---
name: price-driver-model
description: The "Float Versus Flow" price-driver model and 100-point rating rubric Ben commissioned 31 Aug 2026, plus the open decision on replacing Frame 1 of his standing three-frame format. Read before any project rating, valuation or comp question, alongside [[gem-hunter-doctrine]] and [[crypto-token-research]].
sources: [cowork]
aliases: [float-versus-flow, rating-rubric, what-drives-price]
---

# Price driver model

Artifact: https://claude.ai/code/artifact/6e6fe189-1521-4de4-a535-e042932395fc

## What Ben asked for, 31 Aug 2026
- [stated] Rejected the traditional-company lens on crypto projects: said the sector routinely gives chains billion and ten-billion dollar valuations with almost nothing happening on chain, so revenue and usage framing misprices it.
- [stated] Defines "successful" as one thing only: price goes up.
- [stated] Wants the actual drivers of price identified from past cycles and other chains, for fundamental projects rather than memes, then a rating system built on those drivers so projects can be measured on the correct things.
- [stated] Believes a new bull cycle is starting now (Aug 2026).

## The rubric that was built
100 points, three blocks, three hard gates. Blocks: SUPPLY 40 (unlock coverage ratio 18, unlock recipient 12, cost basis of float 10), DEMAND 35 (named marginal buyer 15, effective routing yield 12, access channels closed 8), RERATING ROOM 25 (starting market cap 10, entry valuation vs sector band 8, narrative slot purity 7). Modifiers: +5 reflexivity, -10 post-delivery. Gates: coverage ratio above 5x, no nameable buyer beyond a listing-day market maker, reserve test failure. Tiers map to his existing Full / Starter / Probe / Zero.

Headline metric to compute on every screen: **unlock coverage ratio = next-12m unlock notional at spot divided by every nameable dollar bid.** Second: **effective routing yield = (annualised fees x % actually routed to the token) / market cap**, which replaces fees-to-market-cap.

## Open decision for Ben, not yet answered
His standing three-frame format anchors Frame 1 on the sector's revenue or TVL multiple. Both are formally refuted as predictors (TVL: arXiv 2506.03287, all alphas p=0.13-0.99, GRS fails to reject; revenue multiples: SSRN 6901559, rho=0.04, Fama-MacBeth p=0.98). Proposed replacement: Frame 1 becomes the forced-seller calendar. Frames 2 and 3 unchanged. **Ben has not yet accepted or rejected this. Do not change /preferences.md until he does.**

## Audit, 31 Aug 2026 (Ben asked for a full re-check)
All 7 academic citations verified exactly against primary sources, including the two SSRN preprints suspected of being fabricated. 9 industry figures were wrong. The three that mattered: Arbitrum "sequencer revenue $103.1M" was untraceable and DefiLlama's primary series says trailing-1y fees are $11.4M; HYPE's routing yield is 5.1% not 7.1% and it does not lead its peers today, though it was ~34% at entry; and the "Binance listing effect inverted" claim compares a 24-hour pop against an unbounded-hold hand count, so it is not a valid comparison. Also: the Dec 2025 Hyperliquid vote changed supply accounting, it did NOT raise fee routing to 99%. Delphi cites Jupiter as a buyback FAILURE ($3.77 unlocked per $1 bought), not an outperformer, and puts Aave at 0.90x coverage.

Worst error, corrected: revision 1 called a hindsight scoring exercise a "back-test". Every score was assigned after the outcomes were known, on a sample selected on outcome. All weights, the 5x gate threshold and the tier boundaries are conventions, not estimated parameters. The artifact now carries a three-tier evidence split (external and replicated / observed in a selected sample / convention) and a corrections log. The rubric is a consistency device, not a predictor.

STANDING LESSON: subagent research output must be spot-checked against primary APIs before it goes into a deliverable. Three assembled-but-wrong numbers survived into revision 1 because they were plausible and well formatted.

## Anti-signals established, must score zero
TVL as a valuation anchor, cheapness on a revenue multiple, developer activity and GitHub commits, TPS and node counts and partnerships, infra adoption without a toll booth, supply sinks on an already-distributed float, unsized buybacks, exchange listings as a catalyst (the Binance effect inverted: +41% in 2022-23 to -54.6% in 2024), VC quality and funding size (the investor is the seller).

## Known data gaps, do not paper over
Tradeable float is unobservable from public data. No study tests unlock size normalised by daily volume. No published event study exists for perp listings or ETF inclusion. The capital multiplier ranges 5x to 50x across sources. "One winner per narrative" is an observed pattern with no supporting study.

## Out-of-sample test, DeFi Summer 2020 (Ben asked for it 31 Aug 2026)
The framework was built without DeFi Summer, so applying the frozen rubric to it is the closest thing to a real test. Result: 11 of 14 directionally right. Two systematic misses, UNI and AAVE, both gated on emission-against-current-revenue and both became long-run survivors. One uncovered category: YAM was a clean fair launch that died in 36 hours from a rebase bug, and the rubric has no security factor.

Strongest evidence in the whole exercise, verified directly from DefiLlama: 1 Sep to 5 Nov 2020, total DeFi TVL rose 12.2% while the tokens fell 50 to 84%. MKR was +1.6% over that window and is the only token in the cohort with zero emission and 100% revenue routing. Still the only one above its DeFi Summer peak today, +100.6%.

TWO DESIGN FIXES this produced:
1. Replace A1's fuzzy denominator with the EMISSION RATIO = annualised emission notional / annualised protocol revenue. Both inputs on DefiLlama, no judgement needed. Sorted the 2020 cohort almost perfectly: SUSHI 1,227x, CRV infinite, COMP 51x, MKR 0x.
2. B1 trap: a yield-farming reward denominated in the token is a SELLER, not a buyer. True in 10 of 14 DeFi Summer cases. Must score zero, not fifteen.

## Robinhood Chain is NOT a DeFi Summer
Verified via the Merkl API: 31 live incentive campaigns paying about $88,179/day, of which 99.68% is paid in stablecoins, stablecoin vault shares or another chain's token (USDe 47.3%, steakUSDG 43.3%, UNI 8.4%). The only genuine native governance-token emission is EARN at $282/day. DeFi Summer's engine was protocols printing their own tokens for free; here two companies pay cash from a finite budget.
Morpho Blue alone is 67.6% of chain TVL and Steakhouse curates 96.9% of that, neither native. Genuinely RH-native protocols hold $40.8M, 5.6% of chain TVL. The $967M lending stack produced $1.97M of 30d fees and $0 revenue. The largest campaign, $419.6M, is whitelisted "[Robinhood Users Only]" retail earning a cash-sweep rate that cannot compose.
THE INVERSION THAT MATTERS: RH Chain native tokens have no emissions and mostly 100% float, so gate 1 structurally cannot fire on them. That is the opposite of, and better than, a DeFi Summer token. Confirmed in the tape: every VC-backed 2026 launch is low float and below its ATH (Zama -22.3% to MemeCore -80.9%), while PONS at 100.4% float sits 13% off an ATH set 30 Aug.
PONS scored 80 (Full, at the boundary). Its entire C2 and most of B2 rest on five days of revenue acceleration ($109k/day over 30d, $356k over 7d, $807,899 in 24h). Falsifier: if Pons daily revenue does not hold above about $2M/day past mid-September it drops to Probe. Watch also whether the Ethena USDe campaign renews after 3 Sep 2026.

## Micro cap framework, 31 Aug 2026. Artifact "Buying the Tail": https://claude.ai/code/artifact/5036ede8-7fb6-487c-b52b-5192e4dad0be
- [stated] Ben does not care about multibillion dollar projects. He wants the framework focused on micro, small and mid caps, on the DeFi Summer pattern: new projects appearing, new tech, some dead, some scams, some going to high valuations. He asked specifically whether the framework could have called DELTA three weeks ago.

THE DELTA TEST, run honestly. Peer set defined objectively: 15 RH-chain native tokens with a pool at least 3 days old on 10 Aug 2026. Result: **14 of 15 went up, median +912%.** DELTA did +3,942% and finished 3rd. But DELTA ranked 7/15 on market cap, 10/15 on drawdown, 8/15 on volume. It was mid-pack on everything measurable. **The answer to Ben's question is NO, I could not have singled out DELTA, and a rule that ignored the product would have beaten it** (buy the five smallest = median +2,331%, catching VAULT +7,824% at an $18k cap and INJOH +4,348%).

WHAT ACTUALLY RANKED THE RETURNS (Spearman, n=15): market cap -0.771, age -0.675, volume -0.600, pair count -0.550, turnover +0.414, drawdown -0.364. Size dominates. Turnover is the only positive-sign feature and the only one where DELTA stood out (3/15 at 0.52x). Deeper than -55% off ATH: median +2,026% vs +175% shallower, which SUPPORTS Ben's standing rule about not calling tokens late at 60-70% off ATH.
CAVEATS THAT MUST TRAVEL WITH IT: severe survivorship bias (universe built from tokens alive today), n=15, one window, one chain, violent up-tape. The ranking is informative, the +912% level is fiction.

BASE RATES (measured, not asserted): pump.fun 0.996% graduate (n=860,140), 6.73% of graduates alive at 30d, 0.013% of all launches ever reach $1M (1 in 7,750), zero reached $100M. DeFi Summer denominator 23,597 tokens, 1.92% alive in 2026, 7 of 13 food coins peaked on their first traded day. Median 100x takes 490 days with a 93% pre-peak drawdown. Top 1% carries 71-97% of cohort return.

DAY-0 SIGNALS FROM THE DEFI SUMMER RECORD: what works is (1) usage/TVL that arrived BEFORE any token incentive (yearn had $8.06M unbribed on 16 Jul 2020 after 156 days live; every death had a zero-day product-token gap), (2) enumerate the keys and name the holder, (3) find the diff and ask if it touches the money. **What is WORTHLESS: audits (everyone was unaudited, and the null result is now properly tested across 4,108 protocols with PSM and IV), and fair-launch/0% team (YAM copied it verbatim and died in 36 hours).** Same-chain emission clones are 0-for-N.
CORRECTION TO MY OWN EARLIER CLAIM: YFI's supply was NOT visibly final on day 0. An infinite mint key existed and was publicly known; it was surrendered on day 4 by governance. "Fixed supply, final after 9 days" describes the outcome, not what a buyer could verify.

THE FRAMEWORK SHAPE: not a score, because scores imply discrimination the evidence does not support. Stage 1 survival filter (binary), Stage 2 position in the power law (market cap, turnover, drawdown), Stage 3 basket not pick (6-10 survivors weighted small beats one conviction position, and it eases his exit-depth constraint), Stage 4 exit rule (emission schedule is the holding period; exit 30 days before a cliff; a third to two thirds of peak was given back across this cohort).

## THIRD AUDIT, 31 Aug 2026. Ten corrections, all to my own published claims
MY PEER SET WAS INCOMPLETE. Built from a hand-assembled 34-token list rather than a query, it omitted 8 eligible tokens including CHUMP (0x0e0d2c89a5a019fe1cf762e5e33187631dacc21b), which did +15,036% and was the cohort's BEST performer. Also missed RDDT, a second loser. Corrected set n=17, DELTA ranks 4th not 3rd. STANDING RULE: any peer set built by hand is incomplete until it comes from the radar queue.
UP HAD 3.9% FLOAT, so mcap/price does not give a constant supply and its reconstructed 10 Aug market cap was invalid. Removed. Float check (DexScreener mc/fdv) is now mandatory before any historical market-cap reconstruction. 13 of 16 RH tokens are 100% float so the assumption is usually safe, but ARROW 90%, NET 7.9% and UP 3.9% are not.
ONLY MARKET CAP SURVIVES. rho -0.713, p=0.0013, holds under Bonferroni. TURNOVER RETRACTED (+0.368, p=0.15). DRAWDOWN RETRACTED and the SIGN FLIPPED from -0.364 to +0.127, p=0.63. **I told Ben his ATH rule was supported by this data. It is not. That was overreach and it was worse because it told him what he already believed.** His rule may still be right; this dataset says nothing either way.
Market cap and volume correlate at rho +0.87, market cap and age at +0.57. They were never independent findings.

THE WORST ERROR: the audit claim. Landsman/Lyandres/Maydew/Rabetti/Zhang (Journal of Accounting and Economics, 4,108 protocols, PSM and IV) is real and the sample is exact, BUT its outcome variable is HACK OCCURRENCE, not price or returns, AND it finds top-tier audits and bounty programmes DO reduce breach likelihood. I used it to tell Ben to stop checking audits. That inverts the paper. Correct rule: "audited yes/no" is uninformative, WHO audited is informative.

OTHER FIXES: pump.fun graduate survival I quoted (6.73% at 30d) is LOWER than CoinGecko's published all-token survival (10.05% at 30d, 4.55% at 90d, n=18.67M) which is backwards; use CoinGecko's. "Median 490 days to a 100x" and "top 1% carries 71-97%" are self-computed with NO published counterpart, now labelled unverifiable. Solidus 98.6% DOES have a published methodology (tokens falling under $1,000 liquidity); the correct objection is that it measures failure and calls it fraud. Published pump.fun graduation range is 0.5-2%, not 0.8-3%.
VERIFIED AND UNCHANGED: yearn TVL $8,058,140 on 16 Jul 2020 against a series starting 12 Feb 2020, confirming the 156-day product-before-token gap. PancakeSwap $596M. SUSHI -99.18%, PICKLE -99.98%, CREAM -99.81%. Scam rates 50.14% (Xia et al., peer-reviewed) and 76.42%. DeFi Summer denominator 23,597 corroborated by Xia et al.'s independent 21,778.

CONTAINER LIMITS FOUND: the Robinhood RPC (rpc.mainnet.chain.robinhood.com) and Blockscout API both return 403 from the cloud container. Use DexScreener mc/fdv for float instead, or the Mac bridge for chain calls.

## X / marketing thesis tested, 31 Aug 2026. Artifact "The Attention Trade": https://claude.ai/code/artifact/b61b527d-31a8-4021-bda5-6a63dc3cce57
- [stated] Ben's thesis: crypto is full of unsophisticated money, so price is driven mainly by marketing (X mentions, YouTube, KOLs, narrative spin) rather than adoption or usage; who buys it matters because others follow them; things are valued on potential; "lower your IQ, it's a bull market". He asked it be checked against DeFi Summer and past cycles and turned into scoring criteria.

ORIGINAL TEST RUN (the literature says nobody has done this on micro caps): ~1,100 X posts across 16 RH-chain tokens via twitterapi.io, measured 25 Jul to 10 Aug, against the 10 to 31 Aug return. **NOT ONE attention metric survives controlling for log market cap.** Posts partial -0.110, authors -0.020, views +0.003, engagement -0.301 (p=0.26), accounts>50k +0.021, max follower account +0.058, engagement coefficient -0.170. Market cap alone -0.735. Attention correlates with market cap at +0.46, so it is largely a size proxy.
THE TWO CASES THAT TELL IT: **CHUMP, best performer at +15,036%, had ONE post before the move, from its own 154-follower account. Two Chinese KOLs (137,573 and 255,165 followers) posted on 26 Aug, two weeks AFTER.** STONKBROKER had the HIGHEST engagement (4,884) and most views (287,846) with 5 big accounts, and was the ONLY loser at -29%.

THESIS GRADED: premise CORRECT and strong (BIS: ~75% downloaded above $20k BTC, median lost $431 of $900; JFE 200k eToro accounts: same people are momentum in crypto and contrarian in equities). "Marketing drives price" right for 1-5 days, wrong beyond (Merkley et al, Review of Accounting Studies, 35,569 tweets: +1.83% day 1, -1.02% days 2-5, -6.53% day 30, -19% at 3 months). **"Get the right big people on board" is REFUTED IN DIRECTION: 200k+ follower accounts produced -39% at one week, under-50k accounts +25% at one week and +141% at three months.** "Adoption doesn't matter" refuted at 3-6 months (Greenfield horse race, 77 projects: fundamentals +8.7pp R² at 6m, up to 88% of variance; social sentiment "negligible benefit"). "Who buys it matters" half right: +11% probability of investing on a celebrity tweet, but followers earn -0.5% abnormal while those already positioned earn +2.8%. **"Lower your IQ in a bull market" is CORRECT at one month, where fundamentals actively hurt (-0.73pp R²) and beta explains ~95%.**
DEFI SUMMER: the Vitalik/Pickle story inverts. Tweet was 13 Sep 2020 08:02:20 UTC; PICKLE's low was $4.41 on 12 Sep so the rally preceded it, and its ATH $85.24 was set ON 13 Sep. The endorsement marked the top. Also: NO paid KOL market existed in 2020, the rate-card economy is 2023-25, so 2020's amplification was organic and the emissions were the marketing.

SCORING DECISION: do NOT build a marketing-quality score. No published study has ever regressed marketing spend on price outcomes, and no attention metric had incremental power in my own test. Instead: keep X as a CLAIM source per x-claim-audit; ADD ONE KILL CHECK, heavy coverage from 50k+ accounts on a sub-$1M token is a LATE signal (-39% at one week); keep zero-coverage as a ceiling flag; treat any promoted name as a dated trade with the exit in days.
OPEN AND UNTESTED: I measured attention LEVEL, not ABNORMAL attention/acceleration, which is the version with literature support. My acceleration test failed on a pagination artifact (API returns newest-first, 6 pages never reached past 7 Aug for high-volume tokens, producing impossible "zero prior posts" readings). That is the strongest surviving version of Ben's thesis and it remains open.
BUG CAUGHT: my first X run reported zero posts for CASHCAT, PONS, AI, INDEX and STONKBROKER because the date filter discarded the newest-first first page. Always verify a zero-attention result against a bare-ticker query before reporting it.
FLAG FOR BEN: current CHUMP mentions include 9/19/25-follower accounts pushing "holder eligibility is live" and "portal is active" with the correct contract. Classic drainer pattern.

## HARDENED FRAMEWORK, 3 rounds + out-of-sample test, 31 Aug 2026. Artifact "The Scoring Problem": https://claude.ai/code/artifact/cff5c2a3-67e7-4ef5-a7f6-4f166abb8faf
- [stated] Ben corrected me: he never said the BIGGEST people on board, he said the RIGHT people. Different and sharper claim, testable per-account rather than by follower count. Not yet tested that way.
- [stated] He asked for three rounds of build, apply, optimise, then a hardened framework with tested predictive performance.

DATA BUILT PROPERLY AT LAST. Dune over dex.trades on Robinhood Chain (standing query 8562361): 80,304 distinct traded tokens in the raw pull, filtered to 1,596 with >=$150k cumulative volume and >=12 trading days. Supply from erc20_robinhood.evt_Transfer mints minus burns, assuming 18 decimals. 5 origin dates weekly 20 Jul to 17 Aug, 14-day forward returns, 4,967 observations. **Tokens that stop trading are recorded as -95%, not dropped**, so survivorship is handled. Train/test split 50/50 on MD5 of token address. All stats are Spearman WITHIN origin then averaged, which strips the market out.
REGIME SPREAD IS HUGE: 20 Jul only 11.8% of tokens rose over 14d (median -38.4%); 17 Aug 73.6% rose (median +22.7%).

FROZEN FRAMEWORK v3: gate mc<$2M AND drawdown<-50% AND turnover<0.5, then rank by z-scored composite (lmc 1.0, dd 0.6, mom7 0.4, turn 0.4 all inverted), take top 20%.
OUT-OF-SAMPLE RESULT: **mean lift +11.5pp, t=3.00, positive in 5/5 origins. Rank correlation 0.217 train vs 0.219 test, almost no degradation. Win rate 44.1% -> 54.0%. Median -4.6% -> +2.0%.**
BUT THE COSTS: **death rate 11.4% in picks vs 3.7% in universe (3x worse), >+100% rate 5.8% vs 8.9% (selects AGAINST the tail), and mean return +19.6% vs universe +20.6% so it does NOT beat buying everything on mean.**

THE CORE FINDING, and it is structural: **every feature reverses sign between the median objective and the tail objective, consistently 5/5 in both directions.** market cap -0.165 vs +0.091; volume -0.179 vs +0.110; buyers -0.183 vs +0.114; drawdown -0.147 vs +0.076; turnover -0.142 vs +0.057. Median-seeking and tail-seeking are OPPOSITE instruments. Ben hunts 10-50x, which is the tail objective, so the framework optimises the wrong thing.
v4 TAIL-SEEKING WAS BUILT AND FAILED: in-sample 14.0% big-winner rate vs 7.6% base; out of sample 8.8% vs 8.9%. Zero edge. Reported as a failure.
SURVIVAL MODEL IS THE STRONGEST COMPONENT: safest quintile 0.4% death vs riskiest 9.7%, base 3.5%, out of sample. Death predictors run t=-6 to -8, roughly triple the strength of any return predictor. USE IT FOR SIZING AND EXCLUSION.

APPLIED TO NAMED TOKENS at the 10 Aug origin: **the gate REJECTED DELTA (drawdown only -34%, turnover 0.50) which then did +1,043%, and rejected AI which did +353%.** Correctly rejected STONKBROKER (-41.5%). ARROW failed the gate by $18,737 of market cap. Hard gates are arbitrary and should become soft penalties.

CORRECTIONS TO MY OWN EARLIER SESSION CLAIMS: **deep drawdown IS a real signal (t=-2.58, 5/5 consistent) - my retraction of it on the 17-token sample was itself an overcorrection and Ben's original ATH rule stands.** Turnover is consistently NEGATIVE, not positive as I reported twice. Market cap is NOT monotonic; worst forward returns sit in deciles 6-9, so a linear size score fits a curve that is not there.
DATA QUALITY TRAP: one observation showed a +427,846,843% return from a broken VWAP print and accounted for 100% of an unwinsorised basket's return. ALWAYS winsorise at p99 and inspect the max before reporting any basket economics.

## ROUND 4: SMART-MONEY WALLET SIGNAL, 31 Aug 2026. Strongest finding of the whole session.
Ben's correction ("the RIGHT people, not the biggest") pointed at wallets, not follower counts. Built and tested with no lookahead: wallet quality scored ONLY on tokens born before 1 Aug with outcomes to 9 Aug, then applied forward to the 10 and 17 Aug origins. Dune, dex.trades on Robinhood Chain.
METHOD: every wallet buying a token within 48h of its first trade with >=$100. Winner = token later does 5x from its first-48h price. Score wallets on HIT RATE, not count. **The bot filter is the critical step: top wallets by count were early in 270/256/252 tokens, they buy everything. Require 3 to 60 positions.** 43,738 wallets scored, median hit rate 9.1%, **1,264 ELITE at >=50% hit rate with >=4 positions (chance alone gives ~3).**
RESULT, 2,166 obs: **death rate 6.4% -> 0.4%, a 15.5x reduction. Expected 46.5 deaths under the null, observed 3, z = -6.60.** Big-winner (>+100%) rate 11.6% -> 15.8%. At elite>=3, ZERO deaths in 419 obs.
NOT A SIZE ARTIFACT: stratified by market cap the effect is STRONGEST in the smallest band (<$50k: 7.3% -> 0.8% death).
**ELITE>=1 IS THE FIRST CONSTRUCTION IN THIS ENTIRE SESSION THAT BEATS THE UNIVERSE ON MEAN, DEATH AND TAIL SIMULTANEOUSLY**: mean +61.7% vs +51.6%, death 0.4% vs 4.4%, tail 16.1% vs 13.1%.
**AND IT SHOWS FRAMEWORK v3 IS WORSE THAN BUYING EVERYTHING**: v3 mean +21.9% vs universe +51.6%, death 14.2% vs 4.4%, tail 7.3% vs 13.1%. v3's +11.5pp median lift was real but the median was the wrong objective. RETIRE v3 FOR TAIL HUNTING.

DELTA STILL MISSED: zero elite buyers before 10 Aug, then +1,043%. **Four separate approaches have now missed DELTA (v3 gate, v4 tail model, X attention, elite wallets). The honest conclusion is that DELTA was not detectable by any method tested, not that a better method keeps eluding me. Stop retrofitting to it.**
CAUTION ON THE ELITE SIGNAL: elite counts track market cap closely in the named-token table (CASHCAT 78, PONS 62, STONKBROKER 74 vs DELTA 0), so it is weak for RANKING and strong only for the DEATH question. Same trap as X attention.

## NAMED GAPS, in priority order for the next run
1. **Net buy vs sell pressure. The entire panel uses token_bought only, so distribution vs accumulation has never been measured. Biggest gap, one query.**
2. Elite signal in a FALLING tape. It rests on 10 and 17 Aug only, both strong tape. Test 20 and 27 July. If it fails there it was a bull-market artifact.
3. Elite wallet EXITS as a sell trigger, which would close the exit-rule hole every version has had.
4. Deployer/operator hit rate using the same method (fires before any pool exists, ties to [[operator-graph]]).
5. New vs repeat buyers (holder acquisition), never separated.
6. Liquidity depth and trend, not one feature in the panel measures it.
7. Other chains: Solana pump.fun has years of data and the same queries run against it.

## ROUND 5, 31 Aug 2026: the two tests Ben asked for. One signal withdrawn, one found.
**THE ELITE WALLET SIGNAL IS WITHDRAWN. It was a bull-market artifact.** Rebuilt cleanly for July (wallet quality scored only on tokens born before 12 Jul, outcomes to 19 Jul, elite pool 1,188, applied to the 20 and 27 Jul origins). Result in the BEAR tape: death reduction **1.0x, z=+0.16, no effect at all**, and the tail was WORSE (3.2% vs 4.9%). In the August bull tape the same construction gave 16.0x death reduction and z=-6.72. The 15x death reduction reported in round 4 exists only when everything is being bid. Do not size on it.

**BUY/SELL IMBALANCE IS THE ONE THING THAT SURVIVED.** Closing the gap where my entire panel used token_bought only. 7-day net dollar imbalance = (buy_vol - sell_vol)/(buy_vol + sell_vol).
- vs >+100% TAIL: rho +0.064, **t = 9.22, highest of any feature tested all session, 5/5 origins**
- vs DEATH: rho -0.093, **t = -8.45, 5/5 origins**
- Survives controlling for log market cap (tail partial +0.042 t=2.41; death partial -0.037 t=-3.12). Buyer/seller headcount is slightly stronger on the tail: partial +0.070, t=3.42, 5/5.
- **NOT LINEAR. Quintiles 2-5 are indistinguishable. All information is in the bottom fifth.** So it is an EXCLUSION, not a ranking.
THE RULE: drop the worst 20% by 7-day net imbalance. Pooled: death **7.6% -> 2.4% (3.2x)**, tail **4.9% -> 9.1% (1.9x)**, mean +10.1% -> +19.9%. Holds in BOTH regimes. Honest cost: in the bear tape it HURTS the mean (kept -19.4% vs excluded -3.6%) because heavily-sold tokens had already fallen and bounced hardest.

**SIGNS ARE STABLE ACROSS REGIMES.** Splitting July bear vs August bull, every feature keeps the same sign on both tail and death. The sign flip found earlier is between OBJECTIVES (median vs tail), not between market conditions. Magnitude varies, direction does not.

**STACKING FILTERS DESTROYS THE EDGE.** elite>=1 alone in August: tail 16.1%. elite>=1 AND imbalance>0.1: n collapses to 26 and tail falls to 7.7%, below the universe. RULE: one filter at a time, never a stack. Every extra condition is a chance to fit noise.

## SCOREBOARD ACROSS FIVE ROUNDS
Four of five constructions failed a test designed to break them: v3 failed on objective (worse than universe on mean, death and tail), v4 failed out of sample, elite wallets failed on regime, stacked filters failed on sample size. **ONE survived: do not buy what is being sold.** That ratio is the real output. Complex scores on this data do not replicate; a single mechanically computed exclusion does.
NEXT, one feature at a time, each tested in BOTH regimes before admission: (1) elite wallet EXITS as a sell trigger, since their buying is a bull artifact but their selling may still inform and would close the exit-rule hole, (2) new vs repeat buyers, (3) deployer hit rate, (4) liquidity depth, (5) a second chain.

## ROUND 6, 31 Aug 2026: the ceiling measured. Trade flow is exhausted.
Instead of hand-building more features, fitted a gradient-boosted model on all 10 features with LEAVE-ONE-ORIGIN-OUT cross-validation (so out-of-sample across regimes). sklearn is available in the container.
**CROSS-VALIDATED AUC CEILING: death 0.846, touch +100% 0.644, touch +200% 0.642, close above +100% 0.626, beat median 0.631.** That is what this data supports. Death is far more predictable than the tail, confirmed and now quantified.
**GRADIENT BOOSTING BEAT LOGISTIC AND THE BEST SINGLE FEATURE ON EVERY TARGET.** Hand-weighted composites cannot represent the interactions.
**THE METHODOLOGICAL LESSON THAT MATTERS MOST: univariate screening misled me all session. 7-day momentum had rho near zero on its own (I called it weak) and is the TOP feature in the multivariate tail model at 0.17 importance.** Selecting features by univariate correlation is structurally blind. Stop hand-weighting scores; fit a model and validate held-out.
LONGER HORIZON DOES NOT HELP: 30d AUC 0.632 vs 14d 0.644. Signal dilutes, does not strengthen.

**WALLET STRUCTURE ADDS NOTHING. Trade flow is exhausted.** Built top5/top20 buy concentration, new-buyer count and share, buyer/seller headcount. Gain over flow-only: touch100 +0.010, touch200 -0.010, death +0.001, beatmed -0.001. Structure alone is informative (0.577 tail, 0.820 death) but fully redundant. **Every feature tested this session is a different summary of the same trade table.**
Directional keeper: **buying CONCENTRATION IS INVERTED. High top-5 wallet share of buy volume predicts NOT reaching +100% (AUC 0.405). Concentrated buying is a warning, not confirmation.** New-buyer count points the other way (0.591).

**PROFIT-TAKING AT A FIXED MULTIPLE DESTROYS RETURNS. This inverts the natural instinct and is the clearest practical result.** Universe hold 14d +17.6%; sell at +100% +2.5%; sell at +50% NEGATIVE (-2.3%). Every higher threshold beats every lower one, monotonically, which is the signature of a payoff you must not truncate. The give-back is real (16.0% touch +100%, only 8.3% close there, so 48% give it back) but harvesting it costs far more than it saves because the non-give-backs carry the whole return. **To reduce risk cut position count or size, never upside.**
BEST CONFIGURATION FOUND ACROSS SIX ROUNDS: rank by a gradient-boosted tail model, take the top 20%, HOLD, no profit-taking. Cross-validated **+28.2% mean per position vs +17.6% universe (1.6x), death 2.1% vs 3.4%, big-winner rate 21.2% vs 16.0%.** Top decile lift 1.42x; top 20% tail + bottom 20% death gives 24.0% touch rate and 0.0% death.

## WHERE THE NEXT GAIN MUST COME FROM (not more flow features)
1. **Contract bytecode features** (mint/pause/blacklist/setFee/owner/upgrade selectors, tax). My death label is "stopped trading", which is a DIFFERENT failure from "rugged". research-sources already has the selector scan; it has never been joined to an outcome. HIGHEST priority.
2. **Deployer and funder identity graph.** Wallets failed on regime but deployers are a far smaller population that repeats more, and it fires BEFORE any pool exists. Ties to [[operator-graph]]. HIGHEST priority.
3. **Product existence and category from [[rh-chain-radar]].** Ben's own 27 Aug work found payout tokens returned 66x operated-infrastructure tokens over 30 days. **That entire effect lives outside my feature set.** Already built, just needs joining to the panel.
4. Intraday microstructure (everything so far is daily aggregates). Medium.
5. More origins and a second chain (Solana). Medium.
DO NOT: more flow features, longer horizons, profit-taking rules. All three measured and rejected.

## ROUND 7, 2 Sep 2026: first-principles rebuild. Ben's standing instruction is now ALWAYS work from first principles (filed in /preferences.md).

**THE IDENTITY EVERYTHING REDUCES TO.** Constant product: `ΔP/P = 2·dx/X_quote = 4·NetFlow/Liquidity`. Market cap does not appear. There are exactly two terms, depth and net flow, and every claim about token selection is a claim about one of them.

**MARKET CAP IS A DEPTH PROXY, NOT A SIZE LAW.** This reinterprets the strongest finding of rounds 1-6 (mcap rho -0.713 / -0.165). Measured on 98 live top-100 RH tokens, 2 Sep: rho(log mcap, log liquidity) = **+0.753**; log-log fit `log(liq) = 1.62 + 0.73·log(mc)`, so liquidity grows SLOWER than cap and small caps are mechanically thinner. Sensitivity spread p10 to p90 is **17x**, median price move per $100k net buy **161.6%**. ACTION: rank by liquidity and by residual thinness for the cap, not by market cap. Same intent, cleaner variable, no confound with age or volume.
Median convexity by tier (% move per $100k): <$1M 328% · $1-5M 197% · $5-20M 80% · $20-100M 30% · >$100M 9%. Extremes on 2 Sep: DEEP `0x1da24f6bb623b9d1afeae3f3146659a2662d6d27` 4036%/100k on $9.9k liq at a $1.57M cap; STATICS `0x2d8d6f4a93acd7a916a5a654ec8b690ba3b3eadd` 1.7% on $23.0M liq. Convexity is symmetric: it sets POSITION SIZE, never direction. Caveat: assumes constant product, so v3/v4 concentrated liquidity makes real sensitivity worse, and DexScreener liquidity is reported not tradeable.

**DRY POWDER REFUTED. THE CAUSAL ARROW RUNS BACKWARDS.** Hypothesis: bridging stablecoins is a strictly earlier step than buying, so idle stablecoin balance should lead. Tested on 58 days of DefiLlama RH-chain stablecoin supply vs chain DEX volume. rho(stablecoin inflow_t, dex volume_t+1) = **+0.051** (t+2 -0.031, t+3 -0.082). Reverse: rho(dex volume_t, stablecoin inflow_t+1) = **+0.276**. Activity leads capital by ~5x. Chain stables went $678M to $832M in 13 days while DEX volume went $483M to $1.62B/day, volume first.
CONSEQUENCES: (1) there is no sidelines reservoir on this chain and no fuel gauge, so never wait for a stablecoin-supply signal; (2) reflexivity is not a side effect, it IS the flow model, price movement manufactures its own inflow; (3) chasing is structurally correct here, so **detection latency is the whole game**.

**NARRATIVE IS A COORDINATION EQUILIBRIUM: detectable, not predictable.** Before it forms the information exists nowhere, so there is nothing to predict from in principle. After it forms it is self-reinforcing and persistent. The decisive unmeasured number is therefore **narrative half-life vs detection lag**. Run it off the existing Dune panel (1,596 tokens, 5 origins), not the GT API.

**THE REFLEXIVITY COEFFICIENT, the object worth building.** `R = (new dollars per 1% price gain) / (incumbent dollars sold per 1% price gain)`. R>1 self-sustaining, R<1 the move dies. It is a ratio so it is regime-robust by construction, which is exactly what killed the elite-wallet level signal in round 5. Closes named gap 5 (new vs repeat buyers, never separated) and gives the exit rule every version has lacked.

**DATA BUILT 2 Sep, on disk in the container.** (a) daily OHLCV 120d for the top 100 CAs from GeckoTerminal (82 of 100 done when the session ended, `/home/claude/hist.json`); (b) all X posts citing each of the 100 CAs, 4,939 posts / 1,027 accounts / 99 tokens (`xposts.json`); (c) **the properly lagged set: posts with `until:2026-08-19`, 3,509 posts (`xearly.json`)**, which is the feature side for a clean 19 Aug -> 2 Sep forward-return test.
X ACCOUNT STRUCTURE: 1,027 accounts, **204 cover >=3 tokens and are scoreable**, **73 spray accounts at >=15 tokens must be excluded first** (same trap as top-wallets-by-count). Median poster has **970 followers**, p90 23,231, which is the mechanical reason follower count was the wrong axis all along.
**PAGINATION TRAP CONFIRMED AGAIN, third time: advanced_search returns newest-first, so 3 pages of 60 posts reached back past 21 days for only 18 of 97 tokens.** Any historical X feature MUST be pulled with an explicit `until:` date, never by paginating from the present.
NOT YET RUN, the actual test of Ben's "right accounts" claim: score each of the 204 accounts on pre-19 Aug posts against realised 14d forward return, then **split each account's record in half by date, fit on the first half, test on the second**. If the ranking does not hold out of sample the claim is hindsight; if it does, the output is a named watchlist of ~200 checkable handles.

**PROCESS NOTE:** Desktop Commander went unresponsive mid-session and the droplet was lost, forcing the OHLCV pull into the container at ~7 tokens per 2 minutes. Run long GT pulls on the droplet (`/opt/rhtop/`) when the bridge is alive.

## ROUND 7b, 2 Sep 2026: the tests completed. One new predictor admitted, the account claim refuted as tested.

**CONVEXITY MEASURED FROM THE TAPE IS THE NEW PREDICTOR, and it was never in any prior panel (it closes named gap 6, liquidity depth).** Amihud: `convexity = median(|daily return| / daily dollar volume)`, needs no liquidity data. 99 tokens with 120d daily OHLCV, 5 origins 22 Jul to 19 Aug, 14d forward, Spearman within origin.
| feature | mean rho | t | consistency |
|---|---|---|---|
| **convexity (Amihud)** | **+0.303** | **4.19** | 5/5 |
| drawdown from ATH | -0.196 | -5.04 | 5/5 |
| volume level | -0.242 | -3.71 | 5/5 |
| turnover | +0.195 | 3.51 | 5/5 |
| 14d trailing return | -0.074 | -0.73 | 3/5 |
Pooled quintiles by convexity (n=207 token-origins), Q1 = thinnest: median fwd **+264% / +70% / +36% / +41% / +22%**; share above +100% **71% / 44% / 32% / 29% / 28%**; share below -50% **0% / 7% / 2% / 5% / 12%**.
ROBUSTNESS: partial rho controlling for volume **+0.213, t=2.41, 4/5**; restricted to tokens already above $50k/day at origin **+0.200, t=3.15**. So it is NOT a volume proxy and not only a micro-cap artifact.
WHY TO BELIEVE IT DESPITE SURVIVORSHIP (universe is today's top 100, so levels are fiction): the two features present in BOTH this panel and the survivorship-handled Dune panel replicate closely, volume -0.242 vs -0.179 and drawdown -0.196 vs -0.147. The ordering is the finding, the +264% is not. **Next step that settles it: re-run Amihud on the Dune panel where dead tokens are marked -95%.**
BEN'S ATH RULE CONFIRMED AGAIN here, 5/5 origins, t=-5.04. Third measurement, and the round-3 retraction of it stays retracted.
Deliverable written: `/mnt/user-data/outputs/rh_convexity_rank.csv`, 63 ranked tokens with convexity, % move per $100k, cap, liquidity, 14d volume and % off ATH.

**"THE RIGHT ACCOUNTS" TESTED PER-ACCOUNT AND OUT OF SAMPLE. NO DETECTABLE SKILL.** Design with no lookahead: score accounts on posts made BEFORE 5 Aug against the 5-19 Aug return of the tokens they cited, then test them on entirely different posts (5-19 Aug) against 19 Aug-2 Sep. Spray accounts (>=15 tokens) excluded.
Universe median +118% (n=63). Accounts scoring ABOVE median in training: **+82%** (n=65). Accounts scoring BELOW median: **+109%** (n=101). Tightening to top-quintile scorers gives +106% at a 2-token minimum, +62% at 3, +152% at 4, i.e. noise around the universe.
Follower control on the same test set: **50k+ followers +62%, under 5k +109%, universe +118%.** Big accounts underperform, small accounts merely match. Consistent with the 31 Aug attention test, now measured per-account rather than by follower bucket.
**TWO STRUCTURAL REASONS IT CANNOT WORK IN THIS FORM, and they matter more than the numbers:** (1) only **22% of scored accounts posted in both windows**, so the population turns over faster than it can be scored and even real skill would expire before it was measurable; (2) the scoreable population is 158 accounts, too few to separate skill from luck over two windows.
This refutes hit-rate-on-public-posts as the way to find "the right people". It does NOT refute Ben's underlying claim. **The untested version that is structurally better is OPERATORS AND DEPLOYERS rather than posters: far smaller population, repeats across launches, and fires before a pool exists. Ties to [[operator-graph]]. Now the top priority, alongside re-running convexity on the Dune panel.**

## ROUND 7c, 2 Sep 2026: the three recommended builds, run on the full Dune panel. CONVEXITY RETRACTED.
Panel: 11,556 token-origins, 3,714 tokens, 5 origins 22 Jul to 19 Aug, 14d forward, dead tokens kept at -95% (no trade in the last 4 days of the window). Supply summed as double, decimals joined. Per-origin dead 8.8% to 17.1%, touch +100% 6.2% to 19.2%.

**CONVEXITY (Amihud) DOES NOT RANK EX ANTE. The round-7b +0.303 was survivorship.** Full panel: vs fwd +0.048 t 1.7 4/5; **vs touch +100% -0.080 t -4.3 5/5; vs death +0.099 t 5.5 5/5.** Q1 (most convex) pooled: median -2%, death 23.3%, touch 6.1%; Q5: death 5.8%, touch 18.2%. After every ex ante death filter tried (mcap>$100k, vol14>$20k, age>=14d, combinations) convexity vs return is NEGATIVE, -0.086 to -0.160. Mechanism confirmed only post hoc: among survivors +0.201 t 5.7 5/5. **Convexity is symmetric and flow sign is unknowable, so it is a POSITION-SIZING input only. `rh_convexity_rank.csv` is NOT a buy list.**
Replicated on the full panel: Ben's ATH rule (drawdown -0.103 t -5.5 5/5, fourth confirmation), realised vol negative (t -8.2), age POSITIVE here (+0.147 t 3.9 5/5, young tokens die), mcap vs touch +0.102 (doubling comes from deeper, larger tokens).
**ESTABLISHED FILTER, ex ante, the one usable output:** mcap>$100k AND 14d vol>$20k AND age>=14d -> n=688, death 0.7 to 3.6%, touch +100% 22 to 30%, vs universe 13.4% / 10.5%. It is the death model in three lines. A universe to draw from, not a ranking.

**OPERATOR IDENTITY: weak positive, small n, first thing that separates the tail.** Deployer = tx sender of the creation trace (not the factory), 4,190 tokens (q on Dune id 8568856). Clean split: operators scored on distinct tokens observed before 12 Aug, tested on 622 NEW tokens first seen from 12 Aug. Test base touch 14.3%, death 11.4%. Operators with prior hit-rate >=33% (>=2 prior tokens): **n=21, touch 28.6%, death 0.0%**. Prior hit-rate 0: n=49, touch 4.1%, death 18.4%; at >=3 prior tokens **0 doublings in 30, death 23.3%**. Median return NOT separated (-2% vs +3%); the separation is tail and death. Coverage only 9% of new tokens, grows weekly. NOTE: a naive origin-split (same token in train and test) inflated this to 32.2% vs 15.6%; token-level split is mandatory. Prior Relay extraction alone does NOT predict (death 15.9% vs 13.3%, wrong direction on 19 Aug). Factory/infra deployers (>20 prior launches) are the worst group (mean +4%, death 16.7%); first-time deployers mean +25%.
Scored operators to degree-test before attributing: `0x1e060809fc533333647b51ae57de9ec353909c44` (43% on 7), `0xe6f6030215210024816cd261f94a434d60325f47` (40% on 5, 0% death), `0xcdfc08a1c1fbafb355645e5ddc32122e5716ca90` (33% on 6, 0% death). `0x5516b3451d4d6c9f63353fe7bc9537477ecce000` (52 tokens) is almost certainly infrastructure.

**CONVEXITY DECAY AS AN EXIT RULE: REFUTED as a tape signal.** 451 rally obs (trailing 7d >= +50%) across 51 top-100 tokens; convexity change during the rally vs next 7d rho +0.013 t 0.10, terciles non-monotone. Would need true pool depth, not the Amihud proxy.

Summary shape after 7 rounds: the chain will not say WHICH token; it says which not to touch (death, serial-dead operators) and how hard the kept ones move when flow arrives. All measurable edge is exclusion and sizing; the tail is bought as a basket from the established universe and held.
FINAL WRITE