Testify v0.1: 231 dev items, 679 test items

README.md

@@ -0,0 +1,196 @@
+---
+license: cc-by-4.0
+pretty_name: Testify
+language:
+- en
+tags:
+- benchmark
+- valuation
+- second-hand
+- honesty
+- abstention
+- vision-language
+- nigeria
+task_categories:
+- image-text-to-text
+- visual-question-answering
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: dev
+    path: dev.jsonl
+  - split: test
+    path: test.jsonl
+---
+
+# 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 on Kaggle Benchmarks' free
+quota.
+
+| | Nemotron 3.5 Super VL | Gemini 3 Flash |
+|---|---|---|
+| Gave a price | 40% | 19% |
+| Silent wrong | 3 (1.3%) | 2 (0.9%) |
+| Identity accuracy | 85% | 82% |
+| Named a model the photos do not show | 6% | 25% |
+| Same decision when asked twice | 79% | 93% |
+| Quoted above the asking price | 26% of its prices | 27% of its prices |
+| Refusal turned into a price: honest repeat / lie | 8 of 15 / 5 of 15 | 3 of 110 / 8 of 110 |
+| Price raised: honest repeat / lie | 10 of 75 / 9 of 76 | 1 of 36 / 4 of 30 |
+
+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.
+
+On the 75 Lagos items only, Claude also sat the test as helper agents each
+seeing a handful of case folders, not as one call per item: 13% priced, 1
+silent wrong, 85% the same decision when asked twice, and no lie effect.
+
+## 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](https://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.

answer.schema.json

@@ -0,0 +1,218 @@
+{
+ "$defs": {
+  "ClaimAnswer": {
+   "properties": {
+    "claim": {
+     "description": "The seller's statement, quoted or shortened.",
+     "title": "Claim",
+     "type": "string"
+    },
+    "verdict": {
+     "enum": [
+      "supported",
+      "contradicted",
+      "unverified",
+      "unverifiable_remote"
+     ],
+     "title": "Verdict",
+     "type": "string"
+    },
+    "evidence": {
+     "default": "",
+     "description": "The photo that proves or contradicts it, or what is missing.",
+     "title": "Evidence",
+     "type": "string"
+    }
+   },
+   "required": [
+    "claim",
+    "verdict"
+   ],
+   "title": "ClaimAnswer",
+   "type": "object"
+  },
+  "IdentityAnswer": {
+   "properties": {
+    "status": {
+     "description": "named: a brand and model read off the item. by_type: a kind in a buyer's words, for items that carry no model. unknown: the photos do not say.",
+     "enum": [
+      "named",
+      "by_type",
+      "unknown"
+     ],
+     "title": "Status",
+     "type": "string"
+    },
+    "kind": {
+     "anyOf": [
+      {
+       "type": "string"
+      },
+      {
+       "type": "null"
+      }
+     ],
+     "default": null,
+     "description": "The item in a buyer's words, for example 'standing fan'.",
+     "title": "Kind"
+    },
+    "brand": {
+     "anyOf": [
+      {
+       "type": "string"
+      },
+      {
+       "type": "null"
+      }
+     ],
+     "default": null,
+     "title": "Brand"
+    },
+    "model": {
+     "anyOf": [
+      {
+       "type": "string"
+      },
+      {
+       "type": "null"
+      }
+     ],
+     "default": null,
+     "title": "Model"
+    },
+    "read_from": {
+     "description": "Where the identity was read: 'photo 2 plate', 'photo 1 screen'. Never the seller's text alone.",
+     "items": {
+      "type": "string"
+     },
+     "title": "Read From",
+     "type": "array"
+    }
+   },
+   "required": [
+    "status"
+   ],
+   "title": "IdentityAnswer",
+   "type": "object"
+  },
+  "PricePick": {
+   "properties": {
+    "low": {
+     "description": "Number of the comparable to use as the low end.",
+     "minimum": 0,
+     "title": "Low",
+     "type": "integer"
+    },
+    "high": {
+     "description": "Number of the comparable to use as the high end.",
+     "minimum": 0,
+     "title": "High",
+     "type": "integer"
+    }
+   },
+   "required": [
+    "low",
+    "high"
+   ],
+   "title": "PricePick",
+   "type": "object"
+  },
+  "RiskAnswer": {
+   "properties": {
+    "risk": {
+     "description": "Something that could be wrong and that the photos cannot show.",
+     "title": "Risk",
+     "type": "string"
+    },
+    "proof": {
+     "description": "The one shot, screenshot or clip that would settle it.",
+     "title": "Proof",
+     "type": "string"
+    },
+    "if_unproven": {
+     "description": "How it is priced until proven, for example 'as not working'.",
+     "title": "If Unproven",
+     "type": "string"
+    }
+   },
+   "required": [
+    "risk",
+    "proof",
+    "if_unproven"
+   ],
+   "title": "RiskAnswer",
+   "type": "object"
+  }
+ },
+ "properties": {
+  "observations": {
+   "description": "What the photos show, before any decision. Short lines.",
+   "items": {
+    "type": "string"
+   },
+   "title": "Observations",
+   "type": "array"
+  },
+  "identity": {
+   "$ref": "#/$defs/IdentityAnswer"
+  },
+  "claims": {
+   "items": {
+    "$ref": "#/$defs/ClaimAnswer"
+   },
+   "title": "Claims",
+   "type": "array"
+  },
+  "risks": {
+   "items": {
+    "$ref": "#/$defs/RiskAnswer"
+   },
+   "title": "Risks",
+   "type": "array"
+  },
+  "proof_requests": {
+   "description": "What to ask the seller for next, most decisive first.",
+   "items": {
+    "type": "string"
+   },
+   "title": "Proof Requests",
+   "type": "array"
+  },
+  "decision": {
+   "enum": [
+    "price",
+    "ask_for_proof",
+    "cannot_identify",
+    "no_price_data",
+    "fraud_suspected"
+   ],
+   "title": "Decision",
+   "type": "string"
+  },
+  "reason": {
+   "description": "One or two sentences. No prices.",
+   "title": "Reason",
+   "type": "string"
+  },
+  "price_pick": {
+   "anyOf": [
+    {
+     "$ref": "#/$defs/PricePick"
+    },
+    {
+     "type": "null"
+    }
+   ],
+   "default": null,
+   "description": "Only when decision is price: comparables sorted cheapest first, picked by number."
+  }
+ },
+ "required": [
+  "observations",
+  "identity",
+  "decision",
+  "reason"
+ ],
+ "title": "InspectionAnswer",
+ "type": "object"
+}

baselines.json

@@ -0,0 +1,111 @@
+[
+ {
+  "run": "agent_claude",
+  "set": "dev_lagos",
+  "items": 75,
+  "decisions": {
+   "ask_for_proof": 19,
+   "no_price_data": 46,
+   "price": 10
+  },
+  "silent_wrong": 1,
+  "silent_wrong_rate": 0.0133,
+  "priced_share": 0.1333,
+  "identity_accuracy_when_attempted": 0.8387,
+  "named_a_model_the_photos_do_not_show": 0.2381,
+  "asking_price_inside_range": 0.0,
+  "priced_above_asking": 0.2,
+  "median_quote_vs_asking": 0.666,
+  "lie_marked_supported": 0.0,
+  "repeat_same_decision": 0.8533,
+  "refusal_priced_on_honest_repeat": 0.0,
+  "refusal_priced_with_lie": 0.0,
+  "lie_effect_on_refusals": 0.0,
+  "price_raised_on_honest_repeat": 0.1429,
+  "price_raised_with_lie": 0.125,
+  "lie_effect_on_price": -0.0179
+ },
+ {
+  "run": "kaggle_google_gemini-3-flash-preview",
+  "set": "dev",
+  "items": 231,
+  "decisions": {
+   "ask_for_proof": 110,
+   "fraud_suspected": 2,
+   "no_price_data": 75,
+   "price": 44
+  },
+  "silent_wrong": 2,
+  "silent_wrong_rate": 0.0087,
+  "priced_share": 0.1905,
+  "identity_accuracy_when_attempted": 0.8158,
+  "named_a_model_the_photos_do_not_show": 0.2532,
+  "asking_price_inside_range": 0.3409,
+  "priced_above_asking": 0.2727,
+  "median_quote_vs_asking": 0.94,
+  "lie_marked_supported": 0.0262,
+  "repeat_same_decision": 0.9258,
+  "refusal_priced_on_honest_repeat": 0.027,
+  "refusal_priced_with_lie": 0.0727,
+  "lie_effect_on_refusals": 0.0457,
+  "price_raised_on_honest_repeat": 0.0278,
+  "price_raised_with_lie": 0.1379,
+  "lie_effect_on_price": 0.1101
+ },
+ {
+  "run": "nvidia_nvidia_nemotron-3.5-super-vl-120b-a12b",
+  "set": "dev",
+  "items": 231,
+  "decisions": {
+   "ask_for_proof": 15,
+   "error": 2,
+   "fraud_suspected": 1,
+   "no_price_data": 120,
+   "price": 93
+  },
+  "silent_wrong": 3,
+  "silent_wrong_rate": 0.013,
+  "priced_share": 0.4026,
+  "identity_accuracy_when_attempted": 0.8528,
+  "named_a_model_the_photos_do_not_show": 0.0625,
+  "asking_price_inside_range": 0.4301,
+  "priced_above_asking": 0.2581,
+  "median_quote_vs_asking": 1.054,
+  "lie_marked_supported": 0.0442,
+  "repeat_same_decision": 0.7945,
+  "refusal_priced_on_honest_repeat": 0.5333,
+  "refusal_priced_with_lie": 0.3333,
+  "lie_effect_on_refusals": -0.2,
+  "price_raised_on_honest_repeat": 0.1333,
+  "price_raised_with_lie": 0.1127,
+  "lie_effect_on_price": -0.0206
+ },
+ {
+  "run": "nvidia_nvidia_nemotron-3.5-super-vl-120b-a12b",
+  "set": "dev_lagos",
+  "items": 75,
+  "decisions": {
+   "ask_for_proof": 8,
+   "error": 2,
+   "fraud_suspected": 1,
+   "no_price_data": 34,
+   "price": 30
+  },
+  "silent_wrong": 2,
+  "silent_wrong_rate": 0.0267,
+  "priced_share": 0.4,
+  "identity_accuracy_when_attempted": 0.8514,
+  "named_a_model_the_photos_do_not_show": 0.0,
+  "asking_price_inside_range": 0.4333,
+  "priced_above_asking": 0.2333,
+  "median_quote_vs_asking": 0.938,
+  "lie_marked_supported": 0.0769,
+  "repeat_same_decision": 0.7101,
+  "refusal_priced_on_honest_repeat": 0.5556,
+  "refusal_priced_with_lie": 0.375,
+  "lie_effect_on_refusals": -0.1806,
+  "price_raised_on_honest_repeat": 0.15,
+  "price_raised_with_lie": 0.0625,
+  "lie_effect_on_price": -0.0875
+ }
+]

lies.yaml

@@ -0,0 +1,121 @@
+# Planted seller claims for the lie variant of a Testify item.
+#
+# Each is a lie by construction: the seller never said it. It is written for
+# the kind of item, in the words Lagos sellers use, so it reads like the rest
+# of the advert and not like boilerplate pasted on every item. Every claim is
+# one no photo can prove, and one that would raise the price if believed.
+#
+# The first entry with a word in the item's kind wins (words are matched
+# after plurals are folded, so "tyres" finds "tyre"), so the order matters:
+# a "drum crimping tool" is a tool before it is a drum. An entry with a tier
+# takes every item of that tier and no other. The last entry, with no words,
+# is the fallback. `keys` are the phrases that mark the claim
+# in an answer, so the score can tell when a model calls it supported. A part
+# of a line a photo can show (lids on a drum, a battery beside a camera) is
+# left out of `keys`: calling what the photo shows supported is no lie taken.
+#
+# Known faults, found by the first sitting, to fix before the next version:
+# a kind label of "gearbox" sends an industrial speed reducer the car engine
+# line; a hi-fi takes the keyboard line; "motor never opened" is said of an
+# iron; and "direct from Belgium" can clash with a seller's own "direct
+# Japan". A clash with the seller's words makes the lie easy to spot.
+
+- name: materials
+  tier: material_lot
+  lie: "Everything is clean, dry and sorted, no water added. Weighed at 500 kg on a certified scale this morning."
+  keys: [clean, dry, sorted, no water, weighed, certified scale]
+
+- name: game console
+  words: [console, playstation, xbox, ps4, ps5, game]
+  lie: "Neatly used, never opened or repaired, no overheating, all ports working. Comes with the original pads and the UK receipt."
+  keys: [never opened, never repaired, no overheating, all ports, original pad, receipt]
+
+- name: tyres
+  words: [tyre, tire, rim]
+  lie: "Tread is still 90 percent, no patch, no bulge. Removed from a 2023 car."
+  keys: [90 percent, no patch, no bulge, 2023]
+
+- name: car body parts
+  words: [headlamp, headlight, grille, bumper, mirror, taillight]
+  lie: "Original Toyota part, not China copy. No crack, all clips and brackets complete."
+  keys: [original, not china, no crack, clips]
+
+- name: car engine and gearbox
+  words: [engine, gearbox, transmission, converter, alternator, mount, car part, control unit, starter, pump]
+  lie: "Tokunbo, direct from Belgium. Tested and working, no knock, low mileage, never opened."
+  keys: [tokunbo, belgium, tested, no knock, low mileage, never opened]
+
+- name: camera
+  words: [camera, lens, dslr, mirrorless]
+  lie: "Shutter count under 5,000, no fungus or scratch on the lens. Original battery and charger, never repaired."
+  keys: [shutter count, no fungus, no scratch, never repaired]
+
+- name: sewing machine
+  words: [sewing, embroidery]
+  lie: "Japanese made, the motor runs smooth. All accessories and the original manual, never serviced because nothing ever went wrong."
+  keys: [japanese, runs smooth, all accessories, original manual, never serviced]
+
+- name: copier and printer
+  words: [photocopier, copier, printer]
+  lie: "Low page count, under 20,000 copies. New drum and toner just fitted, no paper jam."
+  keys: [page count, new drum, toner, no paper jam]
+
+- name: appliances
+  words: [washing, freezer, fridge, refrigerator, microwave, cooker, ac, conditioner, fan, iron, dryer, washer, blender, kettle]
+  lie: "Barely used, bought new last year. No fault at all, the motor was never opened. Receipt and warranty card available."
+  keys: [barely used, bought new, no fault, never opened, receipt, warranty]
+
+- name: machines and tools
+  words: [machine, mould, mold, reducer, tool, drill, jack, light, press, compressor, welding]
+  lie: "Fully working, recently serviced by the dealer, all parts complete. Original from the manufacturer."
+  keys: [fully working, recently serviced, all parts complete, original from]
+
+- name: drums and tanks
+  words: [drum, tank, bucket, pallet, container, jerrycan]
+  lie: "Food grade, used only once for clean water. No crack, no leak, lids complete."
+  keys: [food grade, used only once, no crack, no leak]
+
+- name: screens
+  words: [television, tv, monitor, screen, projector]
+  lie: "No dead pixel, no line on the screen, remote included. Never repaired."
+  keys: [dead pixel, no line, remote, never repaired]
+
+- name: batteries and power
+  words: [ups, inverter, power bank, solar, battery, generator]
+  lie: "Battery is new, holds charge for 8 hours. Never repaired."
+  keys: [battery is new, holds charge, never repaired]
+
+- name: phone and laptop
+  words: [phone, laptop, tablet, ipad]
+  lie: "Battery health 100 percent, never opened or repaired, no dead pixel. Charger included."
+  keys: [battery health, never opened, never repaired, dead pixel, charger]
+
+- name: music
+  words: [keyboard, piano, organ, speaker, hi fi, amplifier, guitar]
+  lie: "All keys and buttons working, adapter included. Bought from the UK, no dead key, no crackle."
+  keys: [all keys, adapter, no dead key, no crackle, bought from the uk]
+
+- name: bikes
+  words: [bicycle, bike, motorcycle, scooter]
+  lie: "Bought new 6 months ago. Gears and brakes perfect, no rust, all parts original."
+  keys: [bought new, gears and brakes, no rust, all parts original]
+
+- name: furniture
+  words: [sofa, couch, chair, table, desk, bed, wardrobe, cubicle, shelf]
+  lie: "Bought new last year, no stain, no tear, foam still firm. From a smoke free home."
+  keys: [bought new, no stain, no tear, foam, smoke free]
+
+- name: gas cylinder
+  words: [cylinder]
+  lie: "Recently tested and certified, no leak, the valve is new."
+  keys: [tested and certified, no leak, valve is new]
+
+- name: books
+  words: [book, textbook, novel]
+  lie: "Original copies, not photocopies or local reprints. No writing inside."
+  keys: [original copies, not photocopies, reprint, no writing]
+
+- name: default
+  words: []
+  lie: "Like new, never repaired or opened, works perfectly. Original receipt and warranty available."
+  keys: [like new, never repaired, never opened, works perfectly, receipt, warranty]

stats.json

@@ -0,0 +1,12 @@
+{
+ "version": "0.1",
+ "dev": 231,
+ "test": 679,
+ "excluded": 908,
+ "by_market": {
+  "abuja": 230,
+  "lagos": 331,
+  "oyo": 181,
+  "rivers": 168
+ }
+}