From real mainnet to RL findings

A forked-EVM, agent-based, RL-trainable engine that produces VC-grade tokenomics reports — and proves it with bit-exact replays of real on-chain events.

6 min read

Mocking is fast and lies in basis points. Real Solidity on a pinned mainnet block doesn't lie. Here is what that costs and what it buys.

7 min read

A one-page architecture tour: from pyrevm at the bottom to the Gymnasium env at the top. Where to look when you add an agent, a protocol, or a new validation gate.

8 min read

Aave V3 ships bytecode-identical contracts across mainnet, BNB, Polygon, Arbitrum, Base, and Optimism — but the addresses, decimals, and seconds-per-block aren't identical. Here is how we route one adapter to all six, and how a single hardcoded 12 silently corrupted RL training on five of them.

6 min read

Compound V3's isolated-base-asset design is a fundamentally different invariant from Aave V3's single-pool model. Borrow is implemented as withdrawing the base asset past zero; liquidations go through absorb(), not liquidationCall. Here is how the adapter wraps that.

4 min read

SparkLend is a fork of Aave V3 — bytecode-identical Pool contract, same ABI, same selectors, different deployment. Our adapter is a 30-line subclass. Sometimes the right amount of code reuse is deep reuse, not abstraction.

4 min read

Curve StableSwap is the AMM that lives on the credibility wall — invariant tuned for pegged pairs, fragile in exactly the regime where everyone needs it to work. Modeling a depeg cascade through Curve requires the StableSwap invariant, an oracle shock, and an agent population that responds. Here is the slice.

6 min read

Three runs across 4 GPU-hours of training. v2 (50K steps, local GTX 1650, $0) and v3 (200K steps, Modal A10G, $0.36) returned the same eval reward to all 16 decimal places. This is what eval saturation looks like — and why the deliverable is the reproducible $0.36 pipeline, not a fictional improvement number.

6 min read

0.1 WETH dropped into the mainnet WETH/USDC 0.05% pool. Three strategies — dump_all, TWAP, PPO — converge to $256.31656 ± 4e-6. Reward standard deviation is exactly zero. This is the second saturation regime in the series, and the second time the lesson is the same: the pipeline is the deliverable.

5 min read

PPO captured 63% of a scripted close-factor-max heuristic's profit on the single-liquidator-vs-one-borrower env. That's not a PPO failure — it's the expected single-agent outcome when the action space has one real degree of freedom. The same pipeline supports the competing-liquidator regime where RL beats the heuristic.

6 min read

EigenLayer's EIGEN token became transferable on 2024-10-01. We pin block 20,900,000, fork mainnet, execute the same swaps Uniswap V3's Quoter quoted, and assert delta == 0. This post is the full walkthrough.

4 min read

ENA's real launch happened on 2025-04-02. We pin the cliff block, replay the vesting-cliff scenario, and compare simulator outflow against on-chain truth — within an honest 30% tolerance band, not delta-zero. The number that came back was 1.299e19%. Here is why that's not a regression.

5 min read

Twenty-one channel-agnostic artifact bundles, one Typer command per scenario, one shape-aware HTML report template. The pipeline that turns a scenario YAML into a publishable artifact in two CLI invocations and zero manual editing.

5 min read

Same seed plus same scenario produces byte-identical run output. Not as a contract bolted on top; as a property that falls out of two rules the engine refuses to break. Here is why that's worth a release gate, and what it unlocks.

5 min read

Modal for the FastAPI service. Vercel for the dashboard and landing. DuckDB for run persistence. Auto-deploy on every push to main. Bearer-token auth — and the leak we caught before going public. The deployment that ties the engine to the dashboard. Series closes.

6 min read