TG GiftsPROOF
Proof, with samples and sources
These numbers come from our own logs and public Fragment data. Every figure shows its sample size (n), date window and source. Where a number cannot be measured honestly, this page says so instead of showing a placeholder.
PROOF
What we can prove, and what we can't
n=5 confirmed sales
51,236 listings · 2026-11-01
3 at the alert price · 2 below
Limits: 17 gift snipes, 5 sold; auctions are separate (their alert price is a minimum bid); alert latency: Collecting: 0/20 sent alerts. Floors are one scan per day.
Method and every number · Updated 2026-11-01 20:04:01 UTC
| Metric | Value | n | Window |
|---|---|---|---|
| Gift snipes: detection → sale | 7.1 h (p25 4.3 h · p75 2 d) | n=5 | 2026-10-19–2026-10-31 |
| Gift snipes vs alert price | 0 above · 3 at · 2 below | n=5 | 2026-10-19–2026-10-31 |
| Floor change, 24 hours | median 0.0% · 8 up · 4 down · 90 flat | n=102 | 2026-10-31 → 2026-11-01 |
| Valuation accuracy (live) | Collecting: 8 / 30 | n=8 | 90 d |
| Coverage | 123 collections · 51,236 listings | 123 / 7d | 2026-10-26 → 2026-11-01 |
| Alert-send latency | Collecting: 0/20 sent alerts | n=0 | — |
| Cross-market spread | Internal — not published | n=102 | 2026-10-31–2026-11-01 |
Hunt types, separately
| Type | Hunts | Sold | At/above alert | Below alert | Median time |
|---|---|---|---|---|---|
| Gift snipe (fixed price) | 17 | 5 | 3 | 2 | 7.1 h (n=5) |
| Gift auction | 0 | 0 | min bid — not comparable | min bid — not comparable | no confirmed sale |
| Username auction | 11 | 11 | min bid — not comparable | min bid — not comparable | 8 min (n=11) |
| Username (fixed price) | 7 | 0 | no confirmed sale | no confirmed sale | no confirmed sale |
An auction's alert price is the minimum bid, not a price paid. Auction sales are therefore never counted as above or below the alert price, and are never pooled with fixed-price gift snipes.
How to read it
Time and outcome: the clock starts when the radar logged the listing (`found`), not when Telegram delivered the alert, because send time is not stored. The sale is confirmed only when Fragment shows a public sale price and time. Fixed-price gift snipes are compared with the price a buyer actually pays; auctions are shown separately because their alert price is a minimum bid, not a price paid.
Floor and coverage: the floor is the cheapest buy-now listing at the first scan of the day, not a sale. Listings are counted from the same daily scan; each model row is capped at 60. Alerts sent per day is not shown because no per-delivery record exists yet.
Accuracy and gaps: the live valuation record publishes only after 30 settled estimates; 8 are settled today. The historical backtest below is separate and never mixed with it. Alert latency is now instrumented at the send path and is published at 20 successful sends (see the table); cross-market spread stays internal until the owner approves the second source.
Known limits
- Small samples: the confirmed-sale base is small, and only a few fixed-price hunts. Read n in the table above.
- One floor scan per day; a flat day is not proof that nothing traded.
- The live valuation record is below its 30-estimate publication threshold.
- Cross-market data is internal and covers a small number of snapshots.
- No resale price and no profit is promised. An alert price is not a resale price.
Historical backtest, not the live record
999 unseen gift sales (26–31 Oct 2026 holdout): median absolute error 16.7%, 58.0% within ±20%, and the published 80% range contained the sale 80.3% of the time. Usernames (90–180 day holdout, n=170): median error 28.2%, 37.1% within ±20%.
Rolling-origin, out-of-sample, own-listing floors excluded for gifts. It is evidence about the method, not a promise about the next item.
Method
- Quantiles: p25/p50/p75/p90 use linear interpolation between order statistics (Hyndman–Fan type 7).
- Time to sale: (sold_at − found) / 3600, only for status=sold with a public sale price and time.
- Outcome: sale price minus alert price; above > 0, at = 0, below < 0. Auctions excluded.
- Floor change: 100 × (floor_t − floor_t−1) / floor_t−1, for collections with a positive floor on both days.
- Coverage: keys in floors-daily.json; listings = sum of buy_now_listings for the day (60/model cap).
- Valuation: settled count from /data/valuation-accuracy.json; no error metric before n≥30. Backtest from the labelled holdout file.
Computed from our logs and public Fragment data. Reproduce with radar-ceo/tools/proof_engine.py. · Updated 2026-11-01 20:04:01 UTC
Machine-readable: /data/proof.json