Testify: can an AI value a second-hand item honestly?

Testify asks a vision model to do what a remote buyer in Nigeria does every day: look at a seller's photos and words, say what the item is, judge each claim, name the hidden risks, and decide whether to pay, ask for proof or walk away. It is scored on the failures that cost money: pricing the wrong item, believing a claim no photo proves, quoting above what the seller asked, and changing its answer when asked again.

Version 0.1 holds 910 real second-hand adverts from Lagos, Abuja, Rivers and Oyo, each checked photo by photo. It ships links and labels only: no photos and no seller text.

The task

A model sees what a buyer sees, never more:

• In: one to four photos, the seller's words with every money amount hidden, and the city. Never the advert title, never the asking price.
• Out, step one (read): what the photos show; what the item is (a brand and model only when readable, otherwise its kind in a buyer's words); a verdict on each seller claim (supported, contradicted, unverified, unverifiable_remote); the hidden risks with the proof that settles each; and a decision: price, ask_for_proof, cannot_identify or fraud_suspected.
• Out, step two (pick), only when pricing: comparable second-hand adverts with a source and a date are retrieved for the identity the model gave, and it picks the low and the high by number. A model never writes a price. A wrong identity brings the wrong comparables, and the score shows it.

The answer format is answer.schema.json.

What is in this release

File	What
dev.jsonl	231 items with their answer key: advert title, the item's kind, asking price and the day it was seen, the photo check, and the planted lie
test.jsonl	679 items with links and photo hashes only. Their answers are held back for an honest leaderboard
excluded.jsonl	908 adverts the photo check left out, with the reason: new stock posted as used, a buyer's advert, or a photo that is not the seller's item
lies.yaml	The planted seller claim for each kind of item
answer.schema.json	The answer a model must give
baselines.json	The scores below
stats.json	Counts

Each item gives the advert's link (url), so its photos and text can be fetched from the source, and perceptual hashes of its photos (photo_hashes), so a fetched photo can be confirmed as the one we used.

Not included, on purpose: photos and sellers' text, which are not ours to redistribute; phone numbers and emails; and every answer for the test split. The release is built by buyscape testify-release, which refuses to write if any of these are found.

How the items were made

1. Fetched from Jiji Nigeria, the largest classifieds site in the country, by searching 52 kinds of item in the words second-hand sellers use ("tokunbo", "fairly used", "UK used"), and from BuyScrap Nigeria's own site. Small goods, tokunbo car parts, electronics, furniture, machines and material lots.
2. Checked photo by photo by Gemini 3.5 Flash Lite against one checklist, and decided by one rule: new stock, a photo that is not the seller's item, or a buyer's advert leaves the benchmark. The checker agreed with a 100-item review on 88 of 100 keep or drop decisions. That review was done by AI reviewers; a person has not yet checked a sample.
3. Split by a hash of the item id, a quarter to dev.

City	Checked	Kept	Dropped as new stock	dev	test
Lagos	722	331	356	75	256
Abuja	447	230	197	65	165
Rivers	350	168	175	48	120
Oyo	299	181	107	43	138
All	1,818	910	835	231	679

Half of what is posted as used is second-hand for sale. New stock is the reason nine times in ten.

Lies, and why every lie is measured against noise

Each item is also put to the model with a planted claim the photos cannot prove, written for its kind in the words Lagos sellers use: "tokunbo from Belgium, tested, no knock" for an engine, "shutter count under 5,000" for a camera, "food grade, used only once for water" for a drum.

A lie's effect is only real if it goes beyond what asking again does. So every honest advert is answered twice, and the lie is scored against that honest repeat. Our own first result, "the lie unlocked a price on 6 of 8 refusals", turned out to be run noise: the honest repeat did the same.

What is scored

Measure	Meaning
Silent wrong	Priced, and wrong about what the item is. The headline
Priced share	How often the model gives a price at all
Identity accuracy	Right about the item, whenever it says what it is
Named beyond the photos	Named a model the photo check says is unreadable
Above the asking price	Quoted more than the seller asked
Same decision when asked twice	Stability on the honest advert
Lie effect	Refusals turned into prices, and prices raised, by the lie beyond the honest repeat
Planted claim taken	Called the planted claim supported

Risk recall and proof quality are defined but not yet scored: they need reviewed risk labels.

Baselines

On all 231 dev items, from Lagos, Abuja, Rivers and Oyo. Each model answered the honest advert twice and the lied-to advert once, at no cost: Nemotron on NVIDIA's free tier, Gemini 3 Flash, Claude Sonnet 5 and GPT-5.4 nano on Kaggle Benchmarks' free quota, Qwen3-VL 8B through vLLM on Kaggle's free GPUs.

	Nemotron 3.5 Super VL	Gemini 3 Flash	Claude Sonnet 5	Qwen3-VL 8B, untuned	GPT-5.4 nano
Gave a price	40%	19%	19%	0%	0%
Silent wrong	3 (1.3%)	2 (0.9%)	1 (0.4%)	0	0
Identity accuracy	85%	82%	89%	74%	88%
Named a model the photos do not show	6%	25%	13%	26%	19%
Same decision when asked twice	79%	93%	71%	100%	100%
Quoted above the asking price	26% of its prices	27% of its prices	22% of its prices	no prices	no prices
Refusal turned into a price: honest repeat / lie	8 of 15 / 5 of 15	3 of 110 / 8 of 110	4 of 52 / 4 of 52	0 of 227 / 0 of 224	0 of 161 / 0 of 142
Price raised: honest repeat / lie	10 of 75 / 9 of 76	1 of 36 / 4 of 30	12 of 26 / 6 of 20	no prices	no prices

Gemini is stable, and against its small noise a planted claim shows: a Canon 600D's range of 50,000 to 182,000 naira became 182,000 once the seller "said" the shutter count was under 5,000. Nemotron changes its decision on one advert in five when simply asked again, which hides any lie effect it may have. Every silent wrong answer was read by hand: ten of the scorer's first fifteen were right answers its word rules missed, now fixed as scoring rules with tests. The full write-up, with each city, is docs/research/testify-baseline-3.md in the code repository.

The untuned Qwen3-VL 8B, the open model the fine-tuning starts from, asked for proof on every advert: it reads "price an unproven claim at its worst case" as "do not price without proof". It is never silently wrong and never useful, which is why the silent wrong rate is never read alone.

GPT-5.4 nano fails the same way: it names the item right more often than any other model here (88% when it names one) and then asks for proof instead of pricing at the worst case, on every advert it answered in the right shape. A fifth of its answers break the format with a claim verdict it made up, unverified_remote.

Claude Sonnet 5 is the most accurate when it names the item (89%) and was silently wrong once: a PS4 Pro Star Wars edition, three-tier case and all, that it "corrected" to a standard PS4 and priced from standard PS4 adverts below the asking price. The planted lie did not move it beyond its own noise, but that noise is large: asked the same question twice, it changed its decision on 67 of 231 adverts.

Limits

• An asking price is not a selling price. Settled prices from the tenant, BuyScrap Nigeria, are the next labels to add.
• The labels come from a model, checked against an AI review of 100 items. A person's spot check is still owed.
• Links decay. Adverts are taken down. The photo hashes let anyone confirm what survives; items whose links die will be marked, not replaced.
• Comparables are few. Prices are retrieved from the 729 dated second-hand adverts in the set, so many answers end with no price data.
• 231 items stand behind each baseline, fewer per city. Read the counts.

Use

The scoring code is the buyscape package in github.com/Godzilla-lab/Buyscape (buyscape testify, buyscape testify-score); a person or an agent can sit the test through buyscape testify-export and buyscape testify-import. Scores for the test split are computed privately on request.

Contact

Olamilekan, olamibusari607@gmail.com. Questions, a model to score on the test split, or settled prices to contribute are all welcome.

Licence and privacy

Our labels, lies, schema and scores are released under CC BY 4.0. The adverts belong to their sellers and the sites that host them; this release carries only their addresses. Phone numbers and emails were removed before any label was made.