AI, data & research teams
Ground truth for physical-world AI, on demand.
Name a place, a question and your acceptance rule. Our real-world incentivised network goes and looks, AI does routine checks and people review what is unclear, and you receive structured, consented records with photo, GPS and time.
What it does
Built for one clear job.
Models learned the world from the web, and streets, shelves and billboards barely appear there. We close that gap on demand: people in our network are rewarded for answering specified questions at real places, and every answer carries a photo, GPS, time and consent, and is checked before it counts.
Start with a small paid sample against your own acceptance rule. Records arrive as a report, dataset or API-ready file, and the network can return to the same places on a cadence.
How it works
From signal to defensible decision.
Specify the gap
Name the places, the question and the cadence.
Set acceptance
Agree the rule a record must pass before it counts.
Collect & check
Rewarded people capture each answer; AI does routine checks and people review what is unclear.
Deliver the sample
Receive consented records as a report, dataset or API-ready file.
Core capabilities
What you get.
Specified on demand
Your places, your question, your cadence: not a fixed catalogue.
Fresh place observations
Check a specified physical-world claim against new observations rather than treating a model prediction as evidence.
Place & site profiles
What makes a place commercially useful, reusable across decisions.
Provenance and consent
Photo, GPS, time and opt-in consent on every record, pseudonymous by default.
Why timbuk2
Most data describes the world as it was when someone last mapped it. Ours is asked for: a real-world incentivised network sends people to the places you name, every answer is checked before it counts, and each record carries the photo, place, time and consent you need to defend it.
Who it's for
For AI, data and research teams, and the brands they serve
Teams building spatial products, world models or market intelligence who need fresh, specified evidence from places the web barely covers, with provenance and consent they can defend.
Outcomes
What changes for you.
- Fresh evidence from places the web barely covers
- Every record checked against your acceptance rule
- Delivered as a report, dataset or API-ready record
Before commissioning collection
Define ground truth before you collect it.
Illustrative acceptance brief: a spatial-data team has candidate place records and wants to check a specified visible attribute. The team defines the claim, evidence and freshness window before a field sample is commissioned. This is a worked scoping example, not a customer result.
Separate the four jobs
Collection records an observation. Label definition states what the answer means. Quality review checks the record against the rule. Independent validation tests reliability separately. A photograph with GPS does not complete all four jobs by itself.
Specify the test
For each candidate place, write the exact claim being checked and what evidence would support or contradict it. Define the location match, observation date window, image requirements and permitted answer labels. Avoid asking the collector to confirm an expected answer.
Preserve uncertainty
An illustrative review could classify a record as accepted, uncertain or rejected against the agreed rule. Explain why a record fails or remains unclear. A conflicting image, ambiguous location or inaccessible site should not be converted into a confident training label.
Agree the intended use
Place verification, physical-world attribute evaluation and fresh observations have different rights and quality requirements. Discuss the intended training or evaluation use, consent, retention and permitted sharing before licensing a dataset. Training rights are not automatic.
Illustrative requested output fields
Use this as a scoping checklist. Fields, review labels and delivery format must be agreed for your project.
- Candidate record / question
- Which place claim the observation is intended to test, with the label definition agreed in advance.
- Answer / time / location
- The observed response, its freshness and its location context rather than an unsupported model inference.
- Evidence reference
- A reference to the supporting evidence under the project's agreed access and permitted-use terms.
- Review result / reason
- Whether the acceptance rule was met and why an observation is uncertain or rejected. Independent validation remains a separate step.
This is an illustrative requested schema, not a production dataset or an assurance of accuracy. Confirm coverage, fields, rights, cadence and validation design for the sample. Do not treat approved submissions as independent ground truth without the validation your use requires.
Before you start.
What the evidence can tell you, what it cannot, and how to scope the work.
What is ground-truth data collection?
It means checking a question at the real place rather than relying only on an existing map or model. Timbuk2 collects specified observations with photos, location and time, then checks each record against an agreed acceptance rule.
What does an AI or research team receive?
Start with a scoped paid sample: your places, question, cadence and acceptance rule. Accepted records can be delivered as a report, dataset or API-ready file. Coverage, fields and delivery timing are agreed for the project; this is not a claim of complete coverage everywhere.
Can the data be used to train a model?
Model-training rights are not automatic. Consent, permitted use and acceptance terms for that use must be agreed before anything is licensed for training. Participants are pseudonymous and the collection uses opt-in consent.
See it against your own market.
Tell us the geography, question and decision that matter to you. We'll confirm the scope and what evidence would answer it.
Scope a sampleKeep exploring