Verifiable, auditable, correctable.

AI research studio specializing in security & invariants. · caliperforge.com

CaliperForge researches autonomous systems you can verify, audit, and correct. Every shipped harness carries a clean reference where the invariant holds, a planted-bug twin where it fires, and a CI-asserted record of what our own gates caught. The receipts are the proof, not a claim — see what we caught.

AI involvement is disclosed at point of use; full policy at caliperforge.com/ai-disclosure and in ./AI_DISCLOSURE.md.

One operator (Michael Moffett, named below) orchestrates a small team of specialized AI agents — an audit engineer, a Cairo specialist, a Rust / Anchor specialist, a grant writer, a content reviewer, and others as the work calls for them. Each agent is scoped, has a defined output discipline, and cross-checks the others before anything leaves the workspace. The operator reviews every load-bearing decision, runs cold-environment reproductions, and ships under his own name as operator-of-record.

Specialization velocity is the working pattern: a new chain, a new VM, a new specialist and a CI-verified tool with it, in days rather than months. Cairo (snforge) → Solana / Anchor (on Asymmetric Research's Crucible) was roughly one working session. The engine is domain-general; the current portfolio is protocol and smart-contract security.

• euler-earn-invariants — defender-side stateful invariant harness for Euler Earn allocator and vault share accounting, running against real Euler Earn contracts (pinned submodule, no protocol mock). One case ships in the initial cut, encoded as a same-source clean / planted twin pair against Pashov Audit Group's M-01 finding on the PublicAllocator <-> EulerEarn share-tracking boundary (EulerEarn security review, 2025-07-25) and Euler Labs' fix commit f07f6b1c5e1e. Both twins run in CI on every push: the clean twin holds the encoded invariants under a stateful campaign; the planted twin single-hunk-reverts the fix and fires INVARIANT VIOLATED m01_reallocation_movement_matches_request plus INVARIANT VIOLATED m01_reallocation_flowcap_matches_movement. Regression fixture, not an audit or a discovery engine. GPL-2.0-or-later (own code); vendored Euler Earn consumed under its GPL branch. See caliperforge/euler-earn-invariants.

• uniswap-v4-invariants — defender-side invariant harness for Uniswap v4 hooks. Recurring bug classes expressed as stateful invariants on real v4-core, each shipped as a clean and planted-bug twin pair that both run in CI on every push. Latest addition, went public 2026-07-02. Apache-2.0. See caliperforge/uniswap-v4-invariants.

• invariant-atlas — the Exploit→Invariant Atlas: six historical DeFi exploits across four VMs (Cairo, Move/Sui, Solana, EVM), each reconstructed as a same-source clean / planted twin where an invariant property would have caught the bug class in pre-deploy CI. First cross-VM, defender-side, pre-deploy CI benchmark of this shape; clean passes / planted fires, both asserted on every push. Apache-2.0. See caliperforge/invariant-atlas.

• cf-invariants — AI-suggested stateful invariant testing for Cairo 2.x on top of snforge. Twelve reference contracts deployed and Voyager-verified on Starknet Sepolia. Apache-2.0. See caliperforge/cf-invariants.

• chimera-template-pack — reusable Foundry + Recon Chimera scaffold for EVM build-to-win contest entries.

• hyperevm-safety — open-source library of invariants and CI-runnable property tests for HyperEVM lending protocols that consume HyperCore oracle reads. Six HyperCore-boundary invariants in v0.1 (oracle staleness, mark / oracle deviation, szDecimals round-trip, precompile gas DoS, CoreWriter solvency window, Chainlink-compat invariant catching adapters that defeat staleness). All six run as CI-runnable property tests against a clean reference under fuzz. Three fire INVARIANT VIOLATED on the same CI run — the szDecimals round-trip (D-3) and Chainlink-compat (D-6) planted twins, and the JELLY (Mar 2025) mark-price manipulation reproduction (covering oracle staleness and mark / oracle deviation, D-1/D-2) against a minimal lending-market reference. Precompile gas DoS (D-4) and CoreWriter solvency window (D-5) carry inline broken-reference tests demonstrating the bug class deterministically. Built on hyper-evm-lib and the CaliperForge chimera-template-pack. Apache-2.0. See caliperforge/hyperevm-safety.

• bsc-invariants — open-source library of invariants and CI-runnable property tests for BNB Smart Chain DeFi protocols, starting with PancakeSwap v3 (the canonical BSC DEX). Five PancakeSwap v3 invariants ship in v0.0.2: P-1 FeeGrowthGlobalMonotonicity, P-2 TickInBounds, P-3 SqrtPriceX96InBounds, P-4 LiquidityEventConsistency, and P-5 FeeGrowthOutsideConsistency. P-1 ships with a planted-twin CI pair: clean passes, planted leg fires INVARIANT VIOLATED feeGrowth_neverDecreases and exits non-zero on the same CI run. Venus (M2) and Stargate (M3) harnesses are planned, not started. Not an audit; not a runtime monitor. Apache-2.0. See caliperforge/bsc-invariants.

• cf-invariants-verus-bridge-conservation — CI-verified reference for the cross-side conservation invariant class in lock/mint bridges, anchored on the Verus–Ethereum bridge exploit of 2026-05-18 (reported losses USD 11.58M, per Halborn). The conservation rule existed in prose; this repo lifts it to a one-line property: sum_locked_eth − sum_released_eth == sum_minted_verus − sum_burned_verus. Clean reference holds (0 violations); planted-bug twin breaks it (counterexample emitted; invariants_violated: 1 on the scorecard). CI matrix asserts both outcomes on every push. Not an audit; not a forensic report. Chimera-pattern harness (Recon Chimera + Foundry, Echidna), tracking the CaliperForge chimera-template-pack pinning. Apache-2.0. See caliperforge/cf-invariants-verus-bridge-conservation.

• apart-global-south-lost-in-translation — Apart Global South hackathon eval harness: a language-conditioned detection-rate eval measuring whether AI code auditors catch planted bugs across EN, ES, PT, and CS surface code. Uses Atlas planted-bug twins as the ground truth; measures per-language detection rates to surface where auditor accuracy degrades under language shift. Research artifact and eval harness submitted for the Apart Global South track. Apache-2.0. See caliperforge/apart-global-south-lost-in-translation.

• cf-modeleval: planted-twin discrimination-power harness for AI safety properties. Measures whether an evaluation check has falsifiable discrimination power against a stated failure mode (prompt-injection resistance, sycophancy resistance) across three providers (Anthropic, OpenAI, Groq). Paired CLEAN / PLANTED legs with a fixed scorer and committed receipts; CI matrix re-asserts every cell on every push. Methodology contribution: how to make an AI safety eval falsifiable by pairing every CLEAN claim to a PLANTED control. Not a model ranking; not a jailbreak discovery. Apache-2.0. See caliperforge/cf-modeleval.

• solana-invariant-atlas is the flagship Solana catalog: recurring bug classes as clean and planted twin pairs on real Solana programs, running in CI on every push. Apache-2.0. See caliperforge/solana-invariant-atlas. The atlas cites its member ports (each a defender-side regression fixture on the atlas's pinned rails, anchor-lang 1.0.1 and Crucible v0.2.0) and consumes the anchor invariant-authoring crate as a pinned git dependency. Ports live on their own repos, each with its own CI:

  • cf-invariants-jito: Jito tip-distribution port; four class-shaped invariants against a clean reference and four single-hunk planted twins.
  • cf-invariants-jito-tippayment: Jito tip-payment port; three class-shaped invariants against a clean reference and three planted twins.
  • cf-invariants-jito-priorityfee: Jito priority-fee-distribution port; one monotonic-accounting invariant against a clean reference and one planted twin.
  • cf-invariants-pyth: Pyth Solana Receiver port; two class-shaped invariants against a clean reference and two planted twins.
  • cf-invariants-anchor: pinned Anchor invariant-authoring crate the atlas driver consumes (core, suggest, emit, report).

• taiko-equivalence — differential equivalence tests for Taiko's Type-1 Ethereum-equivalence guarantee, encoded as clean / planted twin Foundry projects. AI-augmented, hard-disclosed. Apache-2.0. See caliperforge/taiko-equivalence.

• soroban-invariant-atlas — Soroban (Stellar) stateful invariant atlas: clean / planted twin fixtures over already-public Soroban findings. AI-augmented, hard-disclosed. Apache-2.0 OR MIT. See caliperforge/soroban-invariant-atlas.

• 12 Cairo contracts Voyager-verified
• 5 Solana programs CI-green (Jito tip-distribution; Jito tip-payment; Jito priority-fee-distribution; Pyth Solana Receiver; Anchor reference)
• 1 operator of record

What our own gates caught — published with provenance — at caliperforge.com/what-we-caught.

• Active language coverage: Cairo, Rust / Anchor, Solidity.
• Active chains: Starknet, Solana, HyperEVM, Ethereum / EVM. Base and the Optimism Superchain come online as ecosystem work lands; Move (Sui, Aptos) and Go (Cosmos, Celestia) when contract volume warrants.
• What we author: stateful-invariant suites, AI-suggested invariants tagged in source, CI-verified clean / planted-bug reference examples, contest-ready scaffolds, security tooling on top of incumbent fuzzers (snforge, Crucible).
• What we don't claim: coverage parity or superiority with the underlying fuzzing engines. We author the layer above them.

Operated by Michael Moffett, accountable for every commit, PR, grant application, and bounty claim made under this org. Reachable. KYC-able. Accountable to a named human.

• Operator: Michael Moffett
• Direct: michael@caliperforge.com
• Team: team@caliperforge.com
• Web: caliperforge.com
• X: @caliperforge
• Farcaster: @caliperforge
• Telegram: @caliperforge
• ENS: caliperforge.eth

For grant collaboration, hired-gun engagements, security tooling inquiries, or contribution questions: team@caliperforge.com.