# DC3.1 Socials and project on top of chain (DC Analyst 3)

## READ THIS FIRST (result v1, 7 Oct 2026 10:30 Bangkok)
- **Verdict: nothing social or project adds money yet. One lead is worth a cheap live test.**
- **X with 3+ likes (the only fair way to compare old weeks): DEAD.** AUC 0.33 to 0.55 at trigger, +1 h and +4 h, both halves. Ticker counts even point the wrong way (copycat tickers).
- **X with ALL posts (only possible for the last 7 days, 30 Sep to 4 Oct, 39 coins): LEAD.** Real authors by trigger +1 h: AUC 0.86 for 5x (5 winners, all had 2 to 8 authors). One half only, 5 winners, so NOT proven. Cause of the gap: early posts on winners have under 3 likes, and X search hides 0 to 2 like posts after about 7 days. So this can only be tested by **saving posts live** (ask below).
- **DexScreener paid profile before trigger: DEAD** (AUC 0.54 / 0.47). Paid ever looks good (0.78 in A) but it is paid AFTER the run: leak, not a signal.
- **Project X account, website, telegram (GeckoTerminal info): DEAD** (AUC 0.51 to 0.62, only 11% to 15% of coins have them).
- **Project cards (deep.db, lane radar): NOT TESTABLE on this period.** Cards start 5 Oct 2026; only 1 of 150 sample coins has one. Can be tested forward from now.
- **DexScreener lost the coins:** 139 of 150 tokens return no pair on DexScreener today. Socials from DexScreener only exist for survivors (bias). GeckoTerminal info still had all 150.
- **Cost:** ~57,000 X credits (cap hit, 165 searches, ~345 credits a search incl. the empty re-read). Live cost if adopted: 1 search per stage, 3 stages = ~1,000 credits a coin = ~43,000 a day at 43 triggered coins. GeckoTerminal + DexScreener: free.
- **Next (asks):** (1) Price Predict saves ALL X posts (CA and $TICKER, no like filter) for every triggered coin at T, T+1 h, T+4 h, live, from today. In ~10 days we test the author lead on two halves for ~1,000 credits a coin. (2) Optional now: 60,000 more credits to run the all-posts read on every triggered coin 30 Sep to 4 Oct (~200 coins), which makes the lead 3x bigger but still one week.

## Per coin X recipe (for the Scout and the Coder; reuse as is)
- Tool: xsafe (`import xsafe as xs`), never raw twitterapi.io.
- Query: `(<full CA lowercase> OR $<TICKER>)` (drop the $TICKER part if the ticker is not 2 to 12 letters/digits starting with a letter). Window `since = launch - 2 h`, `until = decision time`. One `xs.search(q, since, until, pages=4)` per coin per stage; `complete=False` means "at least N".
- **Do NOT add min_faves:3 for live use.** Live, the post is young, so read ALL posts; the likes filter only exists to compare old weeks, and it kills the signal (DC3.1).
- Count from `real` only (xsafe 'full'): drop the project's own handle (GeckoTerminal `twitter_handle`); keep a post if it has the CA, or has the $TICKER AND says robinhood / rh / hood chain / long.xyz / pons / fomo / stonk (copycat tickers on other chains and other contracts).
- Features: `auth` = unique authors of kept posts; `fol` = their follower sum; `big` = authors with 10k+ followers; `lead` = minutes from first kept post to decision. Never count `spam` (bot feeds follow price).
- Cost: ~150 credits a page, an empty window costs 2 pages (xsafe re-reads empties once). Typical survivor coin 1 to 4 pages = 150 to 600 credits a stage.
- Code: http://159.223.88.168/hx-c9ois/dc/dc-a3-xrecipe.py

## Results table (sample 150, weights 1 for 3x+, 9.36 for the rest)
- L_ = X posts with 3+ likes. U_ = all X posts (30 Sep to 4 Oct only, so A is empty). _T = known at trigger, _T1 = by trigger +1 h, _T4 = by +4 h.
- auth = unique real authors (CA posts, or $TICKER posts that also say robinhood / rh / a RH launchpad). fol = their follower sum. big = authors with 10k+ followers. tickrh = ticker posts with RH context. spam = posts xsafe removed (bot feeds, shills).
- $ per coin: $100 ticket, +$400 at 5x, -$75 otherwise. Buying every triggered coin = -$67 a coin in half B (the money model is crude and harsh; replace with DC1.1 / DCC.1 book when it lands).
- "Beats random" for U_ rows is NOT valid: U only covers 30 Sep to 4 Oct while random picks come from all of half B.

| feature | share with it | AUC 5x A | AUC 5x B | AUC 3x A | AUC 3x B | $ per coin B (base -67) | beats random B |
|---|---|---|---|---|---|---|---|
| paid_before_T | 0.08 | 0.54 | 0.47 | 0.63 | 0.47 | -75 | 0.00 |
| paid_before_T1h | 0.15 | 0.71 | 0.53 | 0.71 | 0.49 | -64 | 0.48 |
| paid_ever | 0.19 | 0.78 | 0.59 | 0.73 | 0.50 | -57 | 0.94 |
| gt_has_x | 0.12 | 0.51 | 0.62 | 0.55 | 0.47 | -52 | 0.87 |
| gt_has_site | 0.11 | 0.53 | 0.53 | 0.54 | 0.47 | -65 | 0.80 |
| gt_desc_len | 0.15 | 0.52 | 0.61 | 0.54 | 0.48 | -56 | 0.95 |
| L_ca_T1 | 0.03 | 0.47 | 0.49 | 0.54 | 0.48 | -78 | 0.00 |
| L_tickrh_T1 | 0.21 | 0.40 | 0.55 | 0.46 | 0.50 | -59 | 0.96 |
| L_auth_T1 | 0.23 | 0.37 | 0.55 | 0.50 | 0.48 | -61 | 0.80 |
| L_fol_T1 | 0.23 | 0.37 | 0.54 | 0.51 | 0.47 | -61 | 0.81 |
| L_big_T1 | 0.11 | 0.47 | 0.48 | 0.49 | 0.49 | -66 | 0.77 |
| L_auth_T4 | 0.29 | 0.33 | 0.52 | 0.48 | 0.49 | -63 | 0.89 |
| L_big_T4 | 0.16 | 0.45 | 0.50 | 0.50 | 0.50 | -62 | 0.89 |
| L_spam_T4 | 0.22 | 0.50 | 0.48 | 0.43 | 0.49 | -66 | 0.49 |
| U_auth_T | 0.17 |  | 0.79 |  | 0.60 | -52 | 0.99 |
| U_auth_T1 | 0.21 |  | 0.86 |  | 0.60 | -54 | 1.00 |
| U_fol_T1 | 0.21 |  | 0.64 |  | 0.58 | -54 | 1.00 |
| U_big_T1 | 0.09 |  | 0.52 |  | 0.55 | -50 | 0.89 |
| U_tickrh_T1 | 0.20 |  | 0.79 |  | 0.59 | -53 | 0.99 |
| U_spam_T1 | 0.19 |  | 0.71 |  | 0.47 | -58 | 0.98 |
| U_auth_T4 | 0.22 |  | 0.81 |  | 0.61 | -56 | 1.00 |

## Trap list
- **Copycat tickers:** same ticker, different contract, same days (VRAX on 30 Sep and 1 Oct, NUTFLEX twice on 30 Sep). $TICKER search mixes them. Use CA posts, or ticker posts only from authors who also posted the CA.
- **Likes filter hides early real people.** 3+ like posts favour loud, late, paid posts. The winners' first posts are small accounts with few likes.
- **Bot feeds follow price.** Spam/bot posts (xsafe removed) rise on winners (U_spam_T1 AUC 0.71) because they react to the pump; never count them as demand.
- **"Twitter" field is often not the project.** ISWC's website field is someone else's tweet (x.com/ellengu18/status/2099921633905344898). GeckoTerminal handle = whatever the deployer pasted.
- **DexScreener paid after the run** looks predictive on old data (paid_ever). Only the payment time before the decision counts.
- **DexScreener drops dead pairs** (139 of 150): any social table built from DexScreener today is survivors only.

## Files
- Per coin features: http://159.223.88.168/hx-c9ois/dc/dc-a3-1-features.csv.gz (150 rows, full addresses, all features, labels, weights).
- Eval per feature: http://159.223.88.168/hx-c9ois/dc/dc-a3-1-eval.csv
- Raw X posts (xsafe real + removed, per coin, per mode): http://159.223.88.168/hx-c9ois/dc/dc-a3-1-xposts.json
- DexScreener paid orders + socials: http://159.223.88.168/hx-c9ois/dc/dc-a3-1-dex.csv

## Pre-registration (written 10:20 Bangkok, before any X call)
## Coins (provisional, switch to DC1.1 universe when it lands)
- Uniswap v4 launches 15 Sep to 4 Oct from dc/data/launches_15sep_4oct.csv.gz.
- Trigger T1 = sellable cap $250k reached (`ts_250000`) AND 100+ unique buyers. Unique buyers exist only at 1, 5 and 30 min, so trigger time = later of ($250k time, first of those checkpoints with 100+ buyers). Coins that get 100 buyers only after 30 min are missed (known gap).
- Skipped: campaign 5059_7adb2ec7, junk 25533_3080c1dd, bundle 341_6de2bb44, first cap $2M+.
- Result: **852 triggered coins**, 20 days, about 43 a day. Half A = 15 to 24 Sep (430 coins), half B = 25 Sep to 4 Oct (422).
- Outcome x = best sellable peak (7 d peak if the coin crossed $500k, else 6 h peak) / cap at trigger. 5x+ = 12 coins (5 in A, 7 in B). 3x+ = 66.
- **Sample of 150:** all 66 coins with 3x+ (includes all 12 at 5x+), plus 84 random coins under 3x (seed 20261007). Losers carry weight 9.36 so every $/day number is for the full 852, not the enriched sample.
- Known bias: this v4 table misses most big runners (only 3 of the 54 coins that reached $20M are in it; the rest live on Pons V2, v3, v2). Side panel: the 26 "class A" $20M coins (started small) get the same X features at their $250k and $1M times, for description only (not in the $/day test).

## Labels
- Main: 5x+ (honest, from trigger). Second: 3x+ (enough positives for a stable AUC).

## Features (known at trigger time T, then cumulative at T+1h and T+4h)
X, all through xsafe, real people only ('full' filter), and only posts with 3+ likes (xsafe comparable rule: X search drops 0-like posts after ~7 days, so only 3+ like posts compare across dates). Query `min_faves:3` keeps cost down.
- X1 `ca_posts`: posts with the full contract address, launch to T.
- X2 `tick_posts`: posts with $TICKER, launch to T (ticker collision risk, X2b keeps only ticker posts that also say robinhood / rh / a RH launchpad or the CA).
- X3 `authors`: unique real authors (CA or ticker).
- X4 `fol_sum`: follower sum of those authors.
- X5 `big_auth`: authors with 10k+ followers.
- X6 `lead_min`: minutes from first real post to T (none = 0).
- Mood by meaning: phase 2, only if X1 to X6 show signal (costs Anthropic API).
Project X account (handle from DexScreener socials):
- P1 `has_x`, P2 `acct_age_d` at T (clean), P3 `followers_now` (LEAKS: measured today, after the outcome; reported but not used in rules).
DexScreener:
- D1 `paid_before_T` (paid profile order time before T), D2 `boost_before_T`.
Project card (deep.db, only cards read before T+5 min):
- C1 `has_card`, C2 `build`, C3 `demand_fit`, C4 `code_own`, C5 `newness` novel vs copy, C6 `has_site`.
- C7 `cat_heat`: same-family coins triggered in the 24 h before T.

## Money model (until DC1.1 / DCC.1 baseline book lands)
- $100 at trigger. Honest exits (Second Window rules): sell at 5x, else 0.3x stop or pull. Payoff: 5x+ = +$400, all others = -$75 (mix of 0.3x stops and ~0x pulls). Fees ignored at $100 size.
- Baseline = buy every triggered coin. A feature is worth something only if a one-cut rule picked on half A also beats baseline $/day in half B AND beats 95% of random same-size picks in B.
- Cuts tried on A (fixed list): >0, top half, top quarter of the feature.

## Report per feature
- AUC half A / half B (5x and 3x labels), $/day of the A-picked cut in B, availability within 5 min of T, cost per coin (credits, $), traps.

## Budget
- X: max 60,000 credits for DC3.1, `xs.budget()` before the bulk run, spend logged here.


## Changes vs pre-registration
- X job hit the 380 page cap at coin 126 of 150: L features exist for 126 coins (all 66 coins with 3x+, 60 of 84 random losers), U for 39 of 47 coins dated 30 Sep+.
- Mood by meaning not run (no X signal to justify the API cost). Project X account age and followers not pulled (handle exists for only 12%).
- Side panel ($20M coins) not run: budget used.

Version: v1.1 (X recipe added), 7 Oct 2026 10:35 Bangkok, DC Analyst 3. Latest: http://159.223.88.168/hx-c9ois/dc/dc-a3-1-social.md