ERC-8376 Token launch fraud detection and refund standard (Token Launch Abuse Detection and Remediation)

Review · Vaults & DeFi · created 2026-08-07

A standard that scores and discloses deployer misconduct, such as rug pulls, using on-chain data, and requires exchanges to hold sale proceeds in escrow temporarily so that once fraud is confirmed, buyers receive a pro-rata refund.

Why it is needed

Token issuers are anonymous and disappear the moment they grab the money. Exchanges, by contrast, are identifiable, keep operating, and are the only ones holding records of who bought how much. Yet three things have been missing until now. First, whether a deployer locked liquidity, retains minting authority, or has already withdrawn the raised funds can be seen on-chain, but it is not disclosed in a common format buyers can compare. Second, because exchanges hand purchase proceeds straight to the issuer in the same transaction, the money disappears before fraud even comes to light. Third, victims have no real recourse against an anonymous issuer beyond social condemnation. Losses are scattered across many small amounts, making individual lawsuits a losing proposition, but launch fraud leaves more room for recovery than other incident types because the money is still sitting with the exchange.

What to implement

In depth

Three conformance levels

The specification defines three conformance levels: Consumer, Detection, and Full. The Consumer level requires no separate contract deployment and only needs to call checkLaunch to disclose abuseScore, confidence, and evidenceURI at the time of purchase; when there is no report, it must display 'unknown' and must not display 'safe'. This level cannot claim to offer refunds or remediation. The Detection level adds ILaunchDirectory and ILaunchAbuseRegistry, registering launches it lists in the directory, but still does not hold funds and offers no remediation. The Full level must also implement ILaunchEscrow and ILaunchRemediation, making it the only level at which buyers can actually receive a refund. A venue cannot claim a level it does not actually meet, and terms like 'verified' or 'protected' may only be used when remediation actually follows.

11 abuse patterns and keccak256 identifiers

The specification defines eleven deployer abuse patterns, each identified by the keccak256 hash of a namespaced string: PATTERN_HARD_RUG, PATTERN_SOFT_RUG, PATTERN_INSIDER_ALLOCATION, PATTERN_SNIPER_COORDINATION, PATTERN_HONEYPOT, PATTERN_MINT_DILUTION, PATTERN_RETAINED_CONTROL, PATTERN_WASH_LAUNCH, PATTERN_UNLOCK_EXIT, PATTERN_SERIAL_DEPLOYER, and PATTERN_IMPERSONATION. The 'erc.' prefix is reserved for patterns defined by the ERC itself; implementations defining additional patterns must use a different namespace, such as a domain they own. Consumers must treat identifiers they do not recognize as informational only and must not apply automatic blocking actions based on them. PATTERN_HONEYPOT and PATTERN_IMPERSONATION cannot be scored with the baseline 12-signal vector and require a separate extension schema.

The 12 baseline signal vectors and the principle of excluding outcomes

Every detector must compute and disclose, alongside the score, the 12 baseline signals (deployerSupplyShare, insiderAllocationShare, sniperConcentration, lpLockedShare, lpLockRemaining, liquidityRemoved, deployerSellRatio, proceedsWithdrawnShare, privilegedPowers, priorUpheldClaims, supplyInflation, washTradeRatio). Outcomes such as price decline, loss size, or market cap must never be scored by any signal. Because most launches decline even absent abuse, scoring outcomes would classify even normal launches as abuse and destroy the value of the score. A signal whose value cannot be computed must be reported as an explicit unmeasurable signal, represented as the maximum value of uint16 or uint32 rather than zero. The signal vector is versioned, and a report must declare the version under which it was computed.

abuseScore computation and safeguards for categorical signals

abuseScore is a uint8 from 0 to 100, computed as a weighted average in which each signal is normalized from 0 to 1 against an activation threshold and multiplied by a pattern-specific weight. Protective signals are inverted before being added, and unmeasurable signals are excluded from both the numerator and the denominator. A pattern confirmed solely by a categorical signal, such as privilegedPowers, cannot be scored and must be reported as unmeasurable; a pattern whose confirmed signals fall below half of the profile's weight, or one confirmed without any signal describing the underlying conduct, must not be reported above the Elevated band. confidence is a uint8 separate from abuseScore, and automatic blocking actions must not be triggered at low confidence.

Registry authority structure: submission, slashing, relay

In ILaunchAbuseRegistry, each of the three authorities is approved by the party that authority affects. submitReportFor, which files a report under a detector's name, can only be called by the detector itself, a submitter the detector has nominated and who has accepted, or a relay the detector has approved via approveRelay. The authority to forfeit a detector's bond belongs solely to the single detectorAdjudicator designated at registry deployment, which cannot be the venue's remediation contract and cannot be an address that submits reports. A deployer bond forfeited from a rejected claim can only be withdrawn by the remediation contract that received that bond. This structure prevents the party a report accuses from being able to punish the report's author.

Identifying the deployer and the limits of pre-launch reports

The deployer is the address the venue declares via deployerOf(launchId) through ILaunchDirectory, and detectors must use this value rather than inferring it directly from the transaction that created the token. At most venues, a factory creates the token on behalf of the applicant, so inferring the deployer from the creation address would attribute every launch to the venue itself. A report disclosed before a launch has even opened cannot yet exist for 8 of the 12 signals, since those describe conduct during or after the sale, and such signals must be reported as unmeasurable rather than zero. When a launch is not in the directory, or the directory does not record a deployer, every deployer-related signal must likewise be reported as unmeasurable.

Evidence commitments and chain anchoring

The evidenceRoot must be a Merkle root whose leaves can each be verified independently from chain state, and leaves must take the form keccak256(abi.encode(0x00, chainId, blockNumber, txHash, logIndex, fieldId, value)). Leaves and internal nodes must be domain-separated with 0x00 and 0x01 prefixes respectively, and when the number of leaves is odd, the unpaired leaf must be promoted to the next level rather than duplicated. A commitment not anchored to the chain only guarantees internal consistency, not correctness, so anyone must be able to recompute the signal vector from public data to prove a false report, which is what gives a detector's bond real enforcement power.

A Review-stage proposal that may still change

ERC-8376 was written on August 7, 2026 and its status is currently Review. The Review stage means the content is not yet final, and that during community discussion function signatures, signal definitions, or score band thresholds could change, or the proposal itself could stall without further progress.

Original ERC · ← ERCs