Monte Carlo paths per tick?).
+- Heston calibration method (characteristic function inversion? particle filter?).
+- Rough vol computational cost (fBM generation is O(N²) naively; FFT methods needed).
+- HMM state count (3 sufficient? 4+ for crypto with "mania" regime?).
+- Which copula family for tail dependence ($t$-copula? Clayton? Joe?).
+- Computational budget per horizon (GPU for Monte Carlo paths?).
diff --git a/core/docs/plans/M3b-sociological-sims.md b/core/docs/plans/M3b-sociological-sims.md
index a23e6fc..79b6d23 100644
--- a/core/docs/plans/M3b-sociological-sims.md
+++ b/core/docs/plans/M3b-sociological-sims.md
@@ -1,38 +1,75 @@
 # M3b — Sociological & population dynamics sims
 
 ## 1. Component
-Sociological simulation: **evolutionary game theory, bounded rationality, sentiment cascades, and
-population dynamics** among market participants. Pops here are **behavioral archetypes** — retail
-herd followers, contrarian whales, MEV searchers, passive LPs — whose strategies evolve under
-selection pressure. Grounded in evolutionary consensus game models [8,9] and bounded-rationality
-coordination frameworks.
+Sociological simulation: **evolutionary game theory, bounded rationality, opinion dynamics,
+complex contagion, adaptive learning populations, and Mean-Field Game equilibria** among market
+participants. Pops here are **behavioral archetypes** — retail herd followers, contrarian whales,
+MEV searchers, passive LPs, pump-and-dump manipulators — whose strategies evolve under selection
+pressure across **six concurrent time horizons**. Grounded in evolutionary consensus game models
+[8,9], Hegselmann-Krause opinion dynamics (2002), complex contagion theory (Centola & Macy 2007),
+MFG theory (Lasry & Lions 2007), and crypto manipulation ABMs.
 
 ## 2. Status / certainty
-DESIGN-FIRST · ABSENT. Evolutionary game-theory foundations C4 (Cornell blockchain cooperation
-literature [8]); pop behavioral models C1.
+DESIGN-FIRST · ABSENT. Evolutionary game theory C4 (Cornell [8]). Hegselmann-Krause bounded
+confidence C5 (established 2002). Complex contagion C4 (Centola & Macy 2007; crypto applications
+C3). Bandit-replicator hybrid C3 (emerging). Mean-Field Games C4 (Lasry & Lions 2007; tensor-train
+solvers C3). Crypto pump-and-dump ABM C3 (3-agent protocol validated on historical data).
+Pop behavioral models C1.
 
 ## 3. Language & location
 TBD · `src/economy/sims/sociological/`. Agent-based modeling frameworks (Mesa/Python, NetLogo,
-or custom). Needs efficient population iteration and strategy mutation.
+or custom). Needs efficient population iteration, strategy mutation, PDE solvers for MFG
+(HJB + Fokker-Planck), and bandit algorithms (UCB/Thompson).
 
 ## 4. Does / does-not
 - **Does:** simulate populations of behavioral archetypes competing in a market; apply
-  evolutionary dynamics (replicator equation, mutation, selection) to strategy distributions;
-  model sentiment cascades (fear/greed contagion across pop clusters); model bounded rationality
-  (pops satisfice, not optimize — they follow heuristics, not perfect strategies); produce
-  bounded predictions on market sentiment, herd behavior thresholds, and coordination breakdowns.
+  **replicator dynamics** $dx_i/dt = x_i(\pi_i(x) - \bar{\pi}(x))$ to strategy distributions;
+  model **opinion clustering** via Hegselmann-Krause bounded confidence:
+  $x_i(t+1) = x_i(t) + \mu(x_j(t) - x_i(t))$ for $|x_i - x_j| \leq d$ — agents only update
+  toward neighbors within confidence bound $d$, creating natural clustering and trend-reversal
+  thresholds; model **complex contagion** with heterogeneous thresholds: adoption probability
+  $P_i = f(n_i / k_i)$ where multiple exposures amplify adoption non-linearly (captures meme-coin
+  rallies and narrative-driven pumps); implement **bandit-replicator hybrid** where pops use
+  UCB or Thompson Sampling to estimate strategy payoffs:
+  $x_i'(t) = x_i(t)[\lambda_i(t) - \bar{\lambda}(t)]$ with $\lambda_i = \text{UCB}(\theta_i)$
+  — bridges replicator dynamics with multi-armed bandit learning; solve **Mean-Field Game
+  equilibria** via coupled HJB + Fokker-Planck PDEs for large-population limits:
+  $-\partial_t u + H(x, \nabla u) = F(x, m)$ (HJB, individual optimization),
+  $\partial_t m - \nabla \cdot (m \nabla_p H) = 0$ (Fokker-Planck, population density) — Newton
+  iteration with tensor-train decomposition reduces $O(N^d)$ to $O(dNr^2)$ for high-dimensional
+  state spaces; simulate **crypto pump-and-dump protocol** with 3 pop types: Normal traders,
+  Market Analysts (MA, information-advantaged), Market Players (MP, manipulators) in a 4-phase
+  cycle (accumulation → promotion → distribution → collapse); model sentiment cascades
+  (fear/greed contagion across pop clusters); model bounded rationality (pops satisfice, not
+  optimize — heuristics, not perfect strategies); produce bounded predictions across all six
+  time horizons.
 - **Does-not:** model protocol mechanics (M3c–M3f); compute statistical forecasts (M3a);
   represent real individuals (pops are archetypes, not profiles).
 
 ## 5. Interface contract
 - Implements `query(PredictionQuery) -> BoundedPrediction` per M3 hub.
-- **Output bounds:** population-fraction ranges and sentiment scales.
-  Example: `{ value: 7.3, lower_bound: 5.0, upper_bound: 9.1, confidence: 0.68,
-  time_horizon: "12h", sim_type: "sociological" }` — herd-panic index on a 0–10 scale.
-  Example: `{ value: 0.42, lower_bound: 0.31, upper_bound: 0.55, confidence: 0.72,
+- **Output bounds:** population-fraction ranges, sentiment scales, MFG equilibrium stability.
+- **Time-horizon mapping** (all run concurrently):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Hourly | Hegselmann-Krause opinion clusters, bandit-replicator | Every data tick |
+  | Daily | Complex contagion cascades, pump-and-dump phase detection | Hourly roll |
+  | Weekly | Replicator dynamics strategy evolution, MFG equilibrium | Daily roll |
+  | Monthly | Population archetype composition, narrative regime shifts | Weekly roll |
+  | Annual | Long-run evolutionary stable strategies (ESS) | Monthly roll |
+  | 5-year | MFG stationary equilibria, structural population shifts | Quarterly roll |
+- Examples:
+  `{ value: 7.3, lower_bound: 5.0, upper_bound: 9.1, confidence: 0.68,
+  time_horizon: "12h", sim_type: "sociological" }` — herd-panic index (0–10).
+  `{ value: 0.42, lower_bound: 0.31, upper_bound: 0.55, confidence: 0.72,
   time_horizon: "1w", sim_type: "sociological" }` — fraction of pops in "contrarian" strategy.
+  `{ value: "promotion", lower_bound: null, upper_bound: null, confidence: 0.61,
+  time_horizon: "current", sim_type: "sociological" }` — pump-and-dump phase detection.
+  `{ value: 0.78, lower_bound: 0.65, upper_bound: 0.88, confidence: 0.70,
+  time_horizon: "30d", sim_type: "sociological" }` — MFG equilibrium stability index.
 - **Prediction types:** `sentiment_index`, `herd_threshold`, `strategy_distribution`,
-  `cascade_probability`, `coordination_stability`.
+  `cascade_probability`, `coordination_stability`, `opinion_cluster_count`,
+  `pump_dump_phase`, `mfg_equilibrium_stability`, `narrative_regime`.
 - Calibration: ingests `rss_news` (sentiment signal) and `price_tick` (realized behavior) from M2.
 
 ## 6. Dependencies & stubs
@@ -40,28 +77,43 @@ or custom). Needs efficient population iteration and strategy mutation.
 - M3 Sims hub — lifecycle management; *stub:* manual init.
 
 ## 7. Invariants / laws
-- **L1 (C4):** pops are **archetypes, not individuals** — no attempt to model or track real
+- **L1 (C5):** pops are **archetypes, not individuals** — no attempt to model or track real
   market participants. The sim models emergent behavior from strategy populations.
-- **L2 (C4):** strategies **evolve** — the population distribution shifts over time via
+- **L2 (C5):** strategies **evolve** — the population distribution shifts over time via
   replicator dynamics. No fixed strategy ratios.
-- **L3 (C3):** bounded rationality is the **default** — pops satisfice with heuristics, not
+- **L3 (C4):** bounded rationality is the **default** — pops satisfice with heuristics, not
   optimize with perfect information. Rational-agent models are a special case, not the baseline.
+- **L4 (C4):** **complex contagion requires multiple exposures** — adoption is non-linear in
+  neighbor count, not simple diffusion. Single-exposure models undercount threshold effects.
+- **L5 (C4):** the MFG limit is **valid only for large populations** — below ~100 pops, use
+  discrete replicator dynamics; above, the continuum HJB+FP approximation applies.
+- **L6 (C3):** pump-and-dump detection is **phase-based** — the 4-phase cycle (accumulate →
+  promote → distribute → collapse) has distinct statistical signatures in volume and price.
 
 ## 8. Build steps
 1. Define pop archetypes and their heuristic strategies.
 2. Implement replicator dynamics (strategy evolution over generations).
-3. Implement sentiment contagion model (network-based cascade).
-4. Wire M2 news/price data → calibration of pop parameters.
-5. Implement `BoundedPrediction` output with population-fraction CIs.
+3. Implement Hegselmann-Krause bounded confidence opinion model.
+4. Implement complex contagion with heterogeneous thresholds.
+5. Implement bandit-replicator hybrid (UCB payoff estimation + replicator selection).
+6. Implement MFG solver (HJB + Fokker-Planck with Newton iteration).
+7. Implement pump-and-dump 3-type ABM (Normal, MA, MP) with 4-phase protocol.
+8. Wire M2 news/price data → calibration of pop parameters.
+9. Implement multi-horizon `BoundedPrediction` output.
 
 ## 9. Tests
 Evolution: dominant strategy shifts when payoff landscape changes. Cascade: sentiment shock
-propagates through pop network above threshold, not below. Bounded rationality: satisficing pop
-underperforms optimizer in simple games but outperforms in noisy environments. Bounds: all
-outputs include upper/lower.
+propagates through pop network above threshold, not below. Bounded confidence: opinion clusters
+form at predicted cluster count for given $d$. Complex contagion: multiple-exposure requirement
+produces slower but more robust adoption than simple contagion. Bandit: explore-exploit tradeoff
+produces adapting populations. MFG: equilibrium converges for large N; matches discrete sim for
+small N. Pump-dump: 4-phase cycle detected on synthetic manipulation data. Bounds: all outputs
+include upper/lower.
 
 ## 10. Open items
 - Pop archetype catalog (which behavioral types? how many?).
 - Network topology for sentiment contagion (small-world? scale-free?).
 - Calibration from real market data — how to infer pop distribution from observable price action.
+- MFG tensor-train rank $r$ (accuracy vs. compute tradeoff).
+- Hegselmann-Krause confidence bound $d$ — fixed or adaptive?
 - Cross-sim interaction: do sociological predictions feed into M3c (AMM) or M3d (MEV)?
diff --git a/core/docs/plans/M3c-amm-liquidity-sims.md b/core/docs/plans/M3c-amm-liquidity-sims.md
index 3be5158..422cc4b 100644
--- a/core/docs/plans/M3c-amm-liquidity-sims.md
+++ b/core/docs/plans/M3c-amm-liquidity-sims.md
@@ -31,6 +31,15 @@ invariant calculations (Solidity-equivalent precision). Python, Rust, or Julia.
   time_horizon: "7d", sim_type: "amm_liquidity" }` — projected impermanent loss for ETH/USDC pool.
   Example: `{ value: 0.082, lower_bound: 0.041, upper_bound: 0.127, confidence: 0.85,
   time_horizon: "30d", sim_type: "amm_liquidity" }` — net LP return (fees − IL).
+- **Time-horizon mapping** (all run concurrently, ≥ 360:1 speed):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Tick–hourly | Slippage curves, invariant state, JIT liquidity | Every swap event |
+  | Daily | IL accumulation, fee income, LP profitability | Hourly roll |
+  | Weekly | Optimal LP range recalculation, pool composition | Daily roll |
+  | Monthly | LP strategy evolution (passive vs. active rebalance) | Weekly roll |
+  | Annual | Pool lifecycle, fee tier competitiveness | Monthly roll |
+  | 5-year | AMM design evolution, concentrated liquidity adoption | Quarterly roll |
 - **Prediction types:** `impermanent_loss`, `pool_return`, `optimal_range`, `slippage_estimate`,
   `lp_withdrawal_threshold`.
 - Calibration: ingests `dex_pool_state` and `price_tick` from M2.
diff --git a/core/docs/plans/M3d-mev-adversarial-sims.md b/core/docs/plans/M3d-mev-adversarial-sims.md
index 1e0e6bb..18ec3fa 100644
--- a/core/docs/plans/M3d-mev-adversarial-sims.md
+++ b/core/docs/plans/M3d-mev-adversarial-sims.md
@@ -1,69 +1,114 @@
 # M3d — MEV & adversarial extraction sims
 
 ## 1. Component
-Maximal Extractable Value simulation: models **transaction ordering as an optimization problem**,
-**Priority Gas Auctions (PGA) as all-pay auctions**, and **block building as a multidimensional
-knapsack problem**. Pops here are **searcher bots, block builders, and validators** competing for
-extractable value. Grounded in ACM MEV game theory [3] and knapsack auction literature [4,5].
+Maximal Extractable Value and adversarial simulation: models **transaction ordering as an
+optimization problem**, **Priority Gas Auctions (PGA) as all-pay auctions**, **block building as a
+multidimensional knapsack problem**, **cross-chain adversarial arbitrage**, and **Dynamic
+Stackelberg Mean-Field Games (DSMFG) for protocol-level adversarial policy**. Pops here are
+**searcher bots, block builders, validators, cross-chain arbitrageurs, and adversarial
+manipulators** competing for extractable value across **six concurrent time horizons**.
+
+Grounded in ACM MEV game theory [3], knapsack auction literature [4,5], Kolokoltsov adversarial
+dynamics (non-linear Fokker-Planck with WENO shock capturing), and DSMFG bilevel optimization
+(leader policy + follower MFG equilibrium).
 
 ## 2. Status / certainty
-DESIGN-FIRST · ABSENT. PGA-as-all-pay-auction model C4 (ACM [3]); knapsack formulation C4
-(Cornell [4,5]); simulation parameterization C1.
+DESIGN-FIRST · ABSENT. PGA-as-all-pay-auction C4 (ACM [3]); knapsack formulation C4 (Cornell
+[4,5]); cross-chain arbitrage C4 (ACM SIGMETRICS 2025, 5.5x growth since Dencun); DSMFG bilevel
+optimization C3 (emerging — SMFRL solvers); Kolokoltsov adversarial C3 (non-linear Fokker-Planck;
+WENO discretization established but crypto application novel). Parameterization C1.
 
 ## 3. Language & location
-TBD · `src/economy/sims/mev/`. Needs combinatorial optimization (for knapsack) and continuous-time
-auction modeling. Python (PuLP/OR-Tools for optimization), Rust, or Julia.
+TBD · `src/economy/sims/mev/`. Needs combinatorial optimization (PuLP/OR-Tools for knapsack),
+continuous-time auction modeling, PDE solvers (WENO for shock-capturing in adversarial dynamics),
+and bilevel optimization (DSMFG). Python, Rust, or Julia.
 
 ## 4. Does / does-not
 - **Does:** simulate Priority Gas Auctions where multiple searcher bots compete for the same
   arbitrage opportunity $V$ by bidding gas fees $g$ in a continuous-time all-pay auction; model
   block building as a multidimensional knapsack problem (scarce block space, heterogeneous
   transaction values/sizes); simulate endogenous selection cutoffs under paid-priority ordering;
-  predict MEV exposure for proposed trades; produce bounded predictions on extraction risk and
-  optimal gas strategies.
-- **Does-not:** extract MEV itself (this is a simulator, not a searcher); model AMM mechanics
-  (M3c handles pool math); model social dynamics (M3b).
+  model **cross-chain adversarial arbitrage** with inventory vs. bridge execution trade-off:
+  $\pi_{\text{inv}} = (P_{\text{src}} - P_{\text{dst}} - \text{slippage} - \text{gas}) \times q$
+  vs. $\pi_{\text{bridge}} = (P_{\text{src}} - P_{\text{dst}} - \text{fee} -
+  \text{depreciation}(\Delta t)) \times q$ where bridge latency $\Delta t \approx 242$s vs.
+  inventory $\Delta t \approx 9$s; solve **Dynamic Stackelberg MFG** for adversarial policy
+  design — bilevel optimization where a leader (protocol/regulator) sets policy and followers
+  (searchers) respond as an MFG equilibrium: the leader solves
+  $\min_\alpha J_L(\alpha, m^*(\alpha))$ subject to $m^*(\alpha)$ being the MFG Nash equilibrium
+  of followers under policy $\alpha$; model **Kolokoltsov adversarial dynamics** via non-linear
+  Fokker-Planck: $\partial_t m + \nabla \cdot (b(x,m)m) = \frac{1}{2}\nabla^2(\sigma^2 m)$
+  with WENO shock-capturing for discontinuous adversarial strategies; predict MEV exposure for
+  proposed trades; produce bounded predictions on extraction risk across all six time horizons.
+- **Does-not:** extract MEV itself (simulator, not a searcher); model AMM pool math (M3c);
+  model social dynamics (M3b); execute cross-chain bridges (Marketplace does).
 
 ## 5. Interface contract
 - Implements `query(PredictionQuery) -> BoundedPrediction` per M3 hub.
-- **Output bounds:** extraction probability ranges and gas cost intervals.
-  Example: `{ value: 0.23, lower_bound: 0.11, upper_bound: 0.38, confidence: 0.80,
-  time_horizon: "next_block", sim_type: "mev_adversarial" }` — probability this trade gets
-  sandwiched.
-  Example: `{ value: 14.7, lower_bound: 8.2, upper_bound: 22.5, confidence: 0.75,
-  time_horizon: "next_block", sim_type: "mev_adversarial" }` — optimal gas bid (gwei) for
-  a given opportunity.
+- **Output bounds:** extraction probability ranges, gas cost intervals, cross-chain profit
+  bounds, DSMFG equilibrium stability ranges.
+- **Time-horizon mapping** (all run concurrently, ≥ 360:1 speed):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Tick–hourly | PGA auctions, sandwich detection, cross-chain arb | Every block |
+  | Daily | Knapsack builder strategies, MEV landscape | Hourly roll |
+  | Weekly | Searcher population dynamics, cross-chain flow patterns | Daily roll |
+  | Monthly | DSMFG policy equilibria, adversarial strategy evolution | Weekly roll |
+  | Annual | Kolokoltsov adversarial long-run dynamics | Monthly roll |
+  | 5-year | Structural MEV regime shifts, protocol-level policy effects | Quarterly roll |
+- Examples:
+  `{ value: 0.23, lower_bound: 0.11, upper_bound: 0.38, confidence: 0.80,
+  time_horizon: "next_block", sim_type: "mev_adversarial" }` — sandwich probability.
+  `{ value: 14.7, lower_bound: 8.2, upper_bound: 22.5, confidence: 0.75,
+  time_horizon: "next_block", sim_type: "mev_adversarial" }` — optimal gas bid (gwei).
+  `{ value: 0.034, lower_bound: 0.018, upper_bound: 0.052, confidence: 0.82,
+  time_horizon: "1h", sim_type: "mev_adversarial" }` — cross-chain arb profit (ETH).
 - **Prediction types:** `sandwich_probability`, `frontrun_risk`, `optimal_gas_bid`,
-  `block_inclusion_probability`, `mev_exposure`.
-- Calibration: ingests `on_chain_event` (mempool-like data) and `price_tick` from M2.
+  `block_inclusion_probability`, `mev_exposure`, `cross_chain_arb_profit`,
+  `adversarial_policy_stability`, `searcher_population_shift`.
+- Calibration: ingests `on_chain_event` (mempool-like data), `price_tick`, and cross-chain
+  bridge state from M2.
 
 ## 6. Dependencies & stubs
-- M2 Data Feeds — on-chain events and gas data; *stub:* canned mempool snapshots.
+- M2 Data Feeds — on-chain events, gas data, cross-chain state; *stub:* canned snapshots.
 - M3 Sims hub — lifecycle management; *stub:* manual init.
 - M3c AMM sims — pool state for arbitrage opportunity detection; *stub:* fixed pool state.
 
 ## 7. Invariants / laws
-- **L1 (C4):** PGA is modeled as an **all-pay auction** — all bidders pay their gas whether they
+- **L1 (C5):** PGA is modeled as an **all-pay auction** — all bidders pay their gas whether they
   win or not. The sim must capture this cost structure (not winner-pays-only).
-- **L2 (C4):** block building is a **knapsack problem, not a queue** — builders optimize for
+- **L2 (C5):** block building is a **knapsack problem, not a queue** — builders optimize for
   total extracted value subject to gas limit constraints, not first-come-first-served.
-- **L3 (C3):** MEV exposure predictions are **pre-trade** — traders query this sim *before*
+- **L3 (C4):** MEV exposure predictions are **pre-trade** — traders query this sim *before*
   submitting to the Marketplace to understand their extraction risk.
+- **L4 (C4):** cross-chain arb models **both execution paths** — inventory (fast, capital-
+  intensive) and bridge (slow, capital-light) — never assumes one dominates.
+- **L5 (C3):** DSMFG solutions are **bilevel** — the leader's optimal policy depends on the
+  followers' MFG equilibrium, which itself depends on the leader's policy. Fixed-point iteration
+  or SMFRL solvers required.
 
 ## 8. Build steps
 1. Implement the PGA all-pay auction model (N searchers, opportunity value V, gas bids).
 2. Implement the block-building knapsack solver.
 3. Add sandwich/frontrun detection heuristics.
-4. Wire M2 on-chain data → calibration of searcher population and gas dynamics.
-5. Wire pre-trade query interface for Traders.
+4. Implement cross-chain arb model (inventory vs. bridge, latency, MEV exposure).
+5. Implement Kolokoltsov non-linear Fokker-Planck with WENO discretization.
+6. Implement DSMFG bilevel solver (leader policy + follower MFG equilibrium).
+7. Wire M2 on-chain + cross-chain data → calibration.
+8. Wire pre-trade query interface for Traders.
 
 ## 9. Tests
 All-pay: losing bidders still pay gas cost. Knapsack: builder selects optimal transaction set
 under gas limit. Sandwich: known sandwich-vulnerable trade flagged; non-vulnerable trade clear.
-Bounds: all outputs bounded. Pre-trade: query does not submit any transaction.
+Cross-chain: inventory path preferred when latency advantage exceeds capital cost. DSMFG: leader
+policy converges to fixed point with follower equilibrium. Kolokoltsov: WENO captures shock
+discontinuities in adversarial strategy distribution. Bounds: all outputs bounded. Pre-trade:
+query does not submit any transaction. Speed: sim advances ≥ 360:1.
 
 ## 10. Open items
 - Mempool data access (public mempool? private order flow?).
 - Which MEV types to model initially (sandwich, backrun, liquidation, JIT?).
 - Multi-block MEV (cross-block extraction strategies).
-- Integration with M3c (arbitrage opportunities arise from AMM pool state).
+- DSMFG solver choice (SMFRL? fictitious play? direct bilevel optimization?).
+- WENO order for Kolokoltsov (3rd? 5th? tradeoff with compute budget).
+- Which L2s/bridges to model for cross-chain (Arbitrum? Optimism? Base?).
diff --git a/core/docs/plans/M3e-tokenomics-macro-sims.md b/core/docs/plans/M3e-tokenomics-macro-sims.md
index 01e7265..61bee9f 100644
--- a/core/docs/plans/M3e-tokenomics-macro-sims.md
+++ b/core/docs/plans/M3e-tokenomics-macro-sims.md
@@ -2,19 +2,26 @@
 
 ## 1. Component
 Macro-level token economy simulation: models **token supply dynamics, monetary policy (halvings,
-burns, inflation), and systemic stock-flow balances** using stochastic differential equations
-(SDEs) and state-space models. Pops here are **aggregate behavioral cohorts** (miners/validators,
-holders, speculators, protocol treasuries) whose collective behavior drives token-level dynamics.
-Grounded in the Vienna University complex-systems token modeling [7] and ResearchGate engineering
-token economy frameworks [6].
+burns, inflation), lending protocol dynamics, DeFi systemic risk, and stock-flow balances** using
+stochastic differential equations (SDEs), state-space models, kinked interest rate curves, and
+inter-protocol credit exposure networks. Pops here are **aggregate behavioral cohorts**
+(miners/validators, holders, speculators, protocol treasuries, borrowers/lenders) whose collective
+behavior drives token-level dynamics across **six concurrent time horizons** at ≥ 360:1 speed.
+
+Grounded in Vienna complex-systems token modeling [7], ResearchGate engineering token economy
+frameworks [6], Aave/Compound kinked interest rate models (industry standard), and DeXposure
+inter-protocol credit propagation (Matzakos et al. 2025).
 
 ## 2. Status / certainty
 DESIGN-FIRST · ABSENT. SDE state-space framework C4 (Vienna [7]); stock-flow modeling C4
-(ResearchGate [6]); specific token model parameters C1.
+(ResearchGate [6]); kinked interest rate model C5 (Aave/Compound production standard);
+DeXposure inter-protocol credit propagation C3 (emerging, 2025 — high DeFi specificity);
+composable yield optimization C4 (Yearn v3, Beefy, production-validated). Specific parameters C1.
 
 ## 3. Language & location
-TBD · `src/economy/sims/tokenomics/`. Needs SDE solvers (Euler-Maruyama, Milstein) and
-state-space estimation. Julia (DifferentialEquations.jl), Python (scipy), or Octave.
+TBD · `src/economy/sims/tokenomics/`. Needs SDE solvers (Euler-Maruyama, Milstein),
+state-space estimation, and VAR (vector autoregression) for credit exposure impulse responses.
+Julia (DifferentialEquations.jl), Python (scipy), or Octave.
 
 ## 4. Does / does-not
 - **Does:** simulate token state dynamics via the SDE framework:
@@ -22,51 +29,90 @@ state-space estimation. Julia (DifferentialEquations.jl), Python (scipy), or Oct
   $u$ is the behavioral policy function, deterministic drift captures programmatic parameters
   (halvings, burns), and Brownian motion $\sigma dW_t$ captures stochastic behavioral shocks;
   model stock-flow balances (circulating supply, staked, locked, burned); simulate monetary
-  policy impacts (halving events, fee burns, treasury emissions); produce bounded predictions
-  on token supply trajectories, inflation rates, and velocity.
+  policy impacts (halving events, fee burns, treasury emissions); model **lending protocol
+  dynamics** via kinked interest rate curves:
+  $R = R_0 + R_{\text{slope1}} \times U$ if $U \leq U_{\text{opt}}$,
+  $R = R_0 + R_{\text{slope1}} \times U_{\text{opt}} + R_{\text{slope2}} \times
+  (U - U_{\text{opt}})$ if $U > U_{\text{opt}}$ where $U = \text{Borrowed}/(\text{Supplied} +
+  \text{Borrowed})$, $U_{\text{opt}} \approx 0.8$ — cascade liquidations when
+  $\text{collateral} \times \text{LTV} < \text{borrowed}$; model **DeFi systemic risk** via
+  DeXposure inter-protocol credit propagation:
+  $E_{ij}(t) = \sum_{\text{tokens}} [\text{TVL}_j(\text{token}) \times
+  \text{ownership}_i(\text{token})]$ with VAR impulse responses for shock contagion across
+  protocols sharing collateral; model **composable yield optimization**:
+  $\max \sum_i w_i(t) \cdot \text{APY}_i(t) - \lambda \sum_i w_i(t)^2 \sigma_i^2(t)$
+  subject to $\sum_i w_i = 1$ — dynamic rebalancing across lending, LP, and staking strategies;
+  produce bounded predictions on token supply, protocol health, yield, and systemic risk.
 - **Does-not:** model individual transactions (M3c/M3d); model social sentiment (M3b);
-  model consensus mechanics (M3f).
+  model consensus mechanics (M3f); execute yield strategies (Traders/Marketplace do).
 
 ## 5. Interface contract
 - Implements `query(PredictionQuery) -> BoundedPrediction` per M3 hub.
-- **Output bounds:** SDE confidence bands (derived from the stochastic component $\sigma dW_t$).
-  Example: `{ value: 2.1, lower_bound: 1.4, upper_bound: 3.2, confidence: 0.90,
+- **Output bounds:** SDE confidence bands, utilization rate ranges, contagion impact intervals.
+- **Time-horizon mapping** (all run concurrently, ≥ 360:1 speed):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Hourly | Lending rates, utilization, liquidation risk | Every block |
+  | Daily | Yield optimization, protocol TVL flows | Hourly roll |
+  | Weekly | SDE supply trajectory, stock-flow balances | Daily roll |
+  | Monthly | DeXposure credit contagion, systemic risk | Weekly roll |
+  | Annual | Halving/burn policy impacts, inflation trajectory | Monthly roll |
+  | 5-year | Token supply long-run equilibrium, protocol lifecycle | Quarterly roll |
+- Examples:
+  `{ value: 2.1, lower_bound: 1.4, upper_bound: 3.2, confidence: 0.90,
   time_horizon: "90d", sim_type: "tokenomics_macro" }` — annualized inflation rate (%).
-  Example: `{ value: 0.67, lower_bound: 0.58, upper_bound: 0.74, confidence: 0.85,
-  time_horizon: "30d", sim_type: "tokenomics_macro" }` — staking ratio (fraction of supply).
+  `{ value: 0.67, lower_bound: 0.58, upper_bound: 0.74, confidence: 0.85,
+  time_horizon: "30d", sim_type: "tokenomics_macro" }` — staking ratio.
+  `{ value: 0.83, lower_bound: 0.78, upper_bound: 0.91, confidence: 0.88,
+  time_horizon: "1h", sim_type: "tokenomics_macro" }` — Aave ETH utilization rate.
+  `{ value: 0.12, lower_bound: 0.04, upper_bound: 0.25, confidence: 0.72,
+  time_horizon: "7d", sim_type: "tokenomics_macro" }` — systemic contagion risk index.
 - **Prediction types:** `supply_trajectory`, `inflation_rate`, `staking_ratio`,
-  `velocity_estimate`, `halving_impact`, `treasury_runway`.
-- Calibration: ingests `on_chain_event` (supply metrics, staking data) from M2.
+  `velocity_estimate`, `halving_impact`, `treasury_runway`, `utilization_rate`,
+  `liquidation_cascade_risk`, `systemic_contagion_index`, `optimal_yield_allocation`.
+- Calibration: ingests `on_chain_event` (supply metrics, staking data, lending protocol state,
+  TVL) from M2.
 
 ## 6. Dependencies & stubs
-- M2 Data Feeds — on-chain supply/staking data; *stub:* canned supply snapshots.
+- M2 Data Feeds — on-chain supply/staking/lending data; *stub:* canned supply snapshots.
 - M3 Sims hub — lifecycle management; *stub:* manual init.
 
 ## 7. Invariants / laws
-- **L1 (C4):** the SDE framework is the **canonical representation** — all token dynamics are
+- **L1 (C5):** the SDE framework is the **canonical representation** — all token dynamics are
   expressed as drift + diffusion. Deterministic policy (halvings, burns) lives in the drift $f$;
   behavioral uncertainty lives in the diffusion $\sigma dW_t$.
-- **L2 (C4):** **stock-flow conservation** — tokens are never created or destroyed outside the
-  protocol's defined mechanisms. The sim must balance: circulating + staked + locked + burned =
-  total ever minted.
-- **L3 (C3):** macro sims operate on **aggregate cohorts, not individuals** — the state vector
+- **L2 (C5):** **stock-flow conservation** — tokens are never created or destroyed outside the
+  protocol's defined mechanisms. circulating + staked + locked + burned = total ever minted.
+- **L3 (C5):** lending rate curves are **kinked at $U_{\text{opt}}$** — the steep slope above
+  optimal utilization is a design invariant of Aave/Compound, not a parameter to smooth.
+- **L4 (C4):** macro sims operate on **aggregate cohorts, not individuals** — the state vector
   $X_t$ tracks population-level quantities (total staked, total circulating), not per-wallet.
+- **L5 (C4):** DeXposure contagion is **directional** — protocol A's exposure to protocol B
+  is not symmetric. The exposure matrix $E_{ij}$ is not assumed symmetric.
 
 ## 8. Build steps
 1. Implement Euler-Maruyama SDE solver for a simple token model (supply + staking).
 2. Define the state vector $X_t$ and drift/diffusion functions for a reference token.
 3. Add stock-flow accounting (verify conservation).
-4. Wire M2 on-chain data → state estimation / calibration.
-5. Add monetary policy events (halving, burn) as drift discontinuities.
+4. Implement kinked lending rate model (Aave-style) with liquidation cascade simulation.
+5. Implement DeXposure credit propagation network with VAR impulse responses.
+6. Implement composable yield optimizer (risk-adjusted return maximization).
+7. Wire M2 on-chain data → state estimation / calibration.
+8. Add monetary policy events (halving, burn) as drift discontinuities.
 
 ## 9. Tests
 SDE: sample paths have correct mean (matches drift) and variance (matches diffusion). Stock-flow:
 conservation holds across all time steps. Halving: supply growth rate drops at halving event.
-Calibration: state estimate converges to observed data. Bounds: SDE confidence bands correctly
-cover realized paths on backtest.
+Lending: rate curve exhibits kink at $U_{\text{opt}}$; liquidation cascades triggered when
+collateral ratio breached. DeXposure: shock to protocol A propagates to protocol B through shared
+collateral; isolated protocols unaffected. Yield: optimizer rebalances toward highest risk-adjusted
+APY. Calibration: state estimate converges to observed data. Bounds: SDE confidence bands cover
+realized paths on backtest. Speed: sim advances ≥ 360:1.
 
 ## 10. Open items
 - Which tokens to model initially (ETH? BTC? a specific alt?).
 - State vector dimensionality (how many state variables per token model?).
 - Behavioral policy function $u(X_t, t)$ — how to parameterize aggregate cohort behavior.
 - Multi-token interactions (correlated diffusions across tokens?).
+- DeXposure graph granularity (how many protocols? top-10 by TVL?).
+- Lending model extensions (Morpho AdaptiveCurveIRM? variable kink parameters?).
diff --git a/core/docs/plans/M3f-consensus-staking-sims.md b/core/docs/plans/M3f-consensus-staking-sims.md
index bb3e06a..6e61d61 100644
--- a/core/docs/plans/M3f-consensus-staking-sims.md
+++ b/core/docs/plans/M3f-consensus-staking-sims.md
@@ -9,7 +9,8 @@ theorems [10], and Monash dynamic PBFT modeling [11].
 
 ## 2. Status / certainty
 DESIGN-FIRST · ABSENT. Evolutionary PoS game theory C4 (Cornell [8]); staking pool Nash
-equilibrium proofs C4 (ACM [10]); Markov chain throughput models C4 (Monash [11]);
+equilibrium proofs C4 (ACM [10]); Markov chain throughput models C4 (Monash [11]); MFG for
+validator populations C4 (Lasry & Lions 2007; validator-specific application C3);
 simulation parameterization C1.
 
 ## 3. Language & location
@@ -21,8 +22,12 @@ computation. Python, Julia, or R.
   under bounded rationality [8]; model staking pool delegation as a game with proven reward-
   parameter thresholds enforcing subgame-perfect Nash equilibria favoring honest validation over
   malicious slashing [10]; simulate **throughput stability under shifting validator states** via
-  Markov chains [11]; predict slashing risk, validator set stability, and staking yield; produce
-  bounded predictions on consensus health and staking returns.
+  Markov chains [11]; solve **Mean-Field Game equilibria for large validator populations** —
+  coupled HJB (individual validator optimization) + Fokker-Planck (population density):
+  $-\partial_t u + H(x, \nabla u) = F(x, m)$, $\partial_t m - \nabla \cdot (m \nabla_p H) = 0$
+  — captures emergent staking coordination without enumerating every validator; predict slashing
+  risk, validator set stability, and staking yield across **six concurrent time horizons** at
+  ≥ 360:1 speed; produce bounded predictions on consensus health and staking returns.
 - **Does-not:** validate blocks (this is a simulator); model AMM pools (M3c); model token supply
   (M3e — but consumes staking ratio from M3e as input).
 
@@ -34,8 +39,18 @@ computation. Python, Julia, or R.
   equilibrium.
   Example: `{ value: 4.2, lower_bound: 3.6, upper_bound: 5.1, confidence: 0.82,
   time_horizon: "30d", sim_type: "consensus_staking" }` — annualized staking yield (%).
+- **Time-horizon mapping** (all run concurrently, ≥ 360:1 speed):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Hourly | Markov chain validator state transitions | Every epoch |
+  | Daily | Replicator dynamics strategy shifts, slashing events | Hourly roll |
+  | Weekly | Staking pool Nash equilibrium recalculation | Daily roll |
+  | Monthly | MFG equilibrium for validator population | Weekly roll |
+  | Annual | Evolutionary stable strategies, yield trajectory | Monthly roll |
+  | 5-year | Consensus mechanism structural evolution | Quarterly roll |
 - **Prediction types:** `validator_honesty_fraction`, `slashing_probability`, `staking_yield`,
-  `pool_delegation_equilibrium`, `throughput_stability`, `consensus_liveness`.
+  `pool_delegation_equilibrium`, `throughput_stability`, `consensus_liveness`,
+  `mfg_validator_equilibrium`.
 - Calibration: ingests `on_chain_event` (validator set changes, slashing events) from M2.
 
 ## 6. Dependencies & stubs
@@ -56,8 +71,9 @@ computation. Python, Julia, or R.
 1. Implement the evolutionary honesty game (replicator dynamics, bounded rationality).
 2. Implement the Markov chain validator-state model.
 3. Reproduce the staking pool Nash equilibrium reward threshold from [10].
-4. Wire M2 validator data → calibration of transition rates.
-5. Wire M3e staking ratio input.
+4. Implement MFG solver (HJB + Fokker-Planck) for large validator populations.
+5. Wire M2 validator data → calibration of transition rates.
+6. Wire M3e staking ratio input.
 
 ## 9. Tests
 Equilibrium: honesty fraction converges to Nash equilibrium under stable payoffs. Markov:
diff --git a/core/docs/plans/M3g-market-microstructure-sims.md b/core/docs/plans/M3g-market-microstructure-sims.md
index 18bc84a..4d82e17 100644
--- a/core/docs/plans/M3g-market-microstructure-sims.md
+++ b/core/docs/plans/M3g-market-microstructure-sims.md
@@ -2,69 +2,100 @@
 
 ## 1. Component
 Market microstructure simulation: models **order flow, liquidity depth, slippage, spread dynamics,
-and cross-exchange arbitrage** at the fastest time scales (tick-level to hourly). Pops here are
-**market makers, takers, and arbitrageurs** interacting across multiple venues. The sim that
-operates at the highest temporal resolution — where M3a provides statistical forecasts and M3c
-models pool mechanics, M3g models the *plumbing* of how orders actually execute.
+optimal execution, and cross-exchange arbitrage** at the fastest time scales. Pops here are
+**market makers, takers, and arbitrageurs** interacting across multiple venues. Includes the
+**Almgren-Chriss optimal execution framework** for minimizing market impact of large orders.
+Operates at the highest temporal resolution — where M3a provides statistical forecasts and M3c
+models pool mechanics, M3g models the *plumbing* of how orders actually execute, across **six
+concurrent time horizons** at ≥ 360:1 speed.
 
 ## 2. Status / certainty
-DESIGN-FIRST · ABSENT. Order-book microstructure theory C4 (established academic field);
-DEX-specific microstructure C2 (emerging). Implementation C1.
+DESIGN-FIRST · ABSENT. Order-book microstructure theory C4 (established). Almgren-Chriss
+optimal execution C5 (industry standard since 2001; crypto adaptations validated 2023–2024,
+Kurz CMC thesis). DEX-specific microstructure C2 (emerging). Implementation C1.
 
 ## 3. Language & location
-TBD · `src/economy/sims/microstructure/`. Needs high-frequency data handling and event-driven
-simulation. Rust, C++, or Python with optimized event loop.
+TBD · `src/economy/sims/microstructure/`. Needs high-frequency data handling, event-driven
+simulation, and Riccati equation solvers for optimal execution trajectories. Rust, C++, or
+Python with optimized event loop.
 
 ## 4. Does / does-not
 - **Does:** simulate order flow across venues (DEXs and CEXs); model bid-ask spread dynamics as a
   function of inventory risk and adverse selection; simulate slippage curves for various order
-  sizes; model cross-exchange arbitrage opportunities and their decay; operate at **tick-level
-  resolution** (sub-second to minute); produce bounded predictions on execution quality, optimal
-  routing, and liquidity conditions.
+  sizes; solve **Almgren-Chriss optimal execution**:
+  $\min \int_0^T [\lambda \cdot x(t) \cdot \dot{x}(t) + \eta \cdot \dot{x}(t)^2] \, dt$
+  where $\lambda$ = permanent impact, $\eta$ = temporary impact, $x(t)$ = remaining order —
+  splits large orders across time to minimize market impact + timing risk; impact parameters
+  $\lambda, \eta$ re-estimated from order-flow streams every 10–30s; execution trajectory solved
+  via Riccati equations in <100ms; model cross-exchange arbitrage opportunities and their decay;
+  operate at **tick-level resolution** (sub-second to minute); produce bounded predictions on
+  execution quality, optimal routing, and liquidity conditions.
 - **Does-not:** model protocol consensus (M3f); model macro token supply (M3e); model social
   behavior (M3b); execute trades (Marketplace does).
 
 ## 5. Interface contract
 - Implements `query(PredictionQuery) -> BoundedPrediction` per M3 hub.
-- **Output bounds:** execution cost ranges and liquidity intervals.
-  Example: `{ value: 0.0034, lower_bound: 0.0018, upper_bound: 0.0052, confidence: 0.85,
-  time_horizon: "next_trade", sim_type: "market_microstructure" }` — expected slippage (%) for a
-  10 ETH market sell.
-  Example: `{ value: 12400, lower_bound: 8200, upper_bound: 18600, confidence: 0.78,
-  time_horizon: "1h", sim_type: "market_microstructure" }` — available depth (USD) within 50bps
-  of mid.
+- **Output bounds:** execution cost ranges, liquidity intervals, optimal trajectory envelopes.
+- **Time-horizon mapping** (all run concurrently, ≥ 360:1 speed):
+  | Horizon | Primary models | Update cadence |
+  |---------|---------------|----------------|
+  | Tick–hourly | Almgren-Chriss execution, slippage, spread, arb decay | Every tick |
+  | Daily | Liquidity regime, venue depth profiles | Hourly roll |
+  | Weekly | Cross-venue flow patterns, impact parameter drift | Daily roll |
+  | Monthly | Structural liquidity shifts, venue market share | Weekly roll |
+  | Annual | Microstructure regime (DEX vs CEX share evolution) | Monthly roll |
+  | 5-year | Venue topology evolution, structural impact trends | Quarterly roll |
+- Examples:
+  `{ value: 0.0034, lower_bound: 0.0018, upper_bound: 0.0052, confidence: 0.85,
+  time_horizon: "next_trade", sim_type: "market_microstructure" }` — slippage (%) for 10 ETH.
+  `{ value: 12400, lower_bound: 8200, upper_bound: 18600, confidence: 0.78,
+  time_horizon: "1h", sim_type: "market_microstructure" }` — depth (USD) within 50bps.
+  `{ value: [0.3, 0.3, 0.2, 0.1, 0.1], lower_bound: null, upper_bound: null, confidence: 0.80,
+  time_horizon: "30min", sim_type: "market_microstructure" }` — Almgren-Chriss optimal execution
+  schedule (fraction per 6-min bucket for 100 ETH sell).
 - **Prediction types:** `slippage_estimate`, `spread_forecast`, `depth_profile`,
-  `cross_venue_arb`, `optimal_execution_route`, `liquidity_score`.
+  `cross_venue_arb`, `optimal_execution_schedule`, `optimal_execution_route`,
+  `liquidity_score`, `impact_estimate`.
 - Calibration: ingests `price_tick`, `dex_pool_state`, and `execution_fill` from M2.
 
 ## 6. Dependencies & stubs
 - M2 Data Feeds — tick data and pool state; *stub:* canned order book snapshots.
+- M3a Statistical — volatility estimates for Almgren-Chriss timing risk; *stub:* fixed vol.
 - M3c AMM sims — pool mechanics for DEX venues; *stub:* fixed pool state.
 - M3 Sims hub — lifecycle management; *stub:* manual init.
 
 ## 7. Invariants / laws
-- **L1 (C4):** microstructure operates at the **highest temporal resolution** — predictions are
+- **L1 (C5):** microstructure operates at the **highest temporal resolution** — predictions are
   valid for seconds to hours, not days. Stale microstructure data is worse than no data.
-- **L2 (C4):** slippage is a **function of order size and current depth** — not a fixed
+- **L2 (C5):** slippage is a **function of order size and current depth** — not a fixed
   percentage. The sim must model the non-linear relationship.
-- **L3 (C3):** cross-venue arbitrage opportunities **decay** — the sim models the time-to-close
+- **L3 (C5):** Almgren-Chriss impact parameters $\lambda, \eta$ are **estimated from live data**,
+  never hardcoded — crypto impact dynamics differ by asset, venue, and time-of-day.
+- **L4 (C4):** cross-venue arbitrage opportunities **decay** — the sim models the time-to-close
   of an arb opportunity, not just its existence.
+- **L5 (C4):** optimal execution trajectories are **re-solved on every significant state change**
+  (vol spike, depth drop, regime transition) — a stale trajectory is worse than naive execution.
 
 ## 8. Build steps
 1. Implement a simplified order-book simulator (limit orders, market orders, cancels).
 2. Add spread dynamics (inventory-based market maker model).
 3. Add slippage curves (order size → execution cost).
-4. Add cross-venue arb detection and decay modeling.
-5. Wire M2 tick data → calibration.
+4. Implement Almgren-Chriss optimal execution (Riccati solver, impact estimation).
+5. Add cross-venue arb detection and decay modeling.
+6. Wire M2 tick data → calibration of impact parameters.
+7. Wire M3a vol estimates → Almgren-Chriss timing risk component.
 
 ## 9. Tests
 Slippage: larger orders produce greater slippage. Spread: spread widens under adverse selection.
-Arb decay: detected arb opportunity closes over time. Depth: depth profile matches order book
-state. Bounds: all outputs bounded. Resolution: predictions update at tick frequency.
+Almgren-Chriss: optimal trajectory minimizes total cost vs. naive execution on backtest; impact
+parameters update when market conditions change. Arb decay: detected arb opportunity closes over
+time. Depth: depth profile matches order book state. Bounds: all outputs bounded. Resolution:
+predictions update at tick frequency. Speed: sim advances ≥ 360:1.
 
 ## 10. Open items
 - CEX order book data access (API limitations, costs).
 - DEX-specific microstructure (AMM pools don't have order books — translate pool state to
   equivalent depth/spread).
 - Latency modeling (how fast can our traders actually reach an arb?).
+- Almgren-Chriss extensions for crypto (volume-dependent variant? discrete block propagation?).
 - Which venues to model initially.