Review complete. The findings file is written at the repository root and the scratch probe stays under the ignored test/scratch directory.

Outcome: no defects in the Access Control, Trust Gap or Asymmetry areas. The permission surface is two unguarded public functions on an ownerless engine plus a stock ERC-20, and every guard I inverted held.

What I verified:

• Access control. No roles, owners, initializers, proxies or delegatecalls exist. The constructor hands out nothing and uses msg.sender only to mint the fixed supply to the factory. The vendored OpenZeppelin files hash byte-for-byte to upstream v5.1.0, so no hidden mint or admin path. Exported ABIs match the build.
• Asymmetry. The observe and advanceEpoch writers mirror each other: observe adds to the count, histogram and observer table, advance resets count and histogram and invalidates the table by epoch tag. The deadline checks are exact complements at the boundary. Advancing and observing in one transaction is allowed and gives no ordering advantage.
• Trust gap. No value flows through the engine, so there is no seam where authorization and economics interact. The only external call is a single STATICCALL to balanceOf, confirmed from the runtime bytecode, which rejects the aderyn reentrancy lead. The slither uninitialized-local lines are zero-initialised memory arrays and are not defects.

One info-level note recorded with a measured reproduction: a holder of one QOBS unit can recycle it through 1,024 puppet contracts for about 103M gas, fill the epoch, and lock real holders out with EpochFull. This is the approved eligibility rule and is already disclosed in the README and security notes. It moves probability no further than one honest observation would.

Coverage: all five listed entry points answered, four holds and one carrying the info note, plus seven invariant rows. The launch manifest row is marked unreached because no launch.json exists in the tree yet. The deployment handoff prescribes ["$token"] for the engine, which is the correct argument shape for the manifest assignment to follow.

Review complete. The findings record is written to .imd-findings.json, no repository files were changed, and my scratch tests are removed.

Result for my area: the math holds. I found no precision, boundary or numerical-gap defect in QuantumModel or QuantumEngine.

What I verified, beyond reading:

• Square root. A verbatim copy of the Newton routine fuzzed over the model's full domain and at edge inputs always satisfies the floor property.
• Transform and apportionment. 324 vectors generated by the independent Python reference, including basis states, near-basis states, the documented extreme outputs, full and single observation histograms, and a 60-step chained evolution, all match the Solidity output exactly. Observed component bounds were exactly the documented 23,437 and 835,938.
• Overflow and casts. The largest intermediate is about 6.4e19, far below the uint256 limit, and every uint32 cast is bounded by the scale constant. Normalization denominators are provably nonzero at each stage.
• Engine boundaries. The observe and advance timestamp checks are exact complements. The capacity check cannot be skipped because the counter only moves by one or resets. The 1,025-probe bound exceeds the longest possible run of current entries, so the unreachable assert is truly unreachable.
• Static analysis leads. All slither uninitialized-local lines are declared-then-accumulated locals, and the aderyn reentrancy line is a state write after a STATICCALL. None reproduces.

One low finding outside my area. The committed checksum list in docs fails for seven vendored forge-std files when run on the committed tree. The listed hashes match upstream v1.9.7 byte for byte, and the tree's copies differ only in whitespace, so forge-std was reformatted after hashing. No runtime impact, but the security notes claim a check that does not pass. The reproduction is the documented command:

sha256sum -c docs/vendor-sha256.txt

Coverage rows answer all five listed entry points as holding, plus six invariant and documentation rows.

The review is complete. The findings file is written at .imd-findings.json and nothing in the tracked tree was changed.

Outcome. The economics, invariant and flow-gap passes found no fund-loss, custody, or conservation defect. The two contracts do what the brief specifies. I recorded two design-level economic findings, neither of which is a coding error, so the author has no blocking fix.

What I verified in my area

• Probability conservation holds. The largest-remainder apportionment always has enough nonzero remainders, so every transition sums to exactly 1,000,000 and every component stays within 23,437 and 835,938. The Solidity butterfly and Newton square root agree with the independent Python oracle in all 32 vectors.
• Observation counters and the fixed 2,048-slot observer table stay coupled. Epoch tags start at 1, so zero-initialised slots never read as current, and insert-only probing keeps every current entry reachable.
• The engine never moves tokens or ETH. Its only external interaction is a view balance read, so a failing token cannot block advancement.
• The vendored OpenZeppelin ERC20 hashes identically to the upstream v5.1.0 tag. ABI exports match the build. Formatting and all 60 project tests pass.

Findings reported

1. Low. One recycled QOBS captures a whole epoch. The gate is a point-in-time balance read per address, so one token passed through 1,024 puppet contracts fills every slot. Measured cost is about 84k gas per slot, roughly 86M gas per epoch, and a holder of one million QOBS then gets EpochFull. This follows directly from the approved eligibility rule and is already disclosed in the README, so any mitigation needs a requester scope decision.
2. Info. Linear-probe griefing. An attacker who packs 1,023 entries into one contiguous run raises one victim's observe cost from 82,617 to 2,830,841 gas measured cold. It stays within a block and is documented; I report it only to put a number on the bound.

Not reached. No launch.json exists yet, so constructor linkage was checked only against the deployment guidance, which correctly prescribes $token for the engine. The manifest reviewer must confirm that reference once the manifest exists.

The review is complete. The findings file is valid JSON, the cited snippet matches line 18 of the tree, and the working tree is clean apart from the ignored scratch directory.

Scope

Files read in full: src/QuantumEngine.sol, src/QuantumModel.sol, src/LaunchToken.sol, all six test files, both protected floors, README and the four docs. The three assigned guides (Execution Trace, Periphery, First Principles) and the validation gate were applied. forge build and forge test pass with 60 tests, zero failures.

Severity counts: 0 Critical · 0 High · 0 Medium · 1 Low

[L-1] Vendored-dependency integrity record does not match seven delivered forge-std files

Location: docs/vendor-sha256.txt:18 and six sibling lines. Root cause: The record holds upstream forge-std v1.9.7 hashes, but the committed copies of seven files were re-wrapped after vendoring. The diff against upstream is whitespace only. Reproduction: run the command below from the repository root. Seven lines report FAILED instead of OK.

sha256sum -c docs/vendor-sha256.txt

Impact: The record cannot confirm the delivered dependency bytes, which is its stated purpose. Deployed contracts are unaffected, since forge-std is test-only. Fix: Regenerate the record from the tree as committed, or restore the seven upstream files byte for byte.

What held

• observe has its checks in the documented order and writes only after the balance read. That read is a STATICCALL, so the aderyn reentrancy lead cannot execute. Every revert path leaves the count and table untouched.
• The observer table was stressed at exactly 1024 contiguous same-home entries in a scratch test. The probe loop finds the stale slot on its last permitted iteration and never reaches the assert. Worst-case cold lookup costs about 2.8M gas, which the README already documents as a griefing cost with no victim loss.
• advanceEpoch is the exact boundary complement of observe, performs one transition per call, resets count and histogram together, and makes no external call.
• The model library was re-derived independently in Python from the equations in the model document, not from the project's own reference script. Four fresh vectors plus the two documented ones match the Solidity output exactly. The largest-remainder apportionment always sums to one million, and the nonzero-total assert is unreachable from engine state.
• The token inherits five OpenZeppelin v5.1.0 files that are byte-identical to upstream. The protected floor expectations are reproduced by the project's own deployment test.
• Manifest linkage: no launch.json exists yet. The deployment guide specifies the token reference for the engine constructor, and neither constructor grants any role.

Slither's uninitialized-local lines are all memory arrays and counters that Solidity zero-initialises. The timestamp lines are the intended epoch gating. None became findings.

The review is complete. The findings file is in place, the scratch tests are removed, and the working tree is clean apart from the excluded findings file.

Outcome: no critical, high or medium defects. The code implements the approved brief faithfully, and the full in-tree suite passes with 74 tests and no failures.

Findings kept, after reproduction

1. Low, vendor checksum record. Running sha256sum -c docs/vendor-sha256.txt fails for seven forge-std files. I downloaded upstream forge-std v1.9.7 Vm.sol: its hash matches the record exactly, and the committed copy differs only in whitespace, so lib/ was reformatted after hashing. docs/security.md claims the check passed. The two specialist reports from math and flow were merged into this one. forge-std is test-only, and all five OpenZeppelin files are byte-identical to upstream.
2. Info, Sybil capacity capture. One recycled QOBS unit through 1,024 puppet contracts fills the epoch, a million-token holder then gets EpochFull, and the resulting vector equals a single honest observation. Economics and permissions reported the same root cause, so they are merged. I rated it info rather than low because the code implements the brief's stated eligibility rule and the risk is disclosed in the README and security notes. Any mitigation is a scope decision for the requester.
3. Info, probe-run gas griefing. Packing a contiguous run of 1,023 table entries raises one victim's observe cost to about 2.83M gas from about 83k. Bounded, fits in a block, and already documented.

Dropped leads

• Aderyn's high reentrancy flag: balanceOf is a view call compiled to STATICCALL, and the hostile-callback test in the tree confirms no state can change.
• All Slither uninitialized-local lines are default-zero locals. The best index starting at zero is the intended tie-break toward the lower state.
• Timestamp comparisons gate epochs by design and never seed randomness.

My own pass found nothing further. The manifest is schema-valid with $token as the engine's only argument and no owner or hard-coded wallet. The committed ABI exports are byte-identical to the compiler output. The model arithmetic bounds, largest-remainder loop termination and probe-table termination all hold as documented.

Implemented the frontend, static export and documentation. Build, typecheck, 19 tests and manifest/hash checks pass.

Validation report records evidence and limitations. Live swaps are disabled because the pool has zero active liquidity.

Git commit creation was blocked by read-only .git. Design documentation is in docs/DESIGN.md to respect the write scope.