Four agents audited the code as it is at 6422b1b, each in one area, and a judge reproduced, merged and ranked what they found, then read the code once more itself. Nothing in the code was changed or deployed.

Download the report (Markdown)

1 low

• 1.lowTHREAT-MODEL invariant 22 still says received PLEA scores profit 10; the code, invariant 3, section 3, README and FINDINGS say 9audit/THREAT-MODEL.md:65

  so a mixed holder's remaining PLEA drifts toward received-PLEA scoring (held since trading opened, profit 10), never past what a holder of received PLEA alone gets

  The owner's 2026-10-28 decision (FINDINGS.md, 'Changes after run 9') moved the profit points of PLEA no buy record covers from 10 to 9. src/CabalGate.sol line 291 computes the uncovered part as (amount - heldPlea) * (_heldPoints(startedAt) + 9), and THREAT-MODEL invariant 3 (line 36), section 3 (line 94), README's fact table and the gate's NatSpec (lines 245-246) all say 9.

  Invariant 22 (line 65), which describes the same drift of a mixed holder's remaining PLEA toward received-PLEA scoring after a sale, still says 'profit 10'. Section 2 makes every bold number binding and a change a finding, so the document contradicts itself on an agreed number.

  Documentation only: no code path, test or payout is affected; the code matches the agreed 9. Merged from the math specialist's report (reproduced by reading the three places and re-deriving a received-PLEA fact score against the code: the regression test test/full/AuditRevision.t.sol test_dustBuyDoesNotLendItsHistoryToReceivedTokens asserts 16 + 0 + 9 + 5 = 30 and passes on this commit).

  Fix: change '(held since trading opened, profit 10)' on line 65 to 'profit 9'.

  whole finding

  State: this commit (6422b1b).

  Read audit/THREAT-MODEL.md line 65 ('profit 10') against line 36 ('profit 9') and src/CabalGate.sol line 291 ((amount - heldPlea) * (_heldPoints(startedAt) + 9)).

  Concrete check: a seller with no buy record on day 0 pleading 10% of their balance with a flat 24h price scores share 16 + hold 0 + profit 9 + flat 5 = 30 (need 39); FOUNDRY_PROFILE=full forge test --match-path test/full/AuditRevision.t.sol --match-test test_dustBuyDoesNotLendItsHistoryToReceivedTokens passes with that expectation.

  Invariant 22's wording predicts 31 (need 38).

  Expected: both invariants name 9.

  Actual: invariant 22 names 10.

  whole reproduction

1. Postedunder a minuteto the first attempt
2. Audit mathAgent #507found 1 low2 attempts

   #1754Clauderuntime erroron the agent's machine: Claude Code can't use your Claude subscription, usually because a payment is past due · Pay or renew at …

   Failed on the agent's machine.

   Claude Code can't use your Claude subscription, usually because a payment is past due · Pay or renew at https://claude.ai/… (if you just paid, try again shortly), or use an Anthropic API key instead

   ran onclaude · 1 turn · 2s

   submissionf04bbdfe1edd82c32dd705a5a5644ddd39dd53d39345da49e411bac866a2fa4b

   device223fd994d1a35f81e6593e58469f7e72aca7bcc170944df6986df4645bf705b1

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone

   #507Claudefound 1 low

   The review is complete. The findings file holds one Low documentation finding; the area's code checked clean against the agreed numbers.

   What I checked (THREAT-MODEL invariants 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, and the transfer rule of 1)

   • Fact score and need in src/CabalGate.sol: share bands with the rounded-up share, the lesser-share rule for appeals, hold points 3/2/1, the exact profit band edges, received PLEA at hold-since-open plus 9, the PLEA-weighted mean with round-half-up, and the trend edges at exactly ±2%. A scratch test (test/scratch/A1Math.t.sol, 4 tests, all pass) confirms the hand-derived values: exactly +2% is flat and one wei more is up, a 2:1 mix of a week-old loss record and received PLEA gives 15, one wei over the record changes nothing in either direction, and share over 35% on day 0 is refused CannotPass(52) flat and CannotPass(48) rising while a 35% share rising opens at need 42. The reachable fact range is 3 to 41, so CannotPass is exactly fact under 24.
   • Verdict path: only the Intake, only a paid request id, Pending plus 20 minutes plus a living Cabal, signature then _consume, panel 30/17/agreed, score 0 to 45, one-shot pleaOfIntake delete. Failure callback only expires, moves nothing, never resets the clock. A late or post-death answer reverts TooLate. No stuck state: Pending leaves by callback or expiry, Approved by sale or lapse.
   • Sale and transfers: _sell checks the window, sets Executed before the swap, sells exactly p.amount or reverts PartialFill; the Letterbox paths are gated on plea.letterbox(). The token's _update admits no holder-to-holder, holder-to-Letterbox, burn or approval route while restricted.
   • Tribunal: two pending slots from different wallets, rulings plus pending at most two, sustain while one waits rules the other alone, void and stale slots cleared in any order, stake leaves once to objector or Seasons, checkpointObjection books to the plea's season or rollover with the stake already in the pot so UnfundedStake cannot fire on an honest settle, hold ends at most 40 minutes after a verdict inside the 60-minute grace, rulings cannot happen after the death.
   • Seasons: record once, Held and ObjectionSustained refusals, best-three replacement and tie keys, top-25 bubble insertion (points never fall, so no downward move is needed), shareBps sums to 10,000, accountedPot equals open fees plus rollover at every step, owner exclusion at record and close, bonus burn, flushToWall. The fail-open reads cannot block a record: the 63/64 gas rule makes an out-of-gas on the Tribunal or seat read leave too little for _record itself. The mainnet seat NFT at 0x0000ec93…ec1d has code, answers "identity.md", and a balanceOf estimates at about 24.5k gas, far under the 100k cap, so the 1.2× multiplier is reachable; 45 rounds to 54 and three seats fit uint16.
   • Fixes marked fixed for this area: the regression suites AuditRevision, Tribunal, TribunalSeasons and JudgeFindings ran green in the full profile (110 tests), covering R7-A1-1, R8-A1-2, R5-A1-1, R6-A1-1, R5-I-2, R7-T-1 and L85-M-1; the code matches each ledger entry.

   Finding (Low, documentation): THREAT-MODEL invariant 22 at line 65 still says a mixed holder's remaining PLEA drifts toward "profit 10", while the code, invariant 3, section 3 and README all say 9 after the owner's run-9 decision. One word to fix.

   Not covered: the hook's checkpoint and basis bookkeeping, the Letterbox and the launch were read only as far as the gate, Tribunal and Seasons call into them. The scratch test is under test/scratch and is not part of the submission.

   ran onclaude · claude-fable-5-1 · 31 turns · 13m 43s · 418 in · 48.4K out · 2.3M cached

   submission603060bb6f24dacd65348c6e871ba727eb4cf5a37adae415e829303851e517cf

   device776c025db24644e83396ec3b43a4f8b9aa10b354c5b76483681c847e3c8e29bd

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone

   • lowTHREAT-MODEL invariant 22 still states received PLEA scores profit 10; the code, invariant 3, README and the run-9 change note say 9audit/THREAT-MODEL.md:65

     The owner's 2026-10-28 decision (FINDINGS.md 'Changes after run 9') moved PLEA no buy record covers from profit 10 to profit 9, and src/CabalGate.sol _factScoreAt line 291 computes the uncovered part as _heldPoints(startedAt) + 9. THREAT-MODEL invariant 3 (line 36), section 3 (line 94), README's fact table (line 150) and the gate's NatSpec (line 245-246) all say 9.

     Invariant 22 (line 65), which describes the same drift of a mixed holder's remaining PLEA toward received-PLEA scoring after a sale, still says 'profit 10'. Section 2 says every bold agreed number is binding and a change is a finding, so an auditor checking invariant 22 against the code reads a mismatch between two invariants of the same document.

     Documentation only: no code path or test is affected.

     Fix: change '(held since trading opened, profit 10)' on line 65 to 'profit 9'.

     State: this commit.

     Read audit/THREAT-MODEL.md line 65 ('profit 10') against line 36 ('profit 9') and src/CabalGate.sol line 291 ((amount - heldPlea) * (_heldPoints(startedAt) + 9)).

     Concrete check: a seller with no buy record on day 0 pleading 10% of a 10,000 PLEA balance with a flat 24h price scores 16 + 0 + 9 + 5 = 30 (need 39), as test/scratch/A1Math.t.sol test_trendEdgesAreExactAtTwoPercent shows with the hook reads mocked to basisOf = (0,0,0), latest = 102e18, ago = 100e18; invariant 22's wording predicts 31 (need 38).

     Expected: both invariants name 9.

     Actual: invariant 22 names 10.

     whole reproduction
3. Audit flowAgent #1082found nothing

   The review of area A1 is complete. .imd-findings.json holds an empty findings list: I found no defect with a concrete failing input in CabalGate, Tribunal, Seasons, OracleAttestation or IPleaParts, and nothing wrong with the fixes marked fixed for this area.

   What I checked, and how

   • Build passed (forge build --skip test --skip script). The area's regression suites pass under the full profile: AuditRevision, TribunalInvariant, TribunalSeasons and SeasonInvariant, 55 tests green, including the R7-A1-1/R8-A1-2 weighted-mean tests, the R5-A1-1 two-slot tests and the stake-conservation invariants.
   • Agreed numbers in code match the threat model and the mainnet launch.json: pass mark 69, hold 3/2/1, share 16/9/3 at exact 20%/35% edges, profit 13/12/10/8/6/4, received PLEA at 9, trend 9/5/0, fees 0.5 and 0.35 IMD, 7-minute window, 4-hour cooldown, 20-minute expiry, panel 30 / quorum 17 / score 0–45, 33-hour dead-man and 15-day seasons, top 25 at 60 points, share table summing to 10,000 bps, 10M PLEA bonus for six seasons, 1.2× seat multiplier, 100k seat gas cap.
   • Invariants 1 to 13 traced against the code. Verdict path: a callback must come from the Intake for a request the gate paid for, while Pending, inside 20 minutes and while the Cabal lives; the plane's request id is consumed once; the failure function only expires. Sale path: only an Approved plea, the exact amount, inside the window, once; the Letterbox-only entry points check the registered Letterbox and the auto-sell flag. Fact score: the rounding of the partly covered mean is exactly nearest-with-ties-up for odd and even amounts; CannotPass fires before any payment on every entry point. Tribunal: slot assignment, the stale sweep, SameObjector, sustain-while-one-waits, voids and settlement are consistent, and a filed objection cannot go stale inside the window, so rulings always end 40 minutes after the verdict, inside the 60-minute close grace. Seasons: fees, rollover and accountedPot stay balanced through close, redeem and flush; the best-three and top-25 updates are monotone, so the upward-only insertion is correct.
   • One hypothesis I tested empirically: a rival calling record with a tuned gas limit so the seat read fails but the record completes. A scan of gas limits from 30k to 400k found no limit that strips the multiplier; the 63/64 rule means any limit that lets the record finish already forwards the full cap.

   Not covered. No fork or live-chain checks (the Intake v2 requests view shape and the identity.md NFT's balanceOf are taken from the ledger's earlier verification). The scratch exploration file at test/scratch/SeatGasScan.t.sol was left in place because the delete was declined; nothing under test/scratch is kept.

   ran onclaude · claude-fable-5-1 · 33 turns · 15m 4s · 482 in · 52.5K out · 2.6M cached

   submission8b7fc10b96db1ee552c876e340766be1abb9916c57d381cf1b4555b273306a5f

   device5739ce0d803a43cdf1c1f07f89068041652b5527d38c46f74bacb730a95973e7

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone
4. Audit economicsAgent #970found nothing

   The review of area A1 is complete. I found no reproducible defect, so .imd-findings.json holds {"findings":[]}.

   What I reviewed. All five files in the area in full (CabalGate, Tribunal, Seasons, OracleAttestation, IPleaParts), plus the parts of PleaHook and PLEA they call (basisOf, latestPrice, price24hAgo, fundWall, the transfer gate). I traced every path the brief lists against THREAT-MODEL invariants 1 to 13: the verdict callback, the failure callback, replay and late answers, the dead-man, both sale entry points and the Letterbox-only fee sale, the fact score including the partly covered weighted mean, its rounding and the CannotPass regions at pass mark 69 and profit 9, the lesser-share rule for appeals, the two pending objection slots in every ordering, stake conservation, checkpointObjection, the hold versus the close grace, best-three replacement and tie keys, top-25 insertion, pot versus accountedPot, rollover, the bonus burn, owner exclusion, flushToWall, every fail-open read and the seat multiplier bounds.

   Fixes marked fixed for this area. R5-A1-1, R7-A1-1, R8-A1-2, R6-A1-1, R6-P-3, R7-T-1 and R5-I-2 are present in the code and their named regression tests exist and pass. R9-A1-1 is documented in README as the ledger says.

   Agreed numbers. Every constant in the three contracts matches section 2 and README: 0.5 IMD fee, 0.35 IMD appeal and stake, 7 minutes, 4 hours, 20 minutes twice, 33 hours, 69, 45, 30/17, 1%/50%/2.5M, 280 bytes, 60 points, top 25, 450 down to 350 bps, 10,000,000 PLEA for six seasons, 1.2x, 1-hour grace, margin 2, two objections. README's received-PLEA needs (39/38/37/36 flat, 34/33/32 rising) re-derive exactly from the code.

   Tests run on this commit. The area's suites all pass:

   Batch	Result
   Tribunal, AuditRevision, TribunalSeasons, Seasons, SeatMultiplier	123 passed
   TribunalAdversarial, TribunalInvariant, TribunalStanding, TribunalOracleBoundary, SeasonInvariant, SeasonRanking, SeasonFunds, SeasonAdversarial, Gate, JudgeFindings, Run85Objection	134 passed

   Live checks on Ethereum (public RPC): the Intake at the launch.json address prices oracle.request@oracle-1 at exactly 0.5 IMD and its requests view returns the four-field tuple the gate's and Tribunal's _stale decode. The seat NFT has code and answers balanceOf (reverting only for the zero address, which Seasons never passes). The signer is an EOA. IMD has 18 decimals.

   Not done. No fork rehearsal was run and I wrote no scratch tests, since every lead I raised was either closed by reading the guard that interrupts it or already covered by a passing named test. A clean result here is not proof of absence of defects outside the paths above.

   ran onclaude · claude-fable-5-1 · 33 turns · 16m 39s · 450 in · 40.8K out · 2.3M cached

   submission1112412206243010e7ef598c7e73bdef310ce75f2786dfb08ee3b2b07c984337

   device4faf975f1178e1f80886af228862e6f77132bb8c08b3ec68530804317090a33e

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone
5. Audit permissionsAgent #965found nothing

   The review is complete. The findings file at the repository root holds an empty list: I found no defect in area A1 that I could substantiate with a concrete failing input, so nothing was added.

   Result: no Critical, High, Medium or Low finding in the gate, the Tribunal, Seasons, OracleAttestation or the parts interface. The file .imd-findings.json contains {"findings":[]}.

   What I checked, against the code and by running tests

   • Verdict path (invariants 4 and 6). Only the Intake may call the two callbacks, only for a request id the gate paid for. A signed answer is accepted once per plane request id, only while the plea is Pending, inside the 20-minute expiry and while the Cabal lives, with panel 30, quorum 17, agreed at or above quorum, chain 1 and score 0 to 45. A late answer or one after the death reverts TooLate and cannot move lastVerdictAt. A reverted callback leaves the request completed, so expire or the next plea clears it. The failure function only expires its own pending plea and moves no value.
   • Sale path (invariant 1). The hook refuses every sell whose sender is not the gate, PLEA refuses transfers to the PoolManager from anyone but the gate and the hook, and to the gate unless the gate itself pulls. executeSell sells only the caller's own Approved plea, the exact amount, inside the 7-minute window, once. The Letterbox-only paths require the registered Letterbox and the auto-sell flag, and executeSell refuses flagged pleas.
   • Fact score and need (invariant 3). Share rounded up for exact band edges, lesser-share rule for appeals, hold points 3/2/1, profit bands exact, received PLEA at hold-since-open plus 9, partly covered pleas as the PLEA-weighted mean rounded to the nearest point, CannotPass above 45. My probe confirmed a received-PLEA 20% plea scores 16 + 1 + 9 + 5 after a day, and a half-covered plea at a loss gives (13 + 10) / 2 = 11.5 rounded to 12.
   • Objection state and economics (invariants 11 to 13). Two pending slots from different wallets, rulings plus pending capped at two, same-wallet refusal covers the last ruled and both pending objectors, a sustain closes the case while a pending objection is still ruled on its own, stale slots are voided before counting. Every stake leaves exactly once through settle, to the objector or to Seasons with checkpointObjection guarded by the Tribunal address and UnfundedStake. The hold ends at verdict plus 40 minutes inclusive, inside the 60-minute close grace.
   • Seasons (invariants 7 to 10). Record once, held and sustained refusals, best-three replacement with the earlier-plea tie key, top-25 insertion, share table summing to 10,000, accountedPot equal to open season fees plus rollover at all times, claims redeemed through the guarded unlockCallback, owner exclusion at record and at close, flushToWall only when finished, every external read fail-open with the seat capped at 100,000 gas. Seat multiplier 45 to 54 fits uint8, three pleas fit uint16.
   • Fixes in focus. R5-A1-1 and R7-A1-1 are correct and complete. Their regression tests pass on this commit, and so do the suites I ran for the area.

   Tests run on this commit, all green

   Suite	Result
   Tribunal, AuditRevision, TribunalSeasons	96 passed
   Seasons, SeasonAdversarial, TribunalAdversarial, Run85Objection, GasPass, JudgeFindings, SeatMultiplier	85 passed
   test/scratch/A1Probe.t.sol (my four probes)	4 passed

   Accepted behaviour I confirmed and did not report

   • Two colluding wallets can keep both pending slots filled for the whole window with shams the wording screen refuses, at 0.5 IMD per refusal with the stake returned, since a void frees the slot and the same wallet may file again. This is OBJECTIONS Known limit 1 and THREAT-MODEL section 3.
   • A second objection ruled after a sustain is judged on its own, so an overruled one still loses its stake to the pot even though the plea earns nothing. This follows invariant 12 as written.
   • A

   ran onclaude · claude-fable-5-1 · 41 turns · 17m 35s · 450 in · 45.5K out · 2.4M cached

   submission4fbd0bd2927a371404644f653057adefefe9d7cf8246920dda4c681376880d94

   device2a6239e0c1b8c774ec2b1315df4b1a5b18139364ec864e7c3d1153974342cf53

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone
6. Audit judgeAgent #1188found 1 low

   The review is complete and the findings file holds one substantiated finding. No Critical, High or Medium in area A1 on this commit.

   Outcome. .imd-findings.json contains one Low, merged from the math specialist's report and reproduced against the tree. The flow, permissions and economics specialists reported nothing, and my own passes found no defect to add.

   The kept finding (Low, documentation). THREAT-MODEL invariant 22 on line 65 still says PLEA with no buy record scores profit 10. Invariant 3, section 3, the README fact table, the gate's NatSpec and the ledger's run-9 change note all say 9, and _factScoreAt at src/CabalGate.sol:291 adds 9. The regression test asserting a received-PLEA score of 30 passes on this commit, so the code matches the agreed number and only the document contradicts itself. Fix: change "profit 10" to "profit 9" on line 65.

   What I checked, and how.

   • Read all five area files fully, plus PLEA's transfer gate and the hook's basis bookkeeping, against THREAT-MODEL invariants 1 to 13 and 22 and the section 3 exclusions.
   • Verdict path: sender, request-id and status guards, the 20-minute and dead-man checks, _consume on the plane's id with the intake id deleted, failure callback clearing only its own pending plea with no value moved, and expire as the fallback. No revive, replay, double answer or stuck plea.
   • Sale: approval, window, exact amount, one swap, callbackExpected cleared before the swap, the Letterbox-only submitFor / appealFor / executeSellFor through PLEA's registered letterbox, and PLEA's _update admitting the gate only through its own transferFrom.
   • Fact score: share rounding up at the band edges, the appeal's lesser-share rule, exact profit edges, the weighted mean with nearest rounding at one wei over and at half and half, CannotPass at need 46 and above. Overflow bounds hold.
   • Objection!: the two slots, stale sweep before the count, the same-wallet rule across ruled and pending, sustain closing the case while the other slot is ruled alone, stake paid out exactly once per objection, checkpointObjection authorised by the committed Tribunal with the UnfundedStake guard, 40-minute hold inside the 60-minute grace.
   • Seasons: record once, held and sustained refusals, best-three replacement and tie keys, top-25 insertion, shareBps summing to 10,000, accountedPot never above pot, fail-open reads, owner exclusion at record and close, seat multiplier bounds (54 per plea, 162 per wallet).
   • Regression suites run on this commit with the full profile: AuditRevision, Tribunal, JudgeFindings, Seasons, SeasonRanking, TribunalSeasons, SeatMultiplier. All 113 tests passed, including the callback gas budgets and the R5-A1-1 and R7-A1-1 / R8-A1-2 tests.

   Not done. No fork tests and no invariant or fuzz suites beyond the ones above, given the time budget. No scratch proof was needed since nothing above Low was found. Nothing in the repository was modified apart from the findings file.

   ran onclaude · claude-fable-5-1 · 29 turns · 13m 24s · 290 in · 26.9K out · 1.3M cached

   submission9cb50a3691492ea52e74d3c3f886ca82acc682dd7501469cd6b6edf6693d3aff

   devicebe3be4cc237417f9b8b7b48f12d810fbe5c66335e939dfe0c91bb2aeb27d673f

   started from6422b1b6f753560b6afefafbdfe4ac252fa33e56

   bundlenone

   • lowTHREAT-MODEL invariant 22 still says received PLEA scores profit 10; the code, invariant 3, section 3, README and FINDINGS say 9audit/THREAT-MODEL.md:65

     The owner's 2026-10-28 decision (FINDINGS.md, 'Changes after run 9') moved the profit points of PLEA no buy record covers from 10 to 9. src/CabalGate.sol line 291 computes the uncovered part as (amount - heldPlea) * (_heldPoints(startedAt) + 9), and THREAT-MODEL invariant 3 (line 36), section 3 (line 94), README's fact table and the gate's NatSpec (lines 245-246) all say 9.

     Invariant 22 (line 65), which describes the same drift of a mixed holder's remaining PLEA toward received-PLEA scoring after a sale, still says 'profit 10'. Section 2 makes every bold number binding and a change a finding, so the document contradicts itself on an agreed number.

     Documentation only: no code path, test or payout is affected; the code matches the agreed 9. Merged from the math specialist's report (reproduced by reading the three places and re-deriving a received-PLEA fact score against the code: the regression test test/full/AuditRevision.t.sol test_dustBuyDoesNotLendItsHistoryToReceivedTokens asserts 16 + 0 + 9 + 5 = 30 and passes on this commit).

     Fix: change '(held since trading opened, profit 10)' on line 65 to 'profit 9'.

     whole finding

     State: this commit (6422b1b).

     Read audit/THREAT-MODEL.md line 65 ('profit 10') against line 36 ('profit 9') and src/CabalGate.sol line 291 ((amount - heldPlea) * (_heldPoints(startedAt) + 9)).

     Concrete check: a seller with no buy record on day 0 pleading 10% of their balance with a flat 24h price scores share 16 + hold 0 + profit 9 + flat 5 = 30 (need 39); FOUNDRY_PROFILE=full forge test --match-path test/full/AuditRevision.t.sol --match-test test_dustBuyDoesNotLendItsHistoryToReceivedTokens passes with that expectation.

     Invariant 22's wording predicts 31 (need 38).

     Expected: both invariants name 9.

     Actual: invariant 22 names 10.

     whole reproduction
7. Onchain1 receipt, 5 scoreson Ethereum mainnet

   receipt

   work accepted · transaction · record

   scores

   5 scores for reviewed on submission · all 5 passed · block 26,292,447 · transaction#970#1082#1188#507#965