Worker T3, 2026-09-26 night run. Reads: 00-MANDATE.md, RULES.md, signal-room/PROVIDERS.md, DATA-AUDIT-2026-09-25.md §1–2. Web research was split across three sub-workers whose full notes, with every URL and fetch date, are kept in data/T3-research-A-macro.md (money in/out, firepower, risk appetite, attention), data/T3-research-B-chains.md (chains, terminals, cap bands) and data/T3-research-C-evidence.md (literature and our public-data computation). This file is the condensed, decision-ready version. Grades: A = primary source fetched (URL + date in the research notes) or our own saved script; B = credible secondary; C = unverified or inference.
scripts/T3-trenches.py, A): 40–52 of our watched tokens doubled inside 24 h each day this week, almost all under $1M, none above $50M; down 16% from the week before, and the drop is Robinhood Chain.Condensed from data/T3-research-A-macro.md (40 metrics) and data/T3-research-B-chains.md. Those files carry the full per-metric rows (endpoint, formula, history start, fetch URL and date). Here each metric is one row: question, formula, best source, rights, cadence and cost, and what it is evidence for and NOT for. Rights shorthand: raw = we may show the number as is (with attribution where stated); derived = show only our own computed aggregate, cite the source; restricted = needs a licence or consent before a paying user sees it; unclear = terms not verified. Live values are from probes on 2026-09-26 (grade A unless marked).
| Source | What the terms say | Consequence for us | Grade |
|---|---|---|---|
| DefiLlama free API | Terms of 24 Jun 2025: licence "for personal, non-commercial purposes"; may not "copy, scrape, harvest or otherwise exploit the Content & Data for commercial purposes without prior written consent"; liquidated damages up to $100,000 per violation. Bridges API now returns 402 (paid, $300/mo). https://defillama.com/terms | "DefiLlama is open" is false for a paid product. Use internally as a cross-check; for display either licence it or recompute from chain data (the adapter code publishes the fee wallets and SQL; addresses are public facts) | A |
| CoinGecko (we pay Basic) | Paid plans carry a commercial licence with "Data provided by / Powered by CoinGecko" attribution; no resale or syndication of the API. https://www.coingecko.com/en/api_terms | Our main legal source for market-wide numbers. Attribution on the page | A |
| DexScreener | "may be used for both non-commercial and commercial purposes", but not "to construct, enhance, or market a product or service whose primary purpose is to compete directly with DEX Screener"; no resale. https://docs.dexscreener.com/api/api-terms-and-conditions | Our ticks come from here. A trader companion is arguably not a DEX Screener competitor, but it is close enough to flag for T1 and Thomas | A (text), C (our reading) |
| CoinGlass | No free tier; commercial use only from Standard $299/mo | Skip unless ETF flows or liquidation totals become essential | A |
| Glassnode / CryptoQuant | Glassnode bans commercial publishing without bought redistribution rights; CryptoQuant data under CC BY-NC | Exchange-reserve metrics are internal-only unless we pay for redistribution | A / B |
| Coinalyze | Free API; "if you use the API/data in public places ... cite the data source" | Free aggregated funding, OI, liquidations with a citation | A |
| alternative.me Fear and Greed | "free to use ... in any way you like, this includes commercial projects" | Only fully open mood gauge (but it lags and double counts) | A |
| growthepie | CC BY 4.0, attribution "Source: growthepie"; contact for high-volume commercial | Free, open L2 activity data, includes Robinhood Chain (DAA about 400k/day) | A |
| Exchange public APIs (Binance, Bybit, OKX, Hyperliquid, Deribit, Coinbase) | Binance reportedly bars services that profit from its market data (B); others unchecked | Show derived aggregates (OI-weighted averages, z-scores) with venue citation, not raw ticks | B / C |
| On-chain data via our own RPC | Public ledger | Anything we compute from chain data ourselves is ours to show | C (legal reading) |
The design consequence: the first version of this layer should be built from our own chain tapes, CoinGecko (paid, attributed), exchange public endpoints (derived), growthepie and Coinalyze. DefiLlama is the best free reference and the one we are not allowed to republish.
| Metric | Question | Formula | Best source | Rights | Cadence / cost | Evidence for / NOT for |
|---|---|---|---|---|---|---|
| Stablecoin total supply | How many dollars are parked on-chain? | sum circulating USD-pegged | DefiLlama stablecoins.llama.fi/stablecoins (history from 2017-11); or our own totalSupply() calls on USDT/USDC per chain |
DefiLlama restricted; own calls raw | hourly; free | For: dollar liquidity on crypto rails. NOT: buying intent (much is Tron payments, DeFi collateral). Live: $312.7B |
| Stablecoin 7 d / 30 d change | Are dollars arriving or leaving? | circ_now minus circ 7 d / 30 d ago | same (prev-day/week/month fields precomputed) | as above | daily; free | For: net new dollars (live +$2.70B 7 d, +$3.47B 30 d). NOT: flow into alts; one issuer can rebalance |
| Supply by issuer | Offshore (USDT) vs onshore (USDC) vs synthetic (USDe)? | per-asset deltas | same, per asset | as above | daily; free | For: who is demanding dollars; USDe growth tracks funding carry. Live USDT $183.7B, USDC $76.0B, USDe $4.9B |
| Supply by chain | Which ecosystem is being funded? | chainCirculating deltas | same, per chain | as above | daily; free | For: trading-venue funding (Solana, Hyperliquid, BSC, Robinhood). NOT: Tron USDT ($92.6B) is payments; show it apart |
| Net issuance (mint / burn) | Did an issuer print or redeem? | mints minus burns, counting treasury outflows not treasury mints | USDT/USDC contract events via our RPC | raw (public ledger) | per block; free | For: primary dollar demand. NOT: "authorised" treasury mints ahead of demand |
| Exchange stablecoin reserves | How many dollars sit at exchanges? | sum at labelled exchange wallets | CryptoQuant ($39–109/mo), Glassnode (about $999/mo + API) | restricted (non-commercial / no redistribution) | daily | For: deployable float. NOT: complete (labelled ERC-20 wallets only; Tron and internal ledgers undercounted) |
| US spot ETF net flow, BTC / ETH / SOL | Are US wrapper buyers adding or pulling? | creations minus redemptions, USD | SoSoValue free key (terms unverified), CoinGlass ($299 commercial), Farside (web only, 403 to scripts) | unclear / restricted | daily, final US evening, show T-1 | For: US institutional demand. NOT: new money only (basis trades). SOL spot ETFs are listed: about $1.44B cumulative, 12 straight weeks of inflows to mid Sept (B) |
| Coinbase premium | Are US buyers paying up? | (Coinbase BTC-USD minus Binance BTC-USDT) / Binance | self-computed from both public tickers | derived | 1 min sample, 1 h average; free | For: US-led spot pressure. NOT: pure demand (embeds USDT/USD deviation); ±0.02% is noise |
| Retail door: app-store ranks | Are newcomers downloading? | rank in US top free (Finance) | Apple RSS feed, no key | raw rank numbers | 4x daily; free | For: retail on-ramp intent (live: Robinhood #20, Coinbase #47, Binance.US #84 in Finance). NOT: flows; ads move ranks |
| Total market cap | How big is the pond? | sum of caps | CoinGecko /global |
raw with attribution | 10 min; in our Basic plan | For: price direction. NOT: flows. Live $2.89T, 24 h −2.85% |
| Market cap change vs stablecoin change | Is a rally funded by new dollars? | show 30 d Δ total cap and 30 d Δ stablecoins side by side | the two above | derived | daily; free | For: funded vs unfunded moves. NOT: a multiplier; never show the ratio (explodes near zero) |
| Metric | Question | Formula | Source | Rights | Evidence for / NOT for |
|---|---|---|---|---|---|
| Stablecoin dominance | What share of the pond is cash? | stablecoin cap / total cap (live 10.8%; USDT.D 6.36%) | self-computed from the two above | derived | For: risk-off positioning only when the numerator grows. NOT: when it rises because prices fall (denominator compression). Always show both moves |
| Stablecoin Supply Ratio (SSR) | How big is BTC relative to the cash that could buy it? | BTC cap / stablecoin cap (live about 5.4) | self-computed | derived | For: relative buying power. NOT: independent information: BTC price sits in the numerator, so "low SSR precedes rallies" largely restates "buy after drawdowns" (§2.2). A 2026 CryptoQuant-sourced analysis found a falling SSR driven by capital leaving, not accumulating (B) |
| SSR oscillator | Is SSR extreme versus its own history? | (SSR minus 200 d SMA) / 2 sd | self-computed | derived | For: regime extremes. NOT: timing; claims rest on small-sample chart reading |
| Stablecoins on exchanges, 7 d netflow | Is buying cash arriving at the till? | labelled-wallet balances | CryptoQuant / Glassnode | restricted | CryptoQuant's own August 2025 note: record reserves "have yet to result in sustained spot BTC accumulation" (B) |
| Stablecoin-margined OI / stablecoin supply | How much of the cash is already levered? | aggregated OI / supply | self-sum from exchange OI | derived | Analyst construct, no validated threshold (C) |
| USDe supply change | Is carry demand rising? | Δ USDe | DefiLlama | restricted | For: positive funding persistence. NOT: spot demand |
"Dry powder" verdict: every firepower ratio in use is a price ratio wearing a flow costume, and no vendor or paper has published an out-of-sample test (§2.2). The brief should show the raw numerator and denominator changes and treat the ratios as context.
| Metric | Formula | Source | Rights | Cadence / cost | Evidence for / NOT for |
|---|---|---|---|---|---|
| BTC dominance and 30 d trend | BTC cap / total cap; 30 d change in points; also ex-stables | CoinGecko /global (live 58.3%) |
raw with attribution | 10 min; paid plan | For: rotation toward or away from BTC. NOT: alt strength while stablecoins grow (they sit in the denominator) |
| ETH/BTC, SOL/BTC | price ratio, daily close and 30 d change | exchange public tickers | derived | free | For: first rungs of beta appetite. NOT: broad alt demand (both have their own ETF flows) |
| Altcoin Season Index, Blockchaincenter | share of top 50 (excl. stablecoins and asset-backed: USDT, DAI, WBTC, stETH, cLINK...) beating BTC over 90 d, 0–100; alt season at 75 or more, Bitcoin season at 25 or less; also 30 d "Month" and 365 d "Year" variants | web only (live 61, Month 78, Year 39, "not alt season") | unclear, cite; self-compute our own | daily; free | For: breadth of alt outperformance. NOT: timing; lags by construction (90 d) |
| Altcoin Season Index, CoinMarketCap | same rule on the top 100 CMC coins, 90 d, 75 / 25 thresholds | web chart | unclear | daily | Same; broader universe, noisier |
| Altcoin Season Index, CoinGlass | method not published | API from $79/mo | restricted | 5 min | Black box: do not show |
| Aggregated funding (annualised) | OI-weighted mean funding x periods per year | Binance, Bybit, OKX, Hyperliquid public (no key); Coinalyze free key | derived + citation | 5 min; free | For: leverage crowding. Live: Binance BTC about +3.7%/yr, mild. NOT: direction; flips within hours |
| Aggregated open interest and 7 d change | sum OI x mark | same | derived | 5 min; free | For: leverage build-up. NOT: direction; inflated by basis trades |
| Liquidations, 24 h long vs short | sum forced fills by side | OKX REST, Binance and Bybit websockets, Coinalyze | derived | stream; free | For: forced flow, squeezes. NOT: complete (Binance pushes at most one per symbol per second) |
| Long/short account ratio | long accounts / short accounts | Binance futures data | derived | 5 min; free | Accounts not dollars; contrarian use is folklore |
| Perp premium | (mark minus index) / index | Binance premiumIndex, Hyperliquid | derived | live; free | Instantaneous leveraged pressure (live −0.04%) |
| Annualised futures basis | (future / spot − 1) x 365 / days | Deribit public (live about 5.0% annualised Dec 26 and Mar 27); CME needs a vendor | derived | hourly; free for Deribit | For: institutional appetite for levered long. NOT: bullishness alone |
| Fear and Greed | vendor composite | alternative.me (commercial use allowed) | raw | daily; free | Built from the metrics above, so it double counts; read 71 "Greed" on a −2.85% day, so it lags |
| Metric | Formula | Source | Rights | Cadence / cost | Evidence for / NOT for |
|---|---|---|---|---|---|
| Chain DEX volume 24 h, 7 d, 7 d-over-7 d | sum daily DEX volume | DefiLlama /overview/dexs/{chain} (internal); GeckoTerminal per-network pools; our tapes |
DefiLlama restricted; GeckoTerminal attribution; ours raw | hourly; free | For: where trading is now. Live 7 d: Solana $20.8B (+19.5%), Ethereum $11.8B (+23%), Robinhood $10.5B (−18.8%), BSC $7.6B (+4.4%), Base $6.9B (+10.7%), Hyperliquid L1 $3.1B (+42%); global $77.4B (+10%). NOT: a forecast (§2.3). DefiLlama's Dune-backed adapters miss the latest day for about 10 h |
| Chain TVL and 7 d change | TVL_t / TVL_t−7d − 1 | DefiLlama /v2/chains, /historicalChainTvl |
restricted | daily; free | For: parked capital. NOT: trading activity (Robinhood: $1.0B TVL, #9, but #3 by volume) |
| Bridge net flow by chain | deposits in minus withdrawals out | DefiLlama bridges (now 402, $300/mo), Artemis (paid), or our own canonical-bridge and Relay/Across event readers | restricted / ours | daily | For: new buying power arriving on a chain. NOT: leading; academic panel finds bridge flows chase recent native-token returns (§2.3) |
| Chain fees | sum fees paid | DefiLlama fees (internal), growthepie for L2s | restricted / CC BY | daily; free | For: real demand vs incentivised TVL. Robinhood fees 7 d $47.8M |
| Active addresses | distinct signers per day | growthepie (CC BY, includes Robinhood: about 400k DAA, T+2 lag), our RPC for Solana/BSC; Artemis ($100–500/mo), Token Terminal (Pro $350/mo) | CC BY / ours / restricted | daily | For: breadth of participants. NOT: people (bots, sybils) |
| Launchpad volume, launches, graduations | per-launchpad daily volume and counts | DefiLlama launchpad category (internal); our own launchpad event readers (pump.fun, Pons, four.meme, Flap portals are published in the adapters) | restricted / ours | hourly | Live 7 d: pump.fun $761M (+34%), StonkFun $334M, Pons V2 (Robinhood) $304M (−47%), LaunchLab (Raydium, bonk.fun) $152M, Flap.sh $124M, Meteora DBC $120M, four.meme $75M, Genius.fun $43M. For: new-token supply and appetite. NOT: leading (§2.4) |
| Trading-terminal volume (degen activity index) | terminal + bot volume, and its share of chain DEX volume | DefiLlama "Trading App" / "Telegram Bot" categories (internal); Dune fee-account SQL; our own fee-account watchers | restricted / derived-ours | daily (Dune) or 5 min (own) | Live 7 d: fomo Wallet $973M, Axiom $803M (Solana $664M, Robinhood $118M, BSC $21M), GMGN $665M (Robinhood $247M, BSC $229M, Solana $147M), Terminal (ex-Padre, pump.fun) $112M, Trojan $25M, Photon $6.5M, BONKbot $4.9M, BullX $3.4M, Banana Gun $3.2M. Share of chain DEX volume: Solana about 9.3–9.8%, Robinhood about 3.5%, BSC about 3.3% (Binance Wallet unmeasured). Photon and BullX figures look like adapter undercounts after fee-wallet rotation (C) |
| Category heat (chain meme sectors) | category cap 24 h change, volume / cap | CoinGecko /coins/categories (1 credit) |
raw with attribution | 5 min; in plan | Live: all memes $36.2B (+1.2%); Solana memes $4.0B (vol $0.93B); robinhood-chain-meme $0.9B (+6.2%); four-meme ecosystem $0.8B; Base memes $0.3B |
How terminal volume is measured on-chain, and how we would do it ourselves. DefiLlama's adapters (source at github.com/DefiLlama/dimension-adapters) measure terminals by fee receivers, not routers, because routers rotate: the GMGN adapter comments that "a hardcoded router would have dropped 99% of that day's volume". The method: take transactions that pay a known terminal fee account, join to DEX trades by transaction id, take the last hop per transaction to avoid double counting multi-hop routes. Published addresses include Axiom's router program FLASHX8DrLbgeR8FcfNV1F5krxYcYMUdBkrP1EPBtxB9 and fee wallets (7LCZckF6..., 4V65jvcD..., CeA3sPZf... and ten more), GMGN EVM fee collectors 0xb8159ba3... and 0xe3ed4c49... plus Solana fee wallets, Photon AVUCZyuT..., Trojan 9yMwSPk9... plus five, BullX 9RYJ3qr5..., Maestro EVM 0xB0999731..., Terminal (ex-Padre) on Solana, ETH, BSC and Base; full list in data/T3-research-B-chains.md §3. Dune dashboards: adam_tehc/axiom, adam_tehc/gmgn, adam_tehc/photon, adam_tehc/trading-bots-on-solana, whale_hunter/dex-trading-bot-wars, whale_hunter/sol-trading-bot, whale_hunter/trojan-trading-bot, kucoinventures/meme-trading-platform-war-on-solana, tk-research/telegram-bots; for Binance Wallet / Alpha on BSC: lifi/binance-wallet-swaps, night_walker/binance-alpha-20-overview, hicrypto/binance-alpha-overview. Our own version: watch about 40 fee accounts (getSignaturesForAddress on Solana, Transfer-to-collector eth_getLogs on BSC and Robinhood), value with our router-price library, and run a daily rotation check (a terminal's fee-account transaction count falling more than 50% while the chain is steady means a new fee wallet to find). Cost: RPC calls only; the maintenance of rotating fee wallets is the real cost.
Robinhood Chain, the equivalents. Arbitrum Orbit rollup, public mainnet 2026-07-01. Terminals present: GMGN (from 2026-07-03), Axiom (from about 2026-07-11), Maestro, fomo / copyfomo, o1.exchange. Launchpads: Pons V2 (dominant, Pons Labs, not a Robinhood product), Pons V1, Flap.sh (portal 0x26605f32...), LetsCash, Frontier, StonkBrokers, BaseStonk, o1 Launchpad, clanker, Bags. DEX mix 7 d: Uniswap V3 $4.36B and V4 $3.51B dominate. Coverage: DefiLlama, GeckoTerminal (network robinhood, 42 DEX ids), DexScreener, Dune (raw EVM tables only, no dex.trades), Bitquery, growthepie. Terminal share of Robinhood volume is low (about 3.5%) because much flow is Pons-native and plain Uniswap.
Band boundaries today (CoinGecko /coins/markets, our Basic key): rank 250 is about $127M, rank 494 about $50M, rank 1,000 about $17M, rank 1,210 about $11M, rank 2,871 about $1.3M; 21,589 active coins, total 24 h volume $109.8B. Bands must be cut on market cap, not rank (rank is null on 6–60 rows per page). CoinGecko is thin below $1M: the real trenches (fresh pump.fun, Pons, four.meme tokens between $5k and $1M) are mostly not listed there, and CoinGecko's on-chain pool lists stop at 200 pools per network (the 200th Solana pool already has a $226M FDV); the cross-network pool filter needs the Analyst plan ($129/mo).
| Metric | Formula | Source | Cadence / cost | Evidence for / NOT for |
|---|---|---|---|---|
| Band volume share | band 24 h volume / all volume; freeze band membership at 00:00 UTC | CoinGecko /coins/markets pages 1–40 (top 10, 11–50, 51–200, 201–1000, under $50M, under $10M) |
40 credits per sweep; hourly sweep about 29k credits/month, fits our 100k Basic | For: where in the stack money trades today |
| Band turnover (band heat) | band volume / band market cap | same above $1M; our tapes and ticks below $1M (§3.3) | same | For: money in relative to size, the mandate's "hottest band". Our week: sub-$1M 0.7–2.2x a day, $1–10M 0.16–1.0x, $10–50M 0.02–0.08x |
| Band heat z-score | (share or turnover − 30 d mean) / 30 d sd | own store | daily; free | For: "hotter than usual" labels. Needs 30 d of own snapshots |
| Runners by band | tokens with a 2x inside 24 h, per band | our ticks now (§3.1); market-wide needs new-pool sweeps | 5 min; free on ours | For: the trenches trader's real question ("are things running?") |
| Risk-down-the-curve spread | equal-weight 201–1000 return minus top-10 cap-weighted return | CoinGecko | daily | For: appetite for smaller caps. NOT: leading (§2.5, the seesaw effect) |
| Breadth by band | share of band up on the day | CoinGecko price_change_percentage_24h |
daily | For: many or few |
| Published comparators | CoinDesk 80 (top 100 outside CD20), CoinDesk Memecoin (CDMEME, 50 most liquid memes, equal weight), MarketVector Large / Mid / Small, Kaiko Large / Mid / Small (dual cap and liquidity score), Bitwise 10, CoinGecko categories | licensed indices; we compute our own equivalents from CoinGecko | Nobody publishes a sub-$50M, sub-$10M or sub-$1M heat index; that gap is ours to fill |
X remains the main venue for crypto price talk and caller networks; Telegram carries caller channels and bot-driven meme trading; Farcaster is small (Neynar took over operations in Jan 2026, B); Google search is the only outside-the-bubble gauge. Published indices: Kaito mindshare (enterprise, about $10k/yr+, gamed by incentive campaigns), LunarCrush ($90–900/mo, social volume, coincident with price), Santiment social volume (Pro $49/mo real time, one of few covering Telegram; spikes mark tops as often as starts), Google Trends (official API is application-gated alpha), X API (pay per use about $5 per 1,000 posts read). For this layer the only attention number worth one line is "where the talk is" handed over from T4; the volume already tells us where the money followed.
Full literature tables, per-paper quotes and fetch URLs: data/T3-research-C-evidence.md. Our own computation: scripts/T3-evidence-stables.py (run from this directory: python3 scripts/T3-evidence-stables.py, cached data in data/T3-ev-*.json, add --refresh to re-download; outputs data/T3-ev-stables-results.csv, T3-ev-activity-results.csv, T3-ev-weekly-xcorr.csv, T3-ev-rotation.csv). Inputs are free, no-key public series: DefiLlama stablecoin supply (2017 to 2026-09-24), DefiLlama total, Solana and pump.fun DEX volume, Binance daily klines for BTC, ETH, SOL, DOGE. I reran it at the end of this session and it reproduces.
Short version: none of the five flow metrics has credible evidence of leading the next weeks of alt or meme activity. They move with price the same week, and several follow it. The only leading structure with academic support is short-horizon price momentum inside a size band, which is a property of prices, not of flows. That changes how the page must be framed (§4).
| # | Metric | Verdict | Grade | Strongest evidence |
|---|---|---|---|---|
| 1 | Money in / out (stablecoin issuance and supply growth; spot ETF flows) | Reverse plus contemporaneous: flows follow price | A (our data) / B (literature) | Wei 2018 (Economics Letters): Tether issuance does not move later BTC returns, only volume for up to 5 days, and grants follow downturns (A). Ante, Fiedler, Strehle 2021 and Kristoufek 2021: issuance follows price declines / reflects demand (B, abstracts). Griffin and Shams 2020 (J. Finance) found a lead, but for one actor in 2017 at hourly scale (A). ETF flows: about +53 bps same-day per $100M, returns strongly cause later flows, a small next-day effect (Lim 2026, B; FalconX 2024, A). Ours, post-2023: 30 d stablecoin growth versus next-30 d BTC, ETH, SOL, DOGE, total DEX volume, Solana DEX volume, pump.fun volume: lead r −0.21 to 0.00, none significant; same-period r +0.14 to +0.41; reverse (price or volume first, then supply) up to +0.51 for DOGE (p 0.0003) and +0.54 for pump.fun (p 0.004). The 2020–2026 sample shows a weak positive lead (r 0.1–0.3), carried by the 2020–21 run. |
| 2 | Stablecoin firepower (SSR, exchange stablecoin reserves) | Unclear, mechanically tied to price | C | Glassnode's 2019 piece introducing SSR has no backtest (A). CryptoQuant asserts buying power; its own Aug 2025 note says record reserves had not produced sustained spot accumulation (B). No peer-reviewed out-of-sample test found (C, two targeted searches). SSR has BTC price in the numerator, so "low SSR precedes rallies" is mostly "buy after drawdowns". We did not backtest it. |
| 3 | Chain flows (bridge net inflow, TVL change, DEX volume by chain) | Contemporaneous; price leads by about a week | A | Cao, Cong et al. 2026 (arXiv 2604.03083, 20 chains, 2022–2025): bridge net inflows are "return-chasing"; more interoperability is followed by lower native-token returns at 30 and 100 days (A). Ours, weekly, post-2023 (195 weeks): Solana DEX volume vs SOL: same week r +0.40 (p < 0.001), activity one week ahead +0.04 (p 0.56). Total DEX volume vs DOGE: price one week ahead of volume +0.21 (p 0.003), volume one week ahead −0.03. At 30 d, a Solana volume surge was followed by weaker SOL (r −0.41, p 0.03, n 28, 2024-06 onward), consistent with blow-off. Solana 2024, Base 2023 (BALD), BSC four.meme Oct 2025 are co-movement stories with the trigger (memes, a CZ post, the BTC rally) first (B). |
| 4 | Terminal and launchpad volume (degen activity) | Contemporaneous; no lead at 1–4 weeks | A for weekly / monthly; C for intraday (untested) | No published lead test exists; Dune dashboards and papers on pump.fun are descriptive or token-level (A/B). Ours: pump.fun weekly volume change vs SOL: same week +0.30 (p 0.001), one week ahead +0.01 (p 0.92); vs DOGE same week +0.32, one week ahead −0.07 (121 weeks). pump.fun growth does not even persist into its own next month (r +0.05). Closest academic analogue: attention proxies forecast returns at weekly horizons (Liu and Tsyvinski 2021, RFS, A), which leaves room for a short daily or hourly lead we have not tested. |
| 5 | Cap-band heat (size, momentum, rotation) | Leading, but cross-sectional and short, and it reverses | A for the factor papers; C for transfer to 2026 memes | Liu, Tsyvinski, Wu 2022 (J. Finance, coins over $1M, 2014–2018): small-minus-big 3.4%/week; 1–4 week momentum 2.5–4.1%/week (A). Dobrynskaya (2,000 coins, 2014–2020): momentum up to 2–4 weeks, then reversal; in small illiquid coins even 1-week momentum reverses (A/B). Jia et al. 2023 "seesaw": large-coin gains predict small-coin losses next period (B). Alt-season index lags by construction; its own promoters call it regime confirmation (B). Ours: the BTC then ETH then small-cap rotation has weak support in 2020–2026 (ETH-minus-BTC past vs DOGE-minus-BTC next r +0.26, p 0.02, n 81) and none after 2023 (all rotation r between −0.08 and +0.15, p > 0.26, n 45). |
Caveats on our computation, stated once: small non-overlapping samples (27 to 81 windows at 30 d, 121 to 195 weeks), one or two market cycles, DOGE and SOL stand in for memes (they are large caps), DefiLlama volume coverage grows over time as adapters are added, and about 70 correlations were computed without correction (expect 3 or 4 "significant" by chance). Nothing was tested below weekly resolution.
What this means for the brief. - Label the macro rows (stablecoins, ETF flows, dominance, chain flows, terminal volume) as the weather now, never as a forecast. No arrow on them should imply "next". - The only defensible "leading" content in Layer 1 is momentum within a band at short horizons, which is our runners count and band turnover (§3), and even that decays within weeks and reverses in the smallest names. The honest phrasing is "heat persists for a while, then fades", and we should measure how long on our own tape once we have a month (not done tonight, see §6). - A regime frame is still useful for a trenches trader: when stablecoins, chain volume and runners are all rising together the pond is warm; when they diverge, say so. That is description, and it is enough.
Script: scripts/T3-trenches.py. Command: python3 projects/caverio/companion-2026-09-26/scripts/T3-trenches.py from the workspace root (optional arg: end day, default 2026-09-25). Output tables: data/T3-trenches-output.md; machine output data/T3-trenches.json. Runtime 20–25 s, read-only plain-file reads of signal-room/data/{ticks,candles-bsc,candles-onchain,trades-bsc}. All numbers in this section are grade A (own store, saved script). Window: 7 UTC days, 2026-09-19 to 2026-09-25 (09-25 runs to about 21:40Z, so it is short by about two hours).
Universe warning, read before any number below. These are the tokens our discovery chose to watch (about 180–215 per day with ticks, across Solana, Robinhood Chain and BSC), not the market. The brief can show them as "the trenches we watch", not as "the market". A market-wide version of the same counts needs a pool sweep (GeckoTerminal / DexScreener new-pairs or a chain-wide swap tape), costed in §5.
Definition: token counts on day D if a print at time t2 in D is at least 2x a print at t1 in the preceding 24 h (low before high). Band = dexscreener mcap at the low print. Before the test, lone prints below 0.75x of both neighbours and lone prints above 1.33x of both neighbours are dropped (the audit's phantom test). "Held" additionally requires the next print after the high to be at least 1.8x the low. Brackets: tokens observed in that band that day (band by the day's median mcap), for scale.
| Day (UTC) | < $1M | $1–10M | $10–50M | ≥ $50M | Total 2x | Without phantom filter | Held 2x | Solana / RH / BSC |
|---|---|---|---|---|---|---|---|---|
| 09-19 | 33 (87) | 7 (90) | 1 (18) | 0 (8) | 42 | 42 | 42 | 24 / 14 / 4 |
| 09-20 | 32 (77) | 10 (107) | 0 (11) | 0 (10) | 42 | 44 | 42 | 23 / 17 / 2 |
| 09-21 | 29 (75) | 21 (115) | 1 (20) | 0 (13) | 51 | 52 | 50 | 32 / 16 / 3 |
| 09-22 | 33 (86) | 10 (99) | 3 (24) | 0 (16) | 46 | 55 | 45 | 28 / 16 / 2 |
| 09-23 | 28 (72) | 11 (108) | 1 (23) | 0 (17) | 40 | 45 | 37 | 22 / 14 / 4 |
| 09-24 | 40 (72) | 10 (101) | 2 (25) | 0 (17) | 52 | 56 | 48 | 31 / 17 / 4 |
| 09-25 | 35 (66) | 9 (108) | 2 (25) | 0 (18) | 46 | 48 | 44 | 24 / 18 / 4 |
What it says:
- Under $1M is where the doubling happens: 28–40 tokens a day, against 7–21 in $1–10M and 0–3 in $10–50M. Nothing at or above $50M doubled in any day of the week. Against the tokens observed in each band (an approximate rate, since the numerator bands at the low and the denominator at the day's median), 35–55% of the sub-$1M names we watch double inside a day, 8–18% of $1–10M names, 0–13% of $10–50M names.
- Week on week, the trenches we watch cooled. Same script, previous week (T3-trenches.py 2026-09-18, output data/T3-trenches-output-prevweek.md): 09-12 to 09-18 had 40 to 72 tokens doubling a day, mean 54.4, against mean 45.6 this week (down 16%). The drop is almost all Robinhood Chain: 20–36 doublers a day the week before, 14–18 this week; Solana held at 14–32 a day. This is exactly the kind of sentence the brief should print ("fewer runners than last week, and the drop is Robinhood"), with the universe caveat attached.
- Within this week the count is steady, 40 to 52 a day. There is no visible "hot day" in this week; 09-21 and 09-24 are the high days. A week is too short to call a trend, which is the honest reason the brief needs a month of this series before it shows a "hotter / colder than usual" label.
- Phantom prints matter at exactly this metric: without the filter the count is up to 9 tokens a day higher (09-22: 55 versus 46, a 20% overstatement). 138 lone prints were dropped in the window. Any 2x counter we ship must run the filter.
- The low print is often the discovery move itself (we start watching young tokens), so "2x in 24 h" on our universe partly measures the discovery lane's timing, not only the market's temperature. The market-wide version does not have this bias.
BSC from trades-bsc (deduped on txHash + logIndex, 8,992 duplicate rows removed; WBNB-quoted pools valued at the nearest BNB/USD seen on valued rows, stable-quoted at 1). Robinhood Chain from candles-onchain 1 m volumeUsd.
| Day | BSC pools | BSC volume | BSC top-5 share | RH tokens | RH volume | RH top-5 share |
|---|---|---|---|---|---|---|
| 09-19 | 12 | $5.51M | 71% | 2 | $0.10M | 100% |
| 09-20 | 14 | $12.56M | 92% | 3 | $0.43M | 100% |
| 09-21 | 12 | $5.99M | 92% | 4 | $0.55M | 100% |
| 09-22 | 13 | $6.14M | 92% | 1 | $0.01M | 100% |
| 09-23 | 13 | $4.10M | 88% | 466 | $8.89M | 41% |
| 09-24 | 13 | $1.97M | 87% | 520 | $61.50M | 22% |
| 09-25 | 6 | $1.60M | 100% | 519 | $59.37M | 31% |
Two data facts the build must know: - Robinhood own-tape volume is only comparable from 09-24. 466 of 608 candle files start on 09-23 (the tape's watched set was widened that day), so the jump from $0.5M to $60M is coverage, not market. Read 09-24 and 09-25 only: about $60M a day across about 520 Robinhood tokens, top 5 hold 22–31%. - 47% of BSC swap rows in the window carry no USD valuation (about 350k rows). They are pools quoted in tokenized stocks and other non-BNB assets (BABAB, FXIon, QQQB, oANTHROPIC, TSMB, AAPLB, GOOGLB, BTCB, XAUt and others), which are not trenches pools; they are left out, not guessed. The BSC meme volume we watch is small and concentrated: $2–12M a day in 12–14 pools with the top 5 at 87–92%.
Per token-hour, last vol1hUsd seen in that hour, summed per day; band by the day's median mcap; turnover = band volume / band summed mcap. One token-day was excluded as misreported: a Robinhood token on 09-22 with $1.03B "volume" on $0.2M liquidity and $0.46M mcap (5,000x its liquidity). The filter is volume > 500x max liquidity. This view undercounts: the median token has only 11–18 of 24 hours polled, so read it as relative, not absolute.
| Day | < $1M vol (turnover, n) | $1–10M | $10–50M | ≥ $50M | Top-5 share, under $50M | Under-$50M volume Solana / BSC / RH |
|---|---|---|---|---|---|---|
| 09-19 | $48M (1.41x, 88) | $268M (0.99x, 88) | $7M (0.02x, 19) | $152M (0.05x, 8) | 70% | $267M / $11M / $46M |
| 09-20 | $25M (0.83x, 78) | $157M (0.45x, 105) | $8M (0.03x, 12) | $315M (0.11x, 10) | 65% | $142M / $23M / $25M |
| 09-21 | $74M (2.16x, 76) | $103M (0.30x, 114) | $31M (0.08x, 20) | $510M (0.22x, 12) | 33% | $134M / $15M / $59M |
| 09-22 | $57M (1.64x, 83) | $71M (0.22x, 101) | $31M (0.07x, 24) | $266M (0.02x, 16) | 30% | $75M / $17M / $66M |
| 09-23 | $43M (1.06x, 75) | $54M (0.16x, 105) | $30M (0.06x, 23) | $476M (0.11x, 17) | 22% | $65M / $14M / $47M |
| 09-24 | $47M (1.29x, 75) | $159M (0.53x, 98) | $17M (0.03x, 25) | $113M (0.03x, 17) | 57% | $166M / $20M / $36M |
| 09-25 | $24M (0.68x, 69) | $86M (0.24x, 105) | $29M (0.06x, 25) | $122M (0.01x, 18) | 33% | $88M / $19M / $31M |
What it says: - Turnover is the band-heat number, and it separates the bands by an order of magnitude: sub-$1M names turn over 0.7–2.2x their market cap per day, $1–10M names 0.16–1.0x, $10–50M names 0.02–0.08x. Turnover is the right cap-band heat metric (money in relative to size); raw volume is dominated by a few names (top-5 share 22–70%). - Solana dominates the watched under-$50M volume most days ($65–267M against BSC $11–23M and Robinhood $25–66M). Robinhood was the hottest of the three on 09-21 to 09-23 by turnover-weighted activity, then cooled. - Concentration swings: on 09-19 and 09-20 five tokens were 65–70% of the under-$50M volume (one or two runners carried the day); on 09-23 it was 22% (broad day). This "breadth" number (top-5 share) is cheap and tells a trader whether the day is one runner or many; it belongs on the page.
Principle from §2: almost everything in this layer is descriptive (moves with price the same day or follows it). The page must say so in its own structure: it is the weather report of the pond you trade in, not a forecast. The only items with any claim to leading information at trenches speed are our own tape items (runners count, breadth, band turnover, launchpad graduation pace) and those are leading only in the weak sense that heat persists for hours to days. Each row below is marked L (some leading value, stated how weak), D (descriptive, same-time), or C (confirming, tends to lag).
Order is top to bottom. Views: T = today (rolling 24 h), W = 7 d, M = 30 d.
| # | Number | Form (suggestion for T7/T8) | Views | Type | Source (ours or third party) |
|---|---|---|---|---|---|
| 1 | Runners today: tokens that doubled inside 24 h, split by band (< $1M / $1–10M / $10–50M), versus the same day-count's 7 d and 30 d average | one big number + three small band numbers + "vs usual" arrow | T, W, M | L (weak: heat persists for hours; §2.5) | ours (T3-trenches.py logic, phantom-filtered) |
| 2 | Band heat: turnover (volume / mcap) per band, ranked, with the hottest band named in a sentence; top 10 / 11–50 / 51–200 / 201–1000 from CoinGecko, under $50M / $10M / $1M from our tapes and ticks | ranked bar or seven small gauges | T, W, M | L (weak) / D | ours below $50M; CoinGecko /coins/markets sweep above |
| 3 | Breadth: share of trenches volume in the top 5 tokens ("one runner or many") | single percent + sentence | T, W | D | ours |
| 4 | Chain heat: DEX volume 24 h and 7 d-over-7 d by chain (Solana, Robinhood, BSC, Base, Ethereum), plus our watched-token volume per chain | ranked list with change arrows | T, W, M | D | DefiLlama (internal only, see rights) or own collectors + GeckoTerminal |
| 5 | Launchpad pace: pump.fun / StonkFun / LaunchLab / Pons / four.meme / Flap volume, launches and graduations per day | ranked list + sparkline | T, W, M | D (C at 30 d) | DefiLlama internal, then our own launchpad log readers |
| 6 | Degen activity index: terminal + bot volume (Axiom, GMGN, fomo, Terminal, Trojan, Photon, BullX, Maestro) and its share of chain DEX volume | one index number per chain + league table | T, W, M | D | Dune (fee-account SQL) or our own fee-account watchers |
| 7 | Dollars in or out: stablecoin supply 7 d and 30 d change, split by issuer and by trading chain (Ethereum, Solana, BSC, Hyperliquid, Robinhood; Tron shown apart as payments) | two deltas + by-chain bars | W, M | C (§2.1: follows price after 2023) | DefiLlama stablecoins (rights) or direct on-chain supply calls (free, ours) |
| 8 | Risk appetite strip: BTC dominance and 30 d trend, ETH/BTC, SOL/BTC, an alt-season reading (our own computed, published formula) | strip of four small numbers with trend arrows | T, W, M | D | CoinGecko /global (attribution), exchange klines (free) |
| 9 | Leverage strip: aggregate funding (annualised), open interest 7 d change, 24 h liquidations long vs short | three numbers, colour only on extremes | T, W | D; extremes mildly contrarian (§2) | exchange public APIs (Binance, Bybit, OKX, Hyperliquid), Coinalyze free with citation |
| 10 | ETF flows: BTC, ETH, SOL US spot net flow, yesterday and 5 d sum | three numbers | W, M | C | SoSoValue free key or CoinGlass (commercial from $299) |
| 11 | Total market cap vs stablecoin change: both 30 d deltas side by side, never as a ratio | two numbers | M | D | CoinGecko + stablecoins |
| 12 | Retail door: Coinbase / Robinhood / Binance app-store ranks | three ranks | W, M | D | Apple RSS (free) |
| 13 | Attention: where the talk is (hand-off to T4's attention layer), one line | sentence | T | D | T4 |
Ship first (eight): 1 Runners today, 2 Band heat, 3 Breadth, 4 Chain heat, 5 Launchpad pace, 7 Dollars in or out, 8 Risk appetite strip, 9 Leverage strip.
Why these eight: 1–3 are ours, computed from stores that already exist, free to run, and nobody else shows them for this universe; they are what a trenches trader actually acts on. 4 and 5 answer "which chain and which launchpad are hot" which the mandate asks for directly and cost nothing to fetch. 7–9 give the macro frame in one line each, cheaply, from free sources; they are context, not signals, and the page should label them "background". ETF flows, app ranks and the degen index wait: ETF flows and app ranks are slow and weak for trenches, and the degen index needs a Dune plan or our own fee-account watchers plus maintenance for rotating fee wallets.
Time views: every ours-computed number (1–3) supports T, W, M from day one only if we backfill; our ticks start 2026-09-10, so the M view is honest from about 2026-10-10. Until then show T and W and label M "building".
Estimates are marked (est). Prices are from the research notes (grades there).
| Block | Source | Cadence | Requests/day (est) | Money/month | CPU/RAM/disk (est) | Rights status |
|---|---|---|---|---|---|---|
| Runners, band heat, breadth (1–3) | our ticks (dexscreener polls, already paid for in bandwidth) + own tapes | every 5 min recompute | 0 new external | $0 | script runs 20–25 s over 7 d of files; incremental version under 1 s per run; under 50 MB RAM; output under 1 MB/day | ours, raw display fine |
| Band heat above $50M | CoinGecko /coins/markets pages 1–40 |
hourly | 40 credits per sweep, about 29k credits/month | $0 extra (fits our 100k Basic) | negligible | raw with CoinGecko attribution |
| Market-wide runners (future) | GeckoTerminal new-pools / DexScreener pairs sweep, or own chain tapes | 5–15 min | 3–10k | $0 on public limits, or CoinGecko Analyst $129 for on-chain endpoints | small | GeckoTerminal / CoinGecko attribution; DexScreener terms unclear (T1) |
| Chain heat (4), launchpad pace (5) | DefiLlama free API (internal cross-check) | hourly | about 50 | $0 free; $300 Pro for bridges | negligible | DefiLlama terms forbid commercial republishing without consent: display requires a deal or our own recomputation |
| Chain heat, own recompute | GeckoTerminal per-network pools, own BSC / RH tapes | hourly | about 500 | $0 | small | GeckoTerminal attribution |
| Degen index (6) | Dune fee-account queries | daily | 30 query runs | Dune Plus about $399 (B) | none local | derived from public chain data, our SQL (check Dune ToS) |
| Degen index, own watchers | about 40 fee accounts via RPC | 5 min | about 12k RPC calls | $0 on free RPC (Helius free 1M credits/mo is enough at 1 credit per call, est) | small | ours |
| Stablecoins (7) | DefiLlama stablecoins, or totalSupply() calls on USDT/USDC per chain |
hourly | 24–300 | $0 | negligible | DefiLlama rights as above; on-chain supply is public fact, ours |
| Risk appetite (8) | CoinGecko /global (Basic key we pay), Binance / Coinbase klines | 10 min | about 300 | $0 extra (existing CoinGecko Basic, 100k credits/month) | negligible | "Powered by CoinGecko" attribution |
| Leverage (9) | Binance fapi, Bybit, OKX, Hyperliquid public; Coinalyze free key | 5–15 min | about 1–2k | $0 | negligible | exchange terms unclear or restrictive for resale: show derived aggregates, cite; Coinalyze free with citation |
| ETF flows (10) | SoSoValue free key; CoinGlass Standard for commercial | daily | 3–10 | $0 or $299 | negligible | SoSoValue terms unverified; CoinGlass commercial only from $299 |
| App ranks (12) | Apple RSS | 4x daily | 4 | $0 | negligible | rank numbers fine |
| LLM | none required for this layer; the daily one-paragraph "state of the pond" sentence can be templated | daily | 0 | $0 |
One rights caveat on our own rows: the ticks behind 1–3 are dexscreener polls, and DexScreener's terms bar use in a product whose primary purpose is to compete directly with DEX Screener (§1.0). The derived counts are ours; the raw prices are theirs. Moving the price basis for 1–3 to our own chain tapes (BSC and Robinhood already exist; Solana would need the tape restarted) removes the question.
Total for the recommended first eight: $0/month in new spend, provided chain heat and launchpad pace are recomputed from GeckoTerminal and our own tapes, or DefiLlama agrees to a licence. Adding bridges and exchange reserves costs $300 (DefiLlama Pro) plus $109 (CryptoQuant Professional, non-commercial licence, so internal only). Adding a clean degen index via Dune is about $399.
Status: DONE