TESTNET BUILD — no mainnet deployment, no real value. Nothing on this site is an offer or a live financial product.

← Audit register

Rounds 3–4 · 20–25 July 2026

Pre-mainnet campaign

Internal review. This is an internal engineering review of source, not an independent external audit, and it does not by itself authorize a production launch. It reviews the source at the baseline below, which is not necessarily identical to the code deployed on any network. No mainnet deployment has been authorized or attempted. Nothing here is a securities-law representation; token characterization is a matter for counsel.

Scope
The whole codebase — contracts, deployment, validation and QA scripts, invariant harnesses, the frontend, and the governing documents — assessed against both a Sepolia redeploy and mainnet readiness.
Method
A manual pre-mainnet review merged with an adversarial red-team fleet whose findings were each put to a refuter panel, then a 24-lens full-codebase register with executable proof-of-concept reproduction, then two go/no-go re-audits of the resulting fix diff.
Internal report
audit-reports/CONSOLIDATED_PRE_MAINNET_AUDIT_2026-07-20.md, PRE_MAINNET_SECURITY_REVIEW_2026-07-20.md, FULL_CODEBASE_AUDIT_2026-07-21.md, FINAL_FIX_AUDIT_2026-07-22.md, REAUDIT_H5_DELTA_2026-07-22.md

Findings and remediation history

What this round was

This was the largest and most adversarial campaign run against the protocol, and the one that found the most serious defects. It ran in four connected stages over six days:

  1. A manual pre-mainnet review of contracts, frontend, deployment configuration, invariants and fork validation, producing the readiness register.
  2. An adversarial red-team fleet — per-contract, specialist and protocol-wide reviewers attacking independently on isolated local forks. Every submitted finding was put to a panel of independent refuters before it was accepted; of fifty-six submissions, thirty-six were refuted and roughly nine distinct findings survived deduplication. The severities published here are the panel-adjudicated values, not the values as filed.
  3. A twenty-four-lens full-codebase register in which each lens attacked independently and was forbidden to run the test suite, followed by a reproduction phase that turned forty-three findings into executable proofs. Every entry was labelled as reproduced or reasoning-only.
  4. Two go/no-go re-audits of the resulting fix diff, each run with the explicit posture that another incomplete fix should be assumed until disproven, and each requiring the original defect to be re-run and shown dead rather than argued dead.

What it found

The headline result was a Critical: under a fully-marked unrealized impairment, queue settlement could burn an entire queued position and pay out exactly zero. It was reproduced seven times.

Its history is the most instructive thing in this whole audit record. The same mechanism had already been filed a week earlier, in Round 2, as an Informational item — a queued request could become sub-wei dust and burn for nothing — and closed on the reasoning that an amount below one wei was immaterial. Its recommendation was never implemented. When the conservative redemption NAV was introduced days later, that change destroyed the premise: the clamp could price a whole position at zero, not at a sub-wei residual, and the burn proceeded anyway.

An Informational finding was accepted on an argument about magnitude rather than a bound in the code, and a later change falsified the argument while the code stayed as it was. That is the failure mode this campaign exists to catch, and it is why findings on this site record the premise of an acceptance rather than only its outcome.

Five further High findings concerned backing inflation through treasury-directed funding, a permanently stranded impairment pool after partial recovery, a dilution mint opened by re-pricing the yield vesting period, a mint gate that proved attestations existed without proving they attested to the facility's terms, and the absence of any on-chain force to mark a receivable down between a missed payment and a servicer default. All five are recorded closed, each by re-running the original reproduction rather than by inspection.

The campaign also recorded that a first round of parallel fixes was rejected in its entirety — every one reviewed as incomplete, nothing applied — with upgrade-safety blindness and knowingly-red test suites named as the two systemic causes.

What it did not close

Two findings were deferred rather than fixed, and both remain live commitments:

  • Absorbed junior capital has no on-chain claim on later recoveries. Accepted in principle, handled off-chain while Forest Road is itself the curator, and identified as a required build before a junior tranche or any third-party curator exists. The architecture document was corrected because it had described an on-chain leg the code does not implement.
  • The realized-loss magnitude is chosen by the servicer rather than bound to an attestation. Confirmed real, pre-existing, and folded into the documented attestation and servicer trust model. Constraining it is a counsel decision rather than a code decision.

One further item is open by design and is the single most important caveat on any testnet result reported anywhere on this site: the testnet deployment is deliberately testnet-shaped. It retains bootstrap administrative privileges, uses a mock stablecoin with no value, and runs with concentration limits open for ramp testing. The production validator refuses every one of those postures. A green result on that stack is evidence about the code, not evidence about a production configuration.

Method note on the numbers

Severities here are adjudicated, and several moved during adjudication — two findings filed High were reduced to Medium, one filed High was reduced to Low, and one was raised. Findings that were reproduced with an executable proof are distinguished throughout the internal reports from those established by reasoning alone; roughly a third of the full-codebase register was reasoning-only and is marked as such there.

Fourteen further reported claims were refuted outright and recorded as refuted, so that a later round would not re-derive them and count them again.

Why these reports are not published in full

The underlying reports for this campaign are held internally rather than published, and the reason is specific rather than reflexive. Several of them contain step-by-step, measured exploitation recipes for defects whose fix status is stated in a different document; the operational posture of the live testnet stack in enough detail to target it; and local filesystem and endpoint detail from the reviewers' own environments. Publishing an exploitation recipe alongside an incomplete statement of what is fixed would be worse than publishing nothing.

What is published is the finding, its adjudicated severity, and its current disposition — which is the part a reader needs in order to judge the protocol.

Limitations

The full-codebase register explicitly recorded what it did not cover: nothing was executed by the lenses themselves, coverage was not measured in that round, the handover script was unread, and economic calibration, most of the frontend, individual test files, manifest drift, observability, servicer-outage liveness, and gas, static-analysis and storage-layout diffs were all out of scope. Two of its findings concerned false greens in the test and invariant layers, both reproduced by mutation — a reminder that a passing suite is evidence only to the extent its assertions bind.

An independent external audit remains a required production gate.

Earlier roundFive-pass source review, Round 2→