Four ways to resolve a freelance dispute, and what each one actually accepts as evidence

Every freelancer knows the sequence. The work is delivered, the client goes quiet for two days, and then: "this isn't what I asked for."

Sometimes they're right. Often they mean "this isn't what I imagined." Either way the money is now stuck between two people who disagree about a document neither of them wrote down properly.

I've been reading the actual mechanics of four dispute systems, because the interesting question isn't who wins. It's what each system treats as evidence, how long you wait for a decision, and — the part I had wrong before I started reading — who carries the risk while you wait. Four systems, four different answers.

1. Upwork fixed-price: a 14-day clock you don't control

The client funds a milestone, you deliver, the client releases. If they go quiet, the important detail is that the clock is already running: delivering through the "Submit Work for Payment" button starts a 14-day countdown to automatic payment. Approving early just ends it sooner. After 14 days the money moves unless the client has actively asked for changes.

If they do dispute it, the fixed-price path runs in stages. There's free dispute assistance (or mediation assistance, depending on whether funds were deposited), then binding arbitration run by an independent third-party company — currently Brief — rather than Upwork itself. You get seven calendar days to agree to arbitration and pay your share of the fee; in most cases the freelancer and client split it evenly, and Upwork contributes only on contracts over $20,000.

The detail that matters most is what happens if one side doesn't engage. If both parties decline arbitration, the funds go to the client. If one party pursues and pays and the other declines, the funds go to whoever participated. Declining isn't neutral — it's a forfeit.

And there's a harder deadline underneath, which the docs tie specifically to the refund-request case: once a client ends a contract with a deposited balance and asks for a refund, you have seven calendar days to review the payment details and file a dispute. Miss it, and the remaining funds go back to the client.

What it accepts as evidence: the brief, the message history, the delivered files, read by a human with context. That is genuinely the right instrument for scope creep, which is a narrative disagreement rather than a factual one.

What it costs: the arbitration fee when you need it, plus a service fee. And the outcome depends on which person reads your file and how well you wrote it up.

2. Fiverr: a 3-day auto-complete, wrapped in a tool that doesn't adjudicate

Fiverr's version is faster and blunter. Requests raised in the Resolution Center go directly to the other party — Customer Support is explicitly not notified during that exchange. Each side has 48 hours to accept or decline, and if nobody responds within 48 hours the system auto-accepts.

Two mechanics matter more than the Resolution Center itself. First, delivery auto-completes the order: if the buyer takes no action within three days of delivery — no acceptance, no revision request, no extension — Fiverr completes the order and starts the clearance clock. Two caveats keep "3 days" from being the whole story: gigs that require physical shipping auto-complete after 14 days rather than 3, and completing an order is not the same as cash in hand — Fiverr holds earnings through a safety clearance period before they can be withdrawn. Fiverr does offer an Early Payout option, but only to Top Rated and Seller Plus Premium freelancers, at a 1% fee, and auto-completed orders qualify only for Top Rated sellers. Second, the Resolution Center is unavailable once an order is complete or cancelled, so that door closes exactly when disputes usually start; for completed orders the documented path is Customer Support. Separately, an order 24 hours past its delivery date can be cancelled by the client without the freelancer's approval.

Fiverr's commission is also reported as a flat fifth of the order with no tiers or volume discount — Fiverr's own payments terms sit behind a bot wall, so I'm relying on consistent secondary reporting there rather than the primary document. Either way, the dispute mechanism isn't the only thing taking a cut.

What it accepts as evidence: very little inside the Resolution Center itself. It's a mutual-agreement tool, not an adjudication tool; its real enforcement mechanism is the cancellation button and the effect a cancellation has on your metrics.

What it costs: cheap and fast when both sides want out. When they don't, the Resolution Center itself can't decide anything — you're escalated to support, which is a human with a queue rather than an independent arbitrator. Its strongest feature is actually the three-day auto-complete, and its weakest is that the tool built for disputes doesn't adjudicate.

3. PayPal: 180 days of hindsight

PayPal inverts the whole problem. There's no milestone and no escrow: the money lands in your account when the client pays, so a client who goes quiet has already paid you. What PayPal adds instead is a second, much longer window in which that payment can be taken back — which is why the risk here isn't ghosting, it's reversal. A buyer can open an "item not received" dispute up to 180 days from the payment date. For "significantly not as described" the window is tighter — 30 days from delivery, or 180 days from payment, whichever comes first — which is actually the more relevant clock for service work, since that's the claim most likely to be filed.

Then there's a second clock almost nobody mentions: an unescalated dispute closes automatically after 20 days. If the buyer does nothing, the dispute evaporates; if they escalate it to a claim, PayPal adjudicates.

Here's where the two things get conflated, including in how most freelancers talk about it. Seller Protection for intangible work does exist, and PayPal's terms spell out the conditions: the transaction must be marked eligible, and you must supply proof of shipment or delivery — for intangible goods, "compelling evidence to show the item was delivered or the purchase order was fulfilled," such as a system of record showing the date and the recipient's address. But Seller Protection is a defence against item-not-received and unauthorised-transaction claims. It is not a defence against "significantly not as described" — and for service work that's the claim you'll actually face. On SNAD the question isn't whether you delivered, it's whether what you delivered matched the description, which is precisely the dispute a record of delivery can't settle.

Which is the practical point: a screenshot of a friendly chat is not that. A chat log is evidence you talked, not evidence you delivered.

What it accepts as evidence: delivery records and the buyer's claim, decided by PayPal under its own programme policies at its discretion. "Significantly not as described" exists as a category precisely because description is subjective — and that's the category a delivery receipt does nothing for.

What it costs: time, and a buyer-weighted default. The upsides are real though — fiat, fast, and a genuine consumer-protection rail that most crypto rails simply don't have.

4. Verdikta: pay against a published rubric, not a mood

A different primitive, and the one I've spent most time reading. Verdikta moves the criteria to before the work: the creator locks ETH in a Base escrow contract and publishes success criteria plus a pass threshold. A jury of AI models on the Verdikta network grades each submission against those criteria and reports a 0–100 score with written justifications.

The properties that actually differ from the three above:

• The criteria, threshold and deadline are fixed at funding. The contract has no cancel and no edit function, so the goalposts cannot move mid-task.
• Passing work is paid without the creator's consent. A qualifying submission settles automatically; the creator cannot simply refuse to release.
• No platform cut. The escrow contract has no owner and takes nothing from the payout or the refund. The creator pays the reward plus gas.
• The verdict is final. As the project's own FAQ puts it: "Can I appeal a verdict? No."
• The hunter pays per attempt. Each evaluation needs a prepay of about 0.00024 ETH at default settings, covering the jury's operators, and it is not returned if the submission fails. At recent ETH prices that's under a dollar, but it's a cost per attempt, not a cost per payout.
• Earliest pass wins, not best pass. Payout follows submission order. Once a submission clears the threshold, a later and possibly better one is recorded as passed-but-unpaid.

The four side by side

Where the Verdikta model genuinely breaks

A comparison that only lists strengths is advertising. Eight real weaknesses:

The rubric is the whole product, and writing one is hard. Most freelance disputes aren't about checkable criteria — they're about scope: "I assumed revisions were included." A criteria-first system has no answer to ambiguity that was never resolved before funding. If the deliverable can't be expressed as testable statements, this model doesn't apply, and pretending otherwise just relocates the argument to the start of the project.

No human appeal. The verdict is final by design. For a function, a dataset or a checklist, fine — correctness is checkable. For brand voice, design direction or long-form writing, a model can misread intent and there's no escalation path. That's a real cost, not a rounding error.

It's an onboarding problem before it's a dispute problem. To take on the work at all, the freelancer needs a crypto wallet, ETH on Base, and enough comfort with a block explorer to confirm that funding actually landed. That excludes a large share of the people who most need dispute protection, and it does it silently — the exclusion doesn't show up as a rejection, it shows up as never applying. Filing a PayPal dispute needs an account you already have; posting to Verdikta needs a wallet you probably don't.

A model jury is an attack surface. The judge reads the submission, and the submission is attacker-controlled text. Anyone who has watched an LLM get talked out of its own instructions knows a rubric is not a sandbox. Publishing the criteria makes this worse, not better: it tells a bad-faith submitter exactly which phrases the scorer rewards. Human arbitration has bias too, but a human is at least awkward to manipulate at scale.

Earliest passing submission wins. Payout follows submission order rather than quality — defensible against queue-jumping by copying, but it rewards whoever reads the rubric fastest.

All-or-nothing. No partial payment for partial work. Traditional escrow does milestones; this does one deliverable and one verdict.

Every submission is public. Verdikta's own FAQ answers "Is my submission public?" with "Yes, treat it as public" — submissions are stored on IPFS with identifiers recorded on-chain, so anyone can read them. That rules out client work under NDA, unreleased code, internal data, anything commercially sensitive. It also means a failed submission stays readable forever, which is a strange property for work you'd rather not have on a permanent public record. Most of the fiat systems here keep a dispute private between the parties and the platform; this one publishes the work by design.

No chargeback in either direction. Great for the freelancer whose client gets buyer's remorse. But the buyer's only lever is rubric design: a criterion rejecting copied or derivative work has to be written in before funding, because once a verdict lands there's no chargeback and no appeal. Consumer protection is replaced by up-front discipline, and the assumption that the creator gets the rubric right on the first attempt is doing a lot of work.

So which one should you actually use

• Short, checkable, single-deliverable work — a function, a dataset, a transcript, a checklist: Verdikta's model wins on speed and cost, provided everyone involved already has a wallet.
• Clients who go quiet — two of these systems already handle it, and don't need a blockchain to do it: Upwork releases after 14 days, Fiverr completes the order 3 days after delivery. If ghosting is your specific fear, the platform you're already on has probably solved it.
• Scope-heavy or taste-driven work — brand, design direction, long-form editorial: human mediation still wins, because the disagreement is about interpretation, not verification.
• High fraud risk in either direction: you still want the fiat rails and their chargeback machinery, flaws included.

The honest summary: this isn't a better Upwork, and the ghosting problem it's usually sold on is already handled by auto-complete clocks on both of the big platforms. What Verdikta actually changes is who holds the judgement — a published rubric and a model jury instead of a support queue — and it trades a human queue for a wallet requirement plus a judge that can be argued with. For narrow, checkable work that's a good trade. For everything else, the older systems are worse in ways that are at least legible.

Genuine question for anyone who has run work through a rubric-based escrow: did writing the criteria up front change what you actually delivered, or only how you described it afterwards?

Mechanics above are from Upwork's Fixed-Price Protection, arbitration and dispute help articles; Fiverr's order-process, Resolution Center and earnings documentation; PayPal's seller protection, dispute-timeframe and escalation pages; and Verdikta's published FAQ documentation, all read this week. The Fiverr commission figure is the one number here I could not verify against Fiverr's own legal pages, so I've marked it as reported rather than quoted. Where a platform's own docs and its marketing disagree, I've used the docs.