# pump.fun launch-tape data: datasheet (Avasis / QuantBase, 6 Nov 2026)

**What this is.** Per-launch evidence about how a pump.fun token was launched: who bought in the first slots, how many of those
buyers were bundled, serial snipers, or shared a funder with each other, and what the creator wallet did before and during
the launch. All of it is derived from public Solana chain data. It is **data and risk tooling, not trading advice**. Nothing here
predicts price, and a verdict is not a buy or sell signal (see "Honest limits").

## Coverage
| Item | Value |
|---|---|
| Chain / venue | Solana, pump.fun bonding-curve launches |
| Token rows (`meme_tokens`) | ~31k-40k launches/day recorded (30 Oct-5 Nov 2026: 30,949 / 31,704 / 30,885 / 34,926 / 34,569 / 40,269 / 34,788) |
| Launch-tape features (`meme_tape_features`) | since 13 Mar 2024 (sparse early), ~70-83% of launches/day in the last week |
| Live trade tape | real-time, ~6 h rolling retention; the first 30 s after a launch is read live |
| Batch features land | ~90 min after a launch (p50, measured) |
| Creator records | lifetime launches and graduations per creator wallet, plus funding-cluster aggregates |
| Holder snapshot | RugCheck top-10 share, only after graduation (81 of 7,430 sample rows) |

## The free sample
- `pumpfun-launch-sample-20261030-20261105.csv.gz` (0.65 MB; 1.8 MB as `.csv` after `gunzip`). `SHA256SUMS` lists the hash of the `.csv.gz` as downloaded and of the `.csv` after `gunzip` (the `.csv` line fails `sha256sum -c` until you have unpacked it).
- 7 full UTC days, 30 Oct 00:00Z to 6 Nov 00:00Z 2026. A deterministic 1-in-32 sample of all launches: a mint is in when
  `int(md5(mint)[0:8], 16) % 32 == 0`, so anyone can re-derive the same set. That gives 7,430 launches (952-1,261 a day).
- 37 columns: the raw evidence fields below + `verdict`, `reason_codes` (pipe-separated) and `verdict_basis`.
- Public chain data only: mint, symbol and creator wallet are on-chain. There is no personal or off-chain data. Symbols are cleaned
  (Unicode NFKC, hidden/control/bidi/tag characters removed, max 16 code points) because launchers can put anything in them.
- Verdicts in the sample are computed from the **batch** features (bundles and funder groups included). The live API answers
  fresh launches from the live 30 s window and says so in `not_evaluated`.

Sample verdict mix: WARN 3,784 (51%), DANGER 2,341 (32%), SAFE 1,305 (18%). 1,820 rows (24.5%) have no batch tape features;
those can only be WARN or DANGER (no evidence is never SAFE).

## Column dictionary (sample CSV)
| Column | Meaning |
|---|---|
| mint, symbol, creator, created_at, graduated_at | token identity, creator wallet, launch time and bonding-curve graduation time (UTC, empty = not graduated) |
| initial_buy_tokens | creator's buy in the create transaction, whole tokens (supply is 1,000,000,000) |
| creator_prior_launches / _graduations | creator's launches and graduations before this one |
| creator_sold_by_10m | the creator sold within 10 min |
| window_s, tape_complete, tape_note | launch-tape window length (s) and whether the tape had a gap |
| pct_supply_first_2_slots | % of supply bought by anyone in the first 2 slots |
| creator_supply_pct | % of supply the creator bought in that window |
| creation_slot_buyers, first2_slot_buyers, early_1s_buyers | distinct non-creator buyers in the creation slot, first 2 slots, first second |
| serial_snipers | early buyers already seen sniping OTHER launches earlier |
| bundle_txs, bundled_wallets, bundled_supply_pct | multi-buyer bundled transactions at launch and the supply they took |
| largest_funder_group, shared_funder_wallets | early buyers that share a funding wallet (largest group, total) |
| creator_sold_in_window | creator sold inside the tape window |
| cohort_deployer_funded | early buyers funded by the deployer wallet |
| name/symbol_collisions_7d, image_collisions | metadata reused by other launches (clone detection) |
| creator_launches / _graduations | creator wallet's lifetime totals |
| cluster_size / _launches / _graduations | the creator's funding cluster (wallets linked by funding) |
| top10_ex_pool_pct | RugCheck top-10 holders excluding the pool (post-graduation only) |
| verdict, reason_codes, verdict_basis | see below |

## How a verdict is computed (published live at `GET https://agents.avasis.ai/v1/sol/launch-intel/info`)
Fixed thresholds, calibrated on 32,539 launches (24 h to 6 Nov 09:00Z): WARN near p90, DANGER near p99.
- Dev launch buy >= 10% WARN / >= 20% DANGER of supply.
- Other wallets' first-2-slot supply >= 15% / >= 30%.
- Bundled wallets >= 3 / >= 6; bundled supply >= 5% / >= 15%.
- Serial snipers >= 4 / >= 11; 1-second buyers >= 6 / >= 16.
- Shared-funder group >= 3 / >= 5; any early buyer funded by the deployer: DANGER.
- Creator launch farm (>= 20 launches, 0 graduations) or a funding-cluster farm (>= 5 wallets, >= 50 launches, < 0.5% graduate): DANGER.
- Serial launcher (>= 3 prior), dev sold early, clone metadata (>= 5 collisions or a reused image), RugCheck top-10 >= 30%: WARN (top-10 >= 50%: DANGER).
- The worst reason wins. No launch evidence is never SAFE (INSUFFICIENT_LAUNCH_DATA). Every reason carries its value and threshold.

Most frequent reasons in the sample: SERIAL_LAUNCHER 2,118, CLONE_METADATA 2,101, DEV_SOLD_EARLY 1,813, CREATOR_LAUNCH_FARM 1,717,
INSUFFICIENT_LAUNCH_DATA 738, EARLY_SUPPLY_CONCENTRATION 689, SERIAL_SNIPERS 638, EARLY_1S_BUYERS 603, CLUSTER_LAUNCH_FARM 372.

Live accuracy of the 30 s window vs the batch features (prod, 230 launches): first-2-slot supply MAE 0.012 pp, non-dev buyer counts
equal on 218/230, dev-sold-in-window equal on 225/230, serial snipers equal on 245/300 (MAE 0.5).

## Honest limits
- **Not a prediction.** In this sample, DANGER launches graduated MORE often (130/2,341 = 5.6%) than WARN (1.0%) or SAFE (1 of 1,305).
  Launches that take off attract snipers, bundles and early concentration, and the post-graduation holder check exists only for
  graduates. A verdict describes how a launch was run, not whether its price will rise or fall. Do not trade on it.
- Snipers, bundles and funder links are heuristics over observed transactions; a wallet seen for the first time has no sniping history.
- Batch features cover ~70-83% of launches; a launch with no trades in its window gets no features.
- 61 sample rows have a tape gap (`tape_complete=false`): counts there may be low.
- The live tape keeps ~6 h; older launches are answered from batch features only.
- The dev's current balance is not in this data (the paid token verdict reads live holders).

## How to get more
| Way | Price | Notes |
|---|---|---|
| Free | $0 | `/v1/sol/launch-intel/info`, `/v1/sol/launch-intel/sample/{trap-check,creator,launch-firehose}` |
| Trap check, one mint | $0.02 | `GET /v1/sol/trap-check/{mint}` (x402, USDC on Base) |
| Creator record | $0.01 | `GET /v1/sol/creator/{wallet}` |
| Launch firehose | $0.002 / poll, or $1 / 7 days, $4 / 30 days | `GET /v1/sol/launch-firehose`; a pass gives SSE streaming |
| Bulk history / custom export | ask | the history above; daily files are possible |

Base URL: `https://agents.avasis.ai`. Contact: see EMAIL_DRAFT.md (sender to be confirmed by Abhay).