Caverio trader companion, overnight programme 2026-09-26, MASTER

MASTER: Caverio trader companion, overnight synthesis (T10)

Worker T10, 2026-09-26, started 01:00 Beirut. Inputs: 00-MANDATE.md, RULES.md, the nine phase-1 reports T1 to T9 (all read in full, all end Status: DONE), wireframes/, and the 09-25 wiki synthesis wiki/syntheses/2026-09-25--what-caverio-can-see-on-one-token-and-which-of-it-is-ours-to-publish.md. No new research and no new computation: every number below is quoted from a report and cited as T<n> §<section>. Where a number is not in any report, this document says so. Where I add my own estimate (mostly front-end build days, which no worker estimated), it is marked T10 estimate, C.

Everything built before tonight is experimentation (Thomas, 00:28). This document plans the build.


1. One-page summary

1.1 What the product is (ten lines, from the mandate)

  1. A trader companion: one view per token that replaces the five apps a meme trader opens to decide buy, no-buy, exit or size.
  2. It gathers what is hard and costly for one person to gather, and shows it. It does not judge.
  3. Layer 1, the market brief: is money coming in or leaving, where is it going by chain and platform, which market-cap band is hottest, today / week / month.
  4. Layer 2, discovery into situations: tokens named by solid callers, X accounts, tracked wallets and Fomo-profiled traders enter a look list, not a buy list.
  5. Every discovered token opens a situation: a token plus a clock, lighting up by category as more happens, and rising on the list.
  6. The situation room shows everything on that token on one clock: market numbers, money flow, holders, attention, callers and their stance, safety, our own record.
  7. The hold stage matters as much as the entry: the room must make it obvious when the reasons for being in are going away.
  8. Layer 3, strategy zones, comes later: descriptive buy / hold / sell windows with a forgiving, honestly measured definition of a win.
  9. Presentation is the proprietary part: which fields, in which form, on which clock.
  10. It launches through a feedback-for-access beta: a free month for useful feedback, extended for more, dropped for none.

1.2 The five most important things the night established

  1. The chain tape has to be the product's only displayed price, and today it is broken on two of three chains. DexScreener's terms forbid a directly competing product and making "any portion" of the API available to third parties (T1 §0.1, §2.5, grade A). Every priced row since 09-23 is DexScreener (T1 §0.1). Meanwhile the BSC tape fell from 155k to 22k trades a day with 57 of 64 pools in error and about 128,000 failed RPC calls a day, and the Solana tape hung for 8 days and exhausted its Helius allowance; neither showed as "down" (T2 §1.2, §1.3, §6.1, grade A). Fixing the tapes is the first critical-path item for both rights and data.
  2. Nothing we can show forecasts the next move; the product has to be honest about that on screen. Stablecoin, ETF, chain and terminal flows move with price or follow it, with zero to negative 30-day lead after 2023 (T3 §2, A). Light count at entry does not predict the 24 h outcome (+42% / +42% / +31% median best for 1 / 2 / 3+ lights; T5 §2.1, A). Triggers mark volatility in both directions: after an alert 24% touched +20% in an hour but 28% touched -20% (T6 §0, §3.2, A). The mandate's "it does not judge; it shows" is the only position the data supports.
  3. Money moves before talk, and that ordering is the proprietary view. Watched wallets are first on 64% of tokens; in 26 of 34 tokens with both, a tracked buy preceded the caller call by a median 8.6 h (T5 §0, §2.1). Callers are late relative to tracked traders (Pow: 32 of 53 called tokens already had a tracked buy, median 18.8 h earlier; T4 §0, §1.2). No competitor of 19 reviewed computes the attention-to-money lag or "who was already in" (T7 §1.4, §4). The two-lane clock strip is the signature component (T7 §4, C05).
  4. Most of Layer 1 cannot be licensed for display but can be rebuilt on-chain at low cost. DefiLlama bars commercial republication with up to $100,000 per violation; Glassnode, Token Terminal, CryptoQuant, Santiment, Kaito and Nansen labels are not displayable (T1 §0.3, §0.7, A). Stablecoin supply, exchange-wallet balances and terminal volume (fee wallets listed in DefiLlama's open adapter repo) are chain facts we can read ourselves (T1 §0.4; T3 §1.4). T3's first-eight brief costs $0 new money on that basis (T3 §5).
  5. "A few days" is not reachable for the product; it is reachable for the foundation. T2 puts the infrastructure lane alone at 28 to 30 builder days, 8 to 9 days to a supervised tape and price store (T2 §7). T4 puts intake with stance and flips at 7 to 8 days (T4 §7). T9's slow column (20-day build) opens beta wave 1 on 18 Oct (T9 §5.3). My roll-up in §4 puts the honest full build at about 75 builder days for one builder, a wave-0 beta cut at about 48 builder days, and the critical path to wave 0 at about 22 to 25 days if a second builder takes the front end and intake in parallel (T10 estimate, C; the infra and intake days are T2's and T4's).

1.3 The three biggest unknowns

  1. Whether Telegram reading plus LLM stance extraction is permissible for a paid product. Telegram's Content Licensing terms prohibit access beyond "ordinary ... use of the Telegram platform as its user" and bar "aggregation" and use of platform data in the "deployment of artificial intelligence" (T1 §2.4, T4 §6.1, A). T4 grades our position "unclear". Caller stance and flips, the one field nobody else has, sit on this question.
  2. Whether the product produces anything a trader would pay for over free terminals. Nothing measured tonight shows lift that survives a random-moment baseline after 24 h except small cells (crossings +8 pts, alerts +4 pts first-passage; T6 §3.3), and the strict rule loses on every trigger set (T6 §0.4). The value case is consolidation and presentation, which only the beta can test (T9 §1.1, §6).
  3. The real build length with one builder, especially the front end. No worker estimated the pages and components; T2's 28 to 30 days is infra only, and my front-end days in §4 are unmeasured guesses, and T2's own Solana and EVM tape rebuilds carry provider decisions not yet made (T2 §7 items 4, 5).

2. Per topic

Format per topic: findings that matter (with the report section); the decision the topic implies; evidence quality; what the worker said it did not do.

2.1 T1: Sources and rights

Findings 1. Displayed price, volume and flow must come from our own chain tape; DexScreener stays internal as a cross-check and phantom detector (T1 §0.1, §2.5 rule 1, §6 rec 1). 2. Clean, cheap commercial licences exist and say so in writing: CoinGecko Basic $35 ("Powered by CoinGecko", cache refreshed at least every 24 h), DexPaprika Dev $30 (free tier is non-commercial), GoPlus free ("Powered by GoPlus"), Solana Tracker (derived display explicitly allowed), Fomo API (in-app use) (T1 §0.2, §2.1). 3. Our current keyless DexPaprika use is the non-commercial Free plan; move to Dev before any user (T1 §2.1, §6 rec 2). 4. Layer 1 aggregators are mostly not displayable (DefiLlama, Glassnode, Token Terminal, CryptoQuant); rebuild on-chain (T1 §0.3, §2.2). 5. Terminal volume is measurable from fee wallets and trade contracts listed in DefiLlama's adapter repo; copies saved in data/t1-research/defillama-adapters/ (T1 §0.4). 6. Telegram: show derived timestamped facts (channel, token, time, stance, price at call from our tape), never message text, media or member data; no training on the text (T1 §0.5, §2.5 rule 2). 7. X: official pay-per-use only (about $0.005 per post read, 3M/month cap); resellers carry X's reported $15,000 per 1M posts liquidated damages (T1 §0.6, grade B for the damages). 8. Wallet identity from Nansen, Arkham, Kolscan, GMGN is not displayable; our own recurrence labels are the only wallet intelligence we can publish (T1 §0.7, §6 rec 10). 9. Fomo API is "an independent, unofficial tool" with "no affiliation to fomo.family"; show derived facts only (T1 §0.8). 10. MVP spend at 50 tokens is about $213/month, against about $85 today (T1 §5).

Decision implied. Adopt the three conservative display rules of T1 §2.5 (no DexScreener number on screen, derived facts only from Telegram, no DefiLlama number without consent) as build constraints now; send the five vendor emails of T1 §6 rec 9 (DexScreener, DefiLlama, Helius, GoPlus, Bubblemaps).

Evidence quality. High for terms: most clauses are quoted from pages fetched 2026-09-26 (A). B or C for Farside, SoSoValue, CryptoQuant, Arkham, GMGN, Token Sniffer, LunarCrush, Birdeye API, Bitquery, Codex and CMC terms, X pricing table and X ToS (T1 §7). Costs are estimates except the tape swap rate.

Did not do. No signups or keyed calls; program IDs and fee wallets not verified on an explorer; no lawyer; costs mostly assumptions stated in scripts/t1_cost_rollup.py (T1 §7).

2.2 T2: Infrastructure and architecture

Findings 1. 75% of the stack's CPU is re-reading, not collecting: price_paths re-parses its 7 GB ledger (52% of a core) and the server rebuilds a 100 MB read model nobody views (45%) (T2 §1.2, §1.3). 2. The tapes are broken silently: BSC coverage down 7x since 09-20, about 128,000 failing calls a day; Solana hung 8 days and spent its Helius allowance (T2 §1.2, §6.1). 3. All five failures this week share one cause: we measure whether a process runs, not whether data arrives. Heartbeats must count progress (T2 §6.1, §6.2). 4. Target: collectors that only fetch and append; an event standard (providerAt, availableAt, sha256, quality) enforced at write; a tape standard with coverage rows; rebuildable derived fields; one pre-built JSON read model per view; a read-only API (T2 §2.3). 5. Storage: DuckDB over Parquet for tape and derived data, SQLite for events, a small relational store for situations, accounts and feedback (T2 §4.3, option B). 6. Solana must move from per-transaction fetch to a filtered stream; per-tx design needs 1.7 / 6.6 / 16.5M calls a month at 50 / 200 / 500 tokens (T2 §5.3, §5.4). 7. Budget at 50 / 200 / 500 tracked tokens, excluding X: about $230–330 / $620–720 / $1,330–1,730 a month; the box carries 500 tokens (T2 §5.2, §5.3). 8. LLM is the second-largest variable cost and must carry a daily dollar cap per stage with "unavailable" shown when hit (T2 §5.4, §8 D6). 9. Gap accounting: every collector writes coverage rows; gaps print on the page; the 14-day gate is six coverage metrics, and today the clock has not started (T2 §6.3, §6.4). 10. Infra lane: 28–30 builder days; items 1–3 (4.5 days) need no design input and can start tomorrow (T2 §7).

Decision implied. Freeze the experimentation stack except the kept collectors; Bolo starts T2 §7 items 1–3 on day one; choose the Solana stream provider (D2) and storage option B (D3) before day 3.

Evidence quality. Measurements of our machine are A (scripts T2-*, run 00:37–00:55 with nine agents also running; per-process numbers clean, interface totals not). Target CPU, RAM, message volumes and LLM token counts are C estimates (T2 §5.1). Model prices B.

Did not do. Touched nothing; did not attribute the watchtower's 608 MB/day upload; disk growth is a delta against rounded sizes; vendor prices not re-verified (T2, "What I did not do").

2.3 T3: Layer 1, follow the money

Findings 1. No follow-the-money metric leads the trenches: post-2023, 30-day stablecoin growth versus the next 30 days of BTC, ETH, SOL, DOGE, DEX and pump.fun volume has lead r between -0.21 and 0.00; reverse (price first) up to +0.54 (T3 §0.1, §2, A). 2. The only leading structure with evidence is short-horizon momentum within a size band, and it reverses within weeks (T3 §0.2, §2 row 5). 3. Band heat is turnover (volume / market cap): sub-$1M 0.7–2.2x a day, $1–10M 0.16–1.0x, $10–50M 0.02–0.08x; nobody publishes a sub-$50M heat index (T3 §0.4, §3.3). 4. Our watched universe: 40–52 tokens doubled inside 24 h each day this week, almost all under $1M, none at or above $50M; down 16% from the previous week, the drop is Robinhood Chain (T3 §3.1). 5. Phantom prints inflate the 2x count by up to 20% on a day; any shipped counter must run the filter (T3 §3.1). 6. Robinhood own-tape volume is comparable only from 09-24; 47% of BSC swap rows are tokenized-stock pools with no USD value (T3 §3.2). 7. Chain heat live 7 d: Solana $20.8B, Ethereum $11.8B, Robinhood $10.5B (-18.8%), BSC $7.6B; Axiom about $803M and GMGN about $665M a week (T3 §1.4). 8. Ship-first eight: runners today, band heat, breadth, chain heat, launchpad pace, dollars in or out, risk-appetite strip, leverage strip, at $0/month new spend if chain and launchpad numbers come from our tapes and GeckoTerminal (T3 §4, §5). 9. Label the macro rows "the weather now", never a forecast (T3 §2, "What this means for the brief").

Decision implied. The brief is a weather report built from our tapes plus attributed CoinGecko plus free exchange and on-chain reads; the M (30-day) view is honest only from about 2026-10-10 and shows "building" until then (T3 §4).

Evidence quality. Own-store numbers A (T3-trenches.py, T3-evidence-stables.py, rerun and reproduce). Literature A/B. About 70 correlations without correction; weekly resolution only (T3 §2 caveats).

Did not do. No market-wide runners count; no test of whether runners or band turnover lead anything (needs a month); Solana own tape not used; no keyed calls; several papers only via abstracts (T3 §6).

2.4 T4: Discovery intake and attention

Findings 1. Callers are an attention signal, not early money: on the tokens both see, tracked traders or wallets were in first (Pow 32 of 53, median 18.8 h; Sugar 25 of 44, median 64 h) (T4 §0.1, §1.2). 2. Order across sources: own wallet watch, then Fomo alert, then thesis, then caller (T4 §1.2). 3. Admitters barely overlap: 73% of 1,609 tokens in 14 days came from exactly one family; 17 from all four (T4 §0.3). 4. Base rates look good and mean little: 30% of calls reached +20% at a mark; median 24 h mark -10.1% (T4 §0.5, §1.1). 5. Stance lives in chat: 69% of messages are outside the current regex detector; every exit in the 30-message sample is conversational; 13 of 30 need thread context (T4 §0.6, §3.2). 6. Ticker resolution is fragile: 234 of 300 priced calls resolved by ticker; 15% of our symbols map to 2+ addresses; a five-rung resolver ladder is proposed (T4 §0.7, §4). 7. Extraction design: stage 0 no model, stage 1 Haiku-class, stage 2 Sonnet-class on 10–15%; separate stance and position fields; evidence span must be a verbatim substring (T4 §3.3, §3.4). 8. Cost about $0.80/day ($25/month) at 20 human channels; about $3/day in a busy week at 40; a $5/day cap is enough (T4 §3.5, C). 9. Caller expansion: 16 watch-now candidates plus Call Analyser as meta-source, added in batches of five with 7-day read-only probation; remove 14 dead sweep channels (T4 §2). 10. Names next to scorecards: numbers and method, never labels; scorecards for channels that have not opted in stay inside the logged-in product (T4 §6.4).

Decision implied. Keep reading public channels via the web preview for the first build, run LLM extraction as inference only with T4 §6.1 mitigations, build the 200-message eval set before trusting stance, and keep scorecards member-only.

Evidence quality. Own-store numbers A (four scripts on a copy of events.sqlite). Caller expansion evidence mostly A for "listed by Call Analyser", but no candidate has an audited win rate (T4 §2). Extraction costs and escalation rates C.

Did not do. No LLM run; no eval set labelled; no lift against a random-token baseline; no own-tape anomaly admitter measured; phantom classification not applied to caller marks (T4 §8).

2.5 T5: Situation model, fields and lights

Findings 1. A situation is a token plus a clock; lights, not stages (T5 §0.1). 2. Eight lights: wallets, traders, callers, attention, tape, holders, momentum, safety (inverted); narrative is a tag, not a light (T5 §0.2, §2.2). 3. 83% of situations never get a second independent signal within an hour; 17% get two or more, 4.4% three or more (T5 §2.1). 4. Light count at entry does not predict the 24 h outcome; the page must never present it as a forecast (T5 §0.4, §2.1). 5. Bands behave differently: under $1M median best 24 h +46 to +56% and 43–47% close down half or more; $1–10M +29% and 21%; $10–50M +10% and 11% (n 27) (T5 §3.1). 6. Hold-stage wallet fields are buildable today on Robinhood Chain only; BSC lacks a holdings pass, Solana lacks both (T5 §0.7, §5). 7. Swap netting lies about holders: SPRING tape says 1,305 wallets hold, RPC says 750 of 861 early buyers exited (T5 §1.8). 8. Two opposite shapes readable at a glance: SPRING bought before anyone talked (876 wallets before the first Fomo buy, 260 recurring); AGRIPPA bought because someone talked (19 before the call, 1,213 in the hour after) (T5 §1.8). 9. Ranking: sum of decayed intensities plus rate of lighting plus band-relative attention, minus safety; half-lives scale by band; no score shown (T5 §2.2, §2.3). 10. Thirteen gaps handed to T2/T4, led by BSC holdings, Solana wallet fields and a Transfer-log indexer (T5 §7).

Decision implied. Build the situation store around T5's eight lights with the evidence-id constraint; build holder fields on Transfer logs or balance reads, never swap netting.

Evidence quality. A for all own-store numbers (three scripts, 14-day window). Band table inherits the ladder's selection and uses candle highs that can include single prints (T5 §3.1 caveats). External evidence two B secondary sources.

Did not do. Did not measure tape, holders, momentum or safety as signals; half-lives not fitted; no RPC calls; only two Robinhood worked examples (T5 §8).

2.6 T6: Zones and the forgiving win

Findings 1. A forgiving win alone measures volatility, not direction (T6 §0.1, §3.2). 2. Defensible public definition: first-passage, +20% net of costs, held 5 minutes, before -30%, within 24 h, always beside the random-moment rate on the same tokens (T6 §0.2, §4). 3. On that measure: crossings 46.6% vs 38.3% random (+8 pts), alerts 34.3% vs 30.5% (+4), discovery 29.3% vs 32.4% (no advantage, n=41) (T6 §0.3, §3.3). 4. The advantage lives in the first 1 to 4 hours: crossings +22 pts at 1 h, alerts +10 (T6 §0.3). 5. The strict rule loses money on every trigger set, and a random moment does about the same (T6 §0.4, §3.4). 6. Robust to cleaning, delay and cost changes (nothing moves the headline by a third); the strict mean is fragile (up to 10^19 % uncleaned): quote medians only (T6 §3.6). 7. A new multi-print misquote trap the phantom test misses; Bolo should add the filter (T6 §1.1). 8. Zones can honestly be descriptive clocks and rulers per source, never buy or sell instructions; a fading alert is not a sell signal (T6 §0.6, §5). 9. Discovery cannot be measured today: median 70 h between a call being seen and its first tick; 41 of 308 calls usable (T6 §2, §6).

Decision implied. Layer 3 ships only as clocks and rulers with n and baseline printed, after discovery price capture exists, and after a lawyer reads it (see §3 C8).

Evidence quality. All A (two scripts, CSVs saved). Price paths are DexScreener prints only; controls are one construction; alerts are not independent (591 tokens behind 3,337 triggers) (T6 §7).

Did not do. No price-level zones; no executable quotes; no candles or tape (retention too short); follower clip ladder not modelled; no per-caller splits beyond two callers (T6 §7).

2.7 T7: Presentation and competitors

Findings 1. All 19 products reviewed show snapshots of one source at a time: candles, a stat-tile column, trades, holders, risk chips (T7 §1.3). 2. Nobody puts two sources on one clock: attention-to-money lag, who was already in before a call, caller stance and flips, early-holder exits under rising attention, live concentration series, print provenance, refusals (T7 §1.4). 3. Default form for anything read accurately is position on a common scale; ban gauges, pies for comparison, composite scores, dual axes, looping motion, stat-tile grids (T7 §2, §5). 4. Glance design: one state word, one hero number with a reference, one direction; everything else is the second look (T7 §3.1, §3.2). 5. Phone density: six elements per card; the d8 phone board was about 970 px per card, more than a screen (T7 §3.3). 6. Palette fails colour-vision safety on green vs rose (CIEDE2000 10.5 under deuteranopia); direction must be carried by position and glyph (T7 §3.4). 7. Proprietary ranking: refusal line (25), provenance caption plus suspect-print marker (25), attention-to-money lag strip (20), who was already in (15), early holders vs attention (15) (T7 §4). 8. Signature component: the two-lane clock strip (attention above, money below, lag bracket) as the backbone for extra lanes (T7 §4). 9. Twenty-three component types C01–C23 with one preferred form each (T7 §6).

Decision implied. Build the component library around C05 ClockStrip, C01 StateLine, C02 HeroNumber, C03 FreshnessCaption and C16 RefusalLine first; run the six five-second tests of T7 §2.2 with wave 0.

Evidence quality. Literature largely A with five self-corrections recorded (T7 §2). Competitor layouts often C because app pages blocked headless fetch; docs A. Palette computation A.

Did not do. No screenshots of gated apps; no user tests; no computation from trading stores (feasibility judged against the wiki inventory); no palette change (T7 §7).

2.8 T8: Wireframes v0

Findings 1. Six pages: Market brief, Explore, Situation room (with Zones tab and hold view), Record, My board, plus Search in the frame (T8 §1). 2. The frame's most important slot is F3 coverage and clocks ("BSC tape 25 s · SOL tape 40 s · X dark since 18 Sep") (T8 §2). 3. Intake is about 400 tokens a day and the old machine opened about 55 situations a day; the look list shows 25 rows, ceiling 50, and tightens itself on busy days (T8 intro, §4, A for the counts). 4. One room skeleton of 14 slots with three band profiles (first hour, trenches, established) that reorder slots and change windows (T8 §5). 5. Hold view: H1 position line and H2 "what changed since you got in", eight hold events in priority order (liquidity pulled first, price level last) (T8 §6). 6. Zones are a tab on the chart, off by default (T8 §7, Q7). 7. Record: what the system saw, caller scorecards, archive, by-lights-at-open, frozen rooms (T8 §8). 8. Thirty data asks DA-01 to DA-30 (T8 §10), ten open questions with recommendations (T8 §11). 9. HTML wireframes for nine pages in wireframes/, generated from one slot table by scripts/T8-wireframes.py; 18 shots with no overflow (T8 Files).

Decision implied. Accept the page map and the one-skeleton-three-profiles model; reconcile T8's light set, decay windows, band edges and zone parameters to T5 and T6 (see §3).

Evidence quality. Mostly design proposals (C); the only measured number is intake volume (A, T8-intake-volume.py). Example numbers in ASCII sketches are illustrative.

Did not do. No visual design; no data wiring; did not wait for T1/T4/T5/T6 (their parameters overrule T8's); no trader test; no phone interaction model; no admin pages; no onboarding or access gates (T8 §12).

2.9 T9: Beta and commercial

Findings 1. The beta replaces the 7-day trial, not the paid offer: screened, waved, 50 seats, existing passcode gate plus a private Telegram group (T9 short version 1, §1.4). 2. Extend / drop rule written as code: extend on 8 active days, 3 of 4 check-ins and 2 useful items in a 28-day cycle; one probation; hard strikes drop; at most 3 extensions (T9 §1.6). 3. Feedback rubric: Specific, Grounded, Actionable, 0–3; useful is 2 or more; negative is not bad (T9 §1.5). 4. No comparable product ties access length to feedback quality; the rubric must be public (T9 short version 2, §1.10). 5. Price: $69 / $600 holds; the fallback should be $49, not $39 (T9 §4.2). 6. Terminals take about 1% per trade; the top eight took about $145M in fees in 30 days (T9 §4.1, A on DefiLlama fees data). 7. Most of the 09-25 plan survives; its dates, the 7-day trial and "the record is the product" do not; add a situations log with its own frozen basis (T9 §5.1). 8. Free is still redistribution: the rights file stays blocking before the first stranger logs in (T9 §3.5). 9. Terms, risk statement, privacy line and data-rights clause drafted for a lawyer; Layer 3 "buy zone" is the feature most likely to read as advice (T9 §3, §3.6). 10. Ship the situation room plus the brief as a strip to the beta; zones held to Gate 1a and a lawyer (T9 §6).

Decision implied. Adopt the beta shape and $69 / $49; name the legal person before wave 1; plan on the slow column or later (see §3 C1).

Evidence quality. Pricing A/B (several vendor pages blocked); beta precedents A/B; beta thresholds and recruiting yields C; legal points B from search snippets.

Did not do. No legal review; did not read T1 (rights deferred); build length borrowed from the 09-25 critic's 12–20 days and flagged for T10 to replace; nothing sent or created (T9 §8).


3. Contradictions and tensions

Each row: what disagrees, which side is better supported, and the recommendation.

# Tension Side A Side B Better supported Recommendation
C1 Build length T9 §5.3: "a few days" (4-day build) column opens wave 1 on 2 Oct; fast 12 days on 10 Oct; slow 20 days on 18 Oct. Its 12–20 is the 09-25 critic's number for a smaller room, borrowed (T9 §8) T2 §7: infra lane alone 28–30 builder days, 8–9 to the foundation; T4 §7: 7–8 days intake T2 and T4 (itemised, dependency-ordered). T9 itself says to replace its number Plan on §4 of this document: wave-0 cut about 48 builder days, full build about 75, one builder; critical path to wave 0 about 22–25 days with a second builder in parallel. From a 27 Sep start that is wave 0 around 20–22 Oct and wave 1 in the last week of October, later than T9's slow column (18 Oct) (T10 estimate, C)
C2 Displayed price for DexScreener-derived numbers T1 §2.5 rule 1: "no DexScreener number reaches a user screen, raw or derived" T3 §5 (derived counts are ours; move to tape later), T4 §6.3 (marks fine, credit "data: DexScreener"), T7 C03 ("price 1 min old, DexScreener"), wiki synthesis (derived fields such as "40% below 1 h high" allowed) T1 (it quoted the clause; the others read it loosely). The clause bars a directly competing product and making "any portion" available to third parties Take T1's rule for anything user-facing. Consequence: T3's runners and band heat, T4's caller marks and T6's zone statistics must be recomputed on our tape before display. T7's FreshnessCaption names our tape or an attributed licensed source, never DexScreener. Lift only with a written licence (email 1 of T1 §6 rec 9)
C3 Telegram read route T1 §6 rec 7: move to one dedicated MTProto user account, read-only; T2 §1.2: "needs a real Telegram client session" T4 §6.1 point 4: MTProto raises exposure a level (API Terms bind directly, account bans); "not recommended for the first build"; the preview covers public channels at 5-minute cadence T4 for the first build (it measured what the preview gives and quoted the API terms). T1's argument (preview cannot see groups) matters only once groups are in scope Preview for wave 0 and wave 1, public channels only. Decide MTProto after the lawyer's read (Q3). Record message-id gaps between polls as measured loss (T4 §5 rule 5)
C4 Caller text on screen T1 §2.5 rule 2: no message text T4 §6.1 point 2: a verbatim evidence span of at most about 25 words plus the post link; T8 §6: quoted phrase up to 12 words "if T1 clears" T1 (conservative, and T8 defers to it) No quoted text on screen for the beta. Show derived stance, clocks, and a link to the public post; keep the evidence span internal for audit and eval. Revisit after the lawyer (Q4)
C5 Caller scorecards public or not T8 Q8: public Record summary includes caller scorecard medians ("the page people will share") T4 §6.4 rule 3: scorecards for channels that have not opted in stay inside the logged-in product; the public record launches with our own strategies' calls T4 (it cites the disputes found and the exposure pattern: characterising conduct) Member-only scorecards for the beta; public Record = situations log plus the deep-pullback record (T9 §5.1). Numbers and method, never labels (T4 §6.4 rule 1)
C6 Light set T5 §2.2: wallets, traders, callers, attention, tape, holders, momentum, safety; narrative is a tag, not a light T8 E5.5: callers, X, wallets, traders, flow, holders, attention, narrative; T7 C18: money, attention, holders, callers T5 (only one with a measured basis: §2.1 signal sequences; and a stated reason narrative has no event id) Use T5's eight. X becomes a source inside the attention light (hatched when dark), narrative a tag on S1/S11. T8's wireframe script regenerates from one slot table, so the fix is one edit
C7 Decay and band edges T5 §2.2: half-lives by band (<1M / 1–10M / >10M), 72 h hard cap; T5 bands <250K, 250K–1M, 1–10M, 10–50M, >50M T8 §4: fixed windows per category (callers 24 h ... narrative 3 d); T8 profiles cut at $10M and 2 h; T6 bands <250K, 250K–1M, 1–5M, 5–20M, ≥20M; T3 bands <1M, 1–10M, 10–50M, ≥50M T5 for decay (outcome-timed). For bands, no side is measured better; they differ by purpose One band table for the product: T5's five bands. T3's Layer 1 rows merge T5's two lowest bands into "under $1M". T6 reruns on T5's edges before any zone is shown. T8 profiles keep the $10M cut (it matches a T5 edge)
C8 Zone parameters T8 Z4 sketch: "win = +40% within 24 h, held 30 min", hide under n=30 T6 §4: +20% net, held 5 min, before -30%, within 24 h, with random-moment baseline; hide under n=20; medians only T6 (measured; T8 says its numbers are placeholders) T6's definitions verbatim; T8 Z4 must print the baseline next to every share
C9 Win rates vs base rates T4 §1.1: 30% of calls reached +20% at a mark; Fomo tracked buys 57% candle-high touch at 24 h; T5 §3.1 median best 24 h +42% to +56% in small bands T6 §3.2–3.3: after controls, triggers raise the chance of a 20% move in both directions; lift at 24 h is +4 to +8 pts or none T6 (controls, cleaning, sensitivity). T4 and T5 both say their numbers are base rates, not wins Never show a touch rate or best-24h median without the matched random-moment rate. Record R4 (T8) and caller scorecards print the baseline column
C10 LLM cost T2 §5.2: $183 / $523 / $1,132 a month at 50 / 200 / 500 tokens (6,300 / 18,000 / 39,000 messages a day; stance on 20%) T4 §3.5: about $25/month at 20 human channels (300 messages a day reaching the model after stage 0) T4 for the beta (anchored to measured channel volumes: our five channels give 28–127 messages a day in total, T2 §5.1). T2 is the ceiling if whole-chat reading of busy groups is added Budget T4's figure for wave 1 with a $5/day cap (T4 §3.5); keep T2's caps (D6) as the ceiling design. Measure the real escalation rate in T4 step 4
C11 What price_paths is T9 §5.1 (B-minus row): "the chain tapes and price paths become the companion's backbone" T2 §1.2: rebuild; it re-scans a 7 GB ledger at 52% of a core; forward returns should be computed once from tape and candles; move to the research bench T2 (measured) Tapes are the backbone; price_paths goes to the research bench (T2 §7 item 13)
C12 BSC tape as a strength T1 §3, T7 §4, wiki synthesis: "BSC continuous since 18 Sep" T2 §1.2: coverage fell 7x (155k rows 09-20 to 22k 09-25), 57 of 64 pools erroring T2 (measured tonight) Treat BSC as broken until T2 §7 item 4 (EVM tape v2) passes the coverage metric
C13 Data asks vs rights T8 DA-01 (X mentions, P0), DA-08 (holder count from a provider), DA-30 (chart source for untaped chains) T1: X only via official API at about $1.2k/month at 50 tokens (§5); Bubblemaps terms unknown, Nansen/Arkham no (§2.4); candles only from our tape, or DexPaprika Dev / CoinGecko with attribution and a 24 h cache rule (§1c) T1 DA-01 drops to P1 and shows "X dark" (T4 §5) until Q8; DA-08 is built from a Transfer-log indexer (T5 gap 3), not bought; DA-30 answered by T1 §1c: our tape first, DexPaprika Dev for untaped chains with "Powered by DexPaprika"
C14 Solana spend T1 §5: Helius Developer $49 now, Business $499 at 500 tokens, or Shyft $99 stream T2 §5.4, D2: move to a filtered stream now; per-tx fetch only for gap repair T2 (the per-tx design already failed; T1 also prefers the stream at scale) Stream now (Sodae free 5 GB for 50 tokens, Shyft about $99 or Chainstack $49 per stream above; T2 §5.3), plus Helius Developer $49 for gap repair and holdings reads
C15 Brief scope for the beta Mandate and T8 §3: full Layer 1 page with twelve blocks T9 §6: brief as a strip; T3 §4: ship first eight T9 on sequencing (one clock), T3 on content Build T3's first eight as the data; render them first as a strip on Explore (T8 E1) and a thin Brief page; grow the page from access-log evidence (T9 §6)
C16 Reliability gate timing 09-25 plan: gate before trials T9 §5.1: gate before paid, running and disclosed during the free beta; T2 §6.4: clock starts only after T2 §7 items 1–2 are live T9 and T2 agree Adopt T2's six metrics as the gate; start the clock when items 1–2 are live; close before paid

4. The build module list

Day estimates are honest working days for one builder including tests and a live soak, the basis T2 §7 and T4 §7 use. Where a report gives the number it is cited; where none does it is T10 estimate (C), reasoned from the size of comparable items in T2 §7. "No design input" means the module can start without waiting for visual design or new product decisions; "decision" names the question in §8 it waits on. W0 marks modules in the wave-0 beta cut (§4.7).

4.1 Infrastructure

ID Module Depends on Specified in No design input? Days W0
I1 Stop the bleeding: retire two crash-looping units, kill five hung wrangler deploys, stop BSC retries on out-of-window pools, free NodeReal key none T2 §7 item 1 yes 0.5 (T2) W0
I2 Supervisor, progress heartbeats, sd_notify watchdog, stall rules, coverage rows, /status none T2 §6.2–6.3, §7 item 2 yes 2 (T2) W0
I3 Event standard library (providerAt, availableAt, sha256, quality), raw sidecar, price collector with unchanged, phantom and misquote flags none T2 §2.3, §7 item 3; T6 §1.1 (misquote filter) yes 2 (T2) W0
I4 EVM tape v2: one address-array getLogs per window per chain, coverage rows; BSC, Base, Robinhood I2, I3; Q7 for Robinhood T2 §7 item 4 yes 2 (T2) W0
I5 Solana tape v2: filtered stream, swap parser (Raydium, Pump AMM, Meteora, Orca), per-tx fetch only for gap repair I2, I3; Q6 provider T2 §7 item 5; T1 §1b yes, after Q6 3–4 (T2) W0
I6 Storage layout and caverio.query: Parquet day close, DuckDB, one candle schema from tape, migrate tapes and events I3, I4 T2 §4, §7 item 6 yes 2 (T2) W0
I7 Holder truth: ERC-20 Transfer-log indexer and balance reads on BSC; Solana token-account reads; holder count, top-N share, deployer share I4, I5 T5 §1.3, §7 gaps 1–3; T8 DA-05, DA-06, DA-08 yes 3 (T10 estimate, C) wave 1 (W0 shows tape-derived net positions, labelled, as T8 S7 allows)
I8 Safety: GoPlus flags on open ("Powered by GoPlus"), LP add/remove/burn/lock events, deployer identity from the create tx I4, I5 T5 §1.5, §7 gaps 4–5; T8 DA-09, DA-10, DA-11; T1 §2.4 (GoPlus) yes 2.5 (T10 estimate, C); W0 lite 1.5 (flags plus LP events) W0 lite
I9 Macro collector and brief series: T3's first eight from our tapes, CoinGecko /global and /coins/markets (attributed), exchange public funding/OI, on-chain stablecoin totalSupply I3 T3 §4, §5; T2 §7 item 11 yes 2 (T2); W0 lite 1 (runners, band heat, breadth, dominance) W0 lite
I10 Terminal-volume watcher: about 40 fee accounts, daily rotation check I3, I5 T3 §1.4; T1 §0.4 yes 1.5 (T10 estimate, C); maintenance is the real cost (T3 §1.4) no
I11 Read models and API: one JSON per view, ETags, beta auth, per-user rate limits M1, M2, I9 T2 §2.3, §4.5, §7 item 12 needs T8 v0 (exists) 3 (T2) W0
I12 Retire the experimentation stack: server, price_paths to bench, holdings, social-sweep/reader, proof services I11 live 3 days T2 §7 item 13 yes 1 (T2) no
I13 Gate instrumentation: daily independent refetch, six metrics on /status I2, I6 T2 §6.4, §7 item 14 yes 1.5 (T2) no (needed before paid)
I14 Price capture at call time for discovery (one read at seenAt, then the situation poll) I3 T6 §6 yes 0.5 (T10 estimate, C) W0

4.2 Intake

ID Module Depends on Specified in No design input? Days W0
N1 Source hygiene: drop 14 dead sweep channels, availableAt on caller rows, silent-vs-blind health, dark-state strings none T4 §5, §7 step 1 yes 0.5 (T4) W0
N2 Dedupe (text hash per channel per 24 h), promo filter, bot-feed parsers N1 T4 §3.4 stage 0, §7 step 2 yes 0.5 (T4) W0
N3 Eval set: 200 messages, two labellers, kappa N2 T4 §3.6, §7 step 3 yes (half is human time) 1 (T4) wave 1
N4 Extractor v1: stages 0/1/2, structured outputs, derived events, cost ledger with cap, Batch backfill of the 1,557 held messages N3 (tuning); Q3 T4 §3.3–3.5, §7 step 4 after Q3 1.5 (T4) wave 1
N5 Resolver ladder, shared by regex path and extractor; Solana mint validation N4 (or regex path alone for W0) T4 §4, §7 step 5 yes 1 (T4) W0
N6 Tune to the eval targets, freeze prompt v1 N3, N4, N5 T4 §3.6, §7 step 6 yes 0.5–1 (T4) wave 1
N7 caller_stance table, flip detector, admin page for unresolved references N6 T4 §3.7–3.8, §7 step 7 yes 1 (T4) wave 1
N8 "Who was already in" field and caller admission rules N5 T4 §1.2, §7 step 8; T5 §1.3 yes 0.5 (T4) W0
N9 Caller expansion in batches of five with 7-day probation N1, N2 T4 §2, §7 step 9 yes 0.5, then ongoing (T4) start W0
N10 Per-channel scorecards (internal first) on tape-priced marks N8, M4 T4 §6.4, §7 step 10; C2, C5 above yes 1 (T4) wave 1
N11 Fomo watcher moved onto the event library under a real unit I2, I3 T2 §1.2 (Fomo row), §7 item 7 yes 1 (T10 estimate, C) W0
N12 Narrative tagging: metadata, name/ticker heuristics, caller-text pass I3, N4 for text T5 §4.1 yes 1 (T10 estimate, C) wave 1

T4's own total for N1–N10 is 7 to 8 days (T4 §7); it replaces T2 §7 items 7–8 (4–5 days) in this roll-up.

4.3 Situation model

ID Module Depends on Specified in No design input? Days W0
M1 Derived fields v1: per-minute net flow, buyers/sellers, new vs returning, size mix, first-hour tape, drawdown, liq/cap, volume vs band I6 T5 §1.1–1.2; T2 §7 item 9 yes (T5 field list exists) 3 (T2) W0
M2 Situation store, eight lights with evidence ids, decay, ranking, lifecycle, reopen as new situation M1 T5 §2, §6; T2 §7 item 10; C6, C7 above yes 2 (T2) W0
M3 15-minute band aggregate table (mentions, flow, buyers per band and pool-age bucket) for band-relative percentiles M1 T5 §7 gap 10; T8 DA-04 yes 1 (T10 estimate, C) W0
M4 Outcome marks on honest fills per situation (1 h, 4 h, 24 h, 72 h, best in 24 h), tape-priced, repeats and phantoms classified I6, M2 T8 DA-27; T6 §1 cleaning rules yes 1.5 (T10 estimate, C) W0
M5 Zone job: T6 clocks and rulers per source and band, n and random baseline, recomputed on tape candles M4, I14, candle retention beyond 12 h (T6 §6) T6 §4–5; T8 DA-29 yes, but gated on Q2 and a lawyer 2 (T10 estimate, C) no (after Gate 1a)
M6 Narrative heat rollup (count, money, attention vs 7-day median; hidden under 10 tokens) N12, M3 T5 §4.2–4.3 yes 0.5 (T10 estimate, C) wave 1

4.4 Presentation (component library)

ID Module Depends on Specified in No design input? Days W0
P1 Core: C01 StateLine, C02 HeroNumber, C03 FreshnessCaption, C04 NumberRow, C16 RefusalLine, C17 WaitingForLine, C18 LightsRow (T5's eight), C23 EndOfList none (fixture JSON) T7 §3, §6 forms yes; colour and type wait for the visual-design step 2 (T10 estimate, C) W0
P2 Signature: C05 ClockStrip, C06 LagBracket, C07 FlowBars, C14 GapHatch, C15 SuspectPrint P1 T7 §4, §6 as P1 3 (T10 estimate, C) W0
P3 Remaining: C08 BulletBar, C09 LaneLine, C10 PeerStrip, C11 StanceTrack, C12 QuantileDots, C13 ZoneBand, C19–C22 P1 T7 §6 as P1 2 (T10 estimate, C) partly (C08, C09)
P4 Five-second tests (six pairs) with wave 0 testers P1, P2 T7 §2.2 n/a 0.5 (T10 estimate, C) W0

4.5 Pages

ID Module Depends on Specified in No design input? Days W0
G1 Frame: nav, search stub, F3 coverage and clocks, time view, filters; passcode session I11 (or fixtures), K4 T8 §2 yes (grey) 1 (T10 estimate, C) W0
G2 Explore: just lit, 25-row list with eight light cells, cooling, coverage note M2, P1 T8 §4 yes 2 (T10 estimate, C) W0
G3 Situation room: 14 slots, three band profiles, S14 unknowns M1, M2, P1, P2 T8 §5 yes 4 (T10 estimate, C) W0
G4 Hold view (H0–H3), My board (M1–M5), user store (follow, entry, last-seen) G3, T8 DA-23 T8 §6, §9 yes 2 (T10 estimate, C) W0
G5 Brief: strip on Explore (W0), then the page with T3's first eight I9 T8 §3; T3 §4; T9 §6 yes 0.5 strip, 2 page (T10 estimate, C) strip W0
G6 Record: situations log with frozen basis and start date, R1, R3, R4, R7; frozen rooms M4, T8 DA-26 T8 §8; T9 §5.1 yes 2 (T10 estimate, C); W0 minimal 1 W0 minimal
G7 Alerts: Telegram bot delivery, quiet hours, dedupe; observations only M2, G4 T8 §6, DA-25; T9 §5.1 alerts row yes 1.5 (T10 estimate, C) wave 1
G8 Zones tab M5, P3 T8 §7; T6 §4–5 no (Q2, lawyer) 1.5 (T10 estimate, C) no
G9 On-demand situation from search I4, I5, M2 T8 DA-24 yes 1 (T10 estimate, C) wave 1

4.6 Beta and commercial

ID Module Depends on Specified in No design input? Days W0
K1 Legal person named; lawyer engaged on the five T9 §3.6 questions; terms, risk, privacy, data-rights clause Thomas T9 §3; Q11 n/a (not builder time) Thomas before wave 1 (wave 0 friends can use drafts, T9 §5.1)
K2 Rights file widened to every displayed source; attribution marks in the component spec; five vendor emails; DexPaprika Dev, Helius Developer T1 T1 §2, §6 recs 2, 3, 5, 9; T9 §3.5 yes Vesper, not builder time before wave 0 strangers
K3 Tally form and webhook endpoint writing beta/applications.jsonl none T9 §1.3, §1.7 yes 0.5 (T10 estimate, C) W0
K4 Gate hook: per-passcode expiry, tester id on every session (for active_days) none T9 §1.7 yes 0.5 (T10 estimate, C) W0
K5 Private group and bot: named invite links, /checkin, /moment, voice via Deepgram, beta/feedback.jsonl K4 T9 §1.5, §1.7 yes 1.5 (T10 estimate, C) W0
K6 Evaluator (beta/evaluate.py, day-21 warning, day-28 decision) and Monday digest K5 T9 §1.6–1.7 yes 1 (T10 estimate, C) wave 1 (first decision is day 28)

4.7 Roll-up, critical path and the two timelines

Totals (one builder, sequential).

Lane All modules Wave-0 cut
Infrastructure 27 (I1–I14; T2 items plus I7, I8, I10, I14) 18 (I1–I6, I8 lite, I9 lite, I11, I14)
Intake 10 (N1–N12) 4 (N1, N2, N5, N8, N9 start, N11)
Situation model 10 (M1–M6) 7.5 (M1–M4)
Presentation 7.5 (P1–P4) 5.5 (P1, P2, P4)
Pages 17.5 (G1–G9, full brief) 10.5 (G1–G4, G5 strip, G6 minimal)
Beta tooling 3.5 (K3–K6) 2.5 (K3–K5)
Total about 75 builder days about 48 builder days

The front-end days (P, G) and several infra additions are my estimates (C). If they are 30% optimistic, as the 09-25 ops critic found for earlier builds ("five rebuilds in seventeen days", T9 §6), the totals are about 98 and 62.

Critical path to wave 0 (dependency chain, assuming a second builder takes P, G and N in parallel against fixture read models): I1 (0.5) → I2 (2) → I3 (2) → I5 Solana stream (3.5) → I6 (2) → M1 (3) → M2 (2) → I11 (3) → G3 wired to live read models and soaked (4) → wave 0. About 22 days. I4 (EVM) runs beside I5; K1 and K2 run beside everything and must finish before strangers. The Solana stream provider decision (Q6) sits on this path on day 3.

Timeline A: "a few days" (what is actually reachable by about 30 Sep). - I1, I2, I3: supervised collectors with progress heartbeats and coverage rows; event standard at write; misquote and phantom flags on prints (4.5 days, T2 §7). - N1, N2: caller source hygiene, dedupe, promo filter (1 day, T4 §7). - K3, K4: beta application form live and the gate hook (1 day, C). - In parallel, not builder time: the five vendor emails, DexPaprika Dev and Helius Developer upgrades, legal person named and drafts to a lawyer, Solana stream provider chosen, clean-up of MASTER.md into a Bolo brief. - Not reachable in a few days: any page a tester could use on live data. T8's HTML wireframes (wireframes/) are the only clickable artefact and carry field names, not values.

Timeline B: the honest full build.

Milestone One builder Two builders (second on front end and intake)
Build start 27 Sep 27 Sep
Foundation (I1–I6) live about 9–10 Oct about 9 Oct
Wave 0 (5 friendly, wave-0 cut) about 14 Nov (48 days) about 19–22 Oct (critical path 22 days plus slack)
Wave 1 (15) about 18 Nov about 26 Oct
Remaining modules (stance, holders truth, alerts, full brief, record) through early Dec through mid Nov
Zones (M5, G8) after Gate 1a and the lawyer same

Calendar dates assume Bolo works every day, as T9 §5.3 does. Two builders means Bolo plus either a second human or dedicated Rook/Vesper sub-agent lanes on P, G and N with Bolo reviewing; which one is a staffing question for Thomas (Q10).


5. Owned fields register

Fields the companion can show that are ours (computed from our tape, RPC reads or our machinery) or derived by us. Consolidated from T5 §1, T3 §3–4, T6 §4–5 and the 09-25 wiki synthesis ("Novel views"). Status: - today = computable from stores that run now, on the chains named; - infra = needs a module from §4 (named); - rights = needs a licence, consent or a lawyer's read before display.

Several "today" fields are computed today from DexScreener prints; under rule C2 they become displayable only once repriced on our tape, so they carry "infra (tape basis)" as well.

Field Layer Source Chains Status Needs Report
Situation clock (first sighting, first source) 2 T5 join over intake events all today M2 for the store T5 §1.1
Pool age 2 tape first swap RH, BSC; SOL after I5 today (RH, BSC) / infra I5; tape from pool open (T5 gap 9) T5 §1.1
Cap band and band moves 1, 2 band table on cap all today, but cap is DexScreener-derived infra (tape basis, C2) T5 §1.1, §3
Price, % from open, drawdown from high 2 own tape swaps RH, BSC today (RH, BSC; BSC coverage degraded) I4, I5 for full coverage T5 §1.1
Volume 1 h / 24 h, volume / liquidity 2 tape RH, BSC today / infra I4, I5 T5 §1.1
Net buy pressure by window, acceleration 2 tape (telemetry/) RH, BSC today I5 for SOL; M1 to rebuild T5 §1.2
Buyer / seller counts, new vs returning wallets, size mix 2 tape RH, BSC today M1 T5 §1.2
First-hour tape (buyers, net USD, top-1 share, time to 100 buyers); sniper and bundle share 2 tape from pool open RH, BSC where watched from open infra T5 gap 9, T8 DA-12 T5 §1.2; T8 S6a
Tracked-wallet activity 2 own on-chain watch RH; SOL/BSC partial today N11 pattern for SOL/BSC T5 §1.2
Fomo trader flow (who, when, side, size bucket; no thesis text) 2 Fomo API RH, SOL, BSC today rights: upstream of fomoapi.io unverified (T1 §0.8) T5 §1.2
Early-buyer retention (held / reduced / exited) 2 RPC balance reads RH only today (RH) / infra I7 for BSC, SOL T5 §1.3
Concentration clock (top-20, top-1 share over time) 2 Transfer logs or balances none reliably infra I7 (tape-only overstates; T5 §1.8) T5 §1.3; wiki view 5
Holder count, deployer / top holder share 2 Transfer-log indexer none infra I7, I8 T5 §1.3
Who was already in (before first call / first covered buy) 2 tape plus intake join RH, BSC today (RH, BSC) N8 to make it live T5 §1.8; T4 §1.2; wiki view 2
Recurring wallets (seen in 3+ launches) 2 holdings cohorts RH today (RH); effect unaudited (0.10 vs 0.27) I7 for others T5 §1.3; wiki view 2
Attention-to-money lag 2 intake times plus tape RH, BSC today (RH, BSC) N8, M1 T5 §1.4; wiki view 1; T7 §4 rank 3
Caller calls (channel, time, cap at call from our tape) 2 caller rows any named chain today for facts; cap at call infra tape basis (C2); T1 §2.5 rule 2 T5 §1.4; T1 §3
Caller stance and flips 2 LLM extraction over public posts any infra and rights N3–N7; Q3, Q4 T4 §3.8; T5 §1.4
Telegram mention counts per hour, attention vs band percentile 2 sweep counts situation tokens today (thin: five channels, X dark) M3; N9 for breadth T5 §1.4
Caller scorecard (marks at 15 m, 1 h, 4 h, 24 h, n, censoring) 2 marks priced chains infra (tape-priced marks) and rights (member-only, C5) M4, N10 T5 §1.4; T4 §6.4; wiki view 7
Suspect-print flag ("no swap at this price") 2 tape vs prints RH, BSC today I3 flags T5 §1.5; T7 §4 rank 2; wiki view 4
Rug signature (liquidity pull, all sells) 2 tape, LP events RH, BSC infra I8 T5 §1.5
Contract flags, LP status 2 GoPlus (attributed) or own checks SOL, EVM; RH unconfirmed infra and rights (GoPlus §6 to confirm) I8; T1 §6 rec 9 T5 §1.5; T1 §2.4
Deployer history 2 create tx plus outcomes all taped infra I8 T5 §1.5
Refusal line (why not lit, value vs threshold) 2 our machinery all today M2 to emit per light T7 §4 rank 1
Freshness caption (source, age, source count) all heartbeats and providerAt all infra I2, I3 T7 §3.5, C03
Previous situations on this token; comparable past situation 2 situation archive all today (from experimentation stores) / infra M2, M4 T5 §1.6
Lights and rank (eight categories, decayed) 2 situation store all infra M2 T5 §2
Narrative tag and narrative heat 2 metadata, heuristics, caller text all taped for money heat infra N12, M6 T5 §4
Runners today by band (2x inside 24 h, phantom-filtered) 1 ticks now; tape later SOL, RH, BSC watched today (DexScreener basis) infra (tape basis, C2) T3 §3.1, §4 row 1
Band heat (turnover) and breadth (top-5 share) 1 ticks and tapes watched universe today (DexScreener basis) infra (tape basis, C2); CoinGecko sweep above $50M T3 §3.3, §4 rows 2–3
Own-tape volume per chain 1 tape RH (from 09-24), BSC today I4 for full BSC T3 §3.2
Stablecoin supply per chain and issuer, net issuance 1 own totalSupply and mint/burn reads EVM, SOL infra I9 T3 §1.1; T1 §1a
Exchange-wallet stablecoin balances 1 balanceOf on exchange proof-of-reserve addresses EVM infra (address lists weak) I9 T1 §1a, §4
Terminal volume per chain (degen activity) 1 fee-wallet watcher SOL, BSC, RH infra I10 T3 §1.4; T1 §0.4
Launchpad pace (volume, launches, graduations) 1 own launchpad log readers per launchpad infra I9 T3 §4 row 5
Quiet-hours map 1, record refusal log and tape all today G6 (R5) T7 §4 rank 11; wiki view 6
Forgiving win (first-passage, with random baseline) 3, record outcome paths all infra (tape basis, C2; candle retention) M4, M5 T6 §4
Zone clock and ruler (time-to-best, drawdown-before-best per source) 3 outcome paths all infra and rights (lawyer, Q2) M5, G8 T6 §5
Situations log (what we opened, when, from which sources, outcome) record situation store plus marks all infra M2, M4, G6 T9 §5.1; T8 §8

6. Rights summary (from T1)

Display rule per source, as T1 found the terms on 2026-09-26 (T1 §1, §2). RAW = may show as-is; DERIVED = only numbers we computed; NO = not without a deal; UNCLEAR = terms silent or ambiguous. Nothing here is a lawyer's reading (T1 §7).

Source Raw Derived Attribution / cache Commercial Rule for the companion Grade
Own chain tape, own RPC reads, own machinery yes yes none yes primary for every displayed price, volume and flow A
DexScreener unclear, treat as no unclear none stated "compete directly" and "any portion ... available for third parties" clauses internal cross-check only; no number on screen, raw or derived, without a written licence A
CoinGecko / GeckoTerminal /onchain (Basic $35, paid) yes in product yes "Powered by CoinGecko", ≥10 font; refresh cache at least every 24 h yes; no resale of API access Layer 1 global, dominance, categories; live charts only, never stored backtests A
DexPaprika (Dev $30) yes on paid plan yes "Powered by DexPaprika", ≥10 pt Free plan is non-commercial fallback candles for untaped chains, after moving to Dev A
Solana Tracker no yes, explicitly no endorsement implied yes option for Solana derived fields A
Birdeye no (dashboard info) unclear n/a API terms not found fallback; ask first A / C
Moralis unclear, leaning no unclear n/a "for use by a third party" barred skip A
Coinglass no probably n/a from Standard $299 derived bands only, full set A-
Coinalyze yes with citation and link yes cite and link free MVP derivatives B
DefiLlama no no without written consent encouraged personal, non-commercial; up to $100,000 per violation internal only; ask for consent A
Dune no as a running feed occasional citation with credit credit query author no substitute or competing use research only A
Glassnode no no n/a public display needs prior authorisation skip A
CryptoQuant no (reported CC BY-NC) no n/a non-commercial skip B
Token Terminal no no n/a no commercial purpose skip A
Artemis, Farside, SoSoValue unclear unclear n/a no terms found; sites block fetch skip; ETF flows are a gap C
Helius not addressed not addressed n/a "benefit of third party" clause read literally bars apps primary Solana infra; one email to confirm A / C
publicnode, Robinhood Chain RPC chain data chain data n/a no resale of access fallback ring members A
Envio HyperSync ours once rebuilt yes credit HyperIndex if its code ships no hosting generated code as a service backfill route A
pump.fun website/API no no n/a bots and commercial use banned on-chain programs only A
Raydium, Jupiter, Meteora, PancakeSwap, Uniswap (on-chain) chain data yes Jupiter label when showing its routing yes on-chain reads A / B
Fomo API no as a standalone dataset yes in-app keep their notices in-app use; upstream unofficial derived facts ("3 tracked traders bought in the first hour"), no thesis text A
Telegram (preview or client) no in practice (no text, media, member data) unclear (terms silent) n/a "ordinary ... use ... as its user"; no aggregation for AI handle, token, time, stance, price at call from our tape; inference only, no training A
X API (official, pay-per-use) only with full Display Requirements, delete within 24 h not addressed avatar, handle, permalink, logo no model training; no profiling or inferring financial status dark for the beta; derived facts only if it returns A
X resellers no no n/a X anti-scraping, reported $15,000 per 1M posts skip B
Santiment no no n/a private use only; resale needs corporate internal research only A
Kaito no no n/a no commercial use skip A
LunarCrush unclear unclear n/a terms page did not render read terms before use C
Nansen labels never smart-money data prohibited "Powered by Nansen API" for other endpoints no competing products skip for display A
Arkham no without consent no n/a no third-party disclosure skip B
Bubblemaps unclear unclear n/a no terms in docs email before use C
GoPlus yes in app with mark yes "Powered by GoPlus" / Go+ Security §6 no commercial use of original data without permission primary safety flags; confirm §6 by email A
RugCheck, Token Sniffer, Cielo, GMGN unclear unclear n/a terms not readable do not display C
Kolscan no automation no n/a scraping banned skip A

Also binding from T4 §6.2 and §6.4: no handle-to-wallet bridge unless the person published the wallet; caller scorecards carry method, price source, censoring caveat and a correction contact, and never characterise conduct.


7. Cost summary: 50 / 200 / 500 tracked tokens

"Tracked" means a live situation: pools taped, price polled, attention read (T2 §5). Three reports costed overlapping scopes; they are shown side by side rather than added, because T2's total already includes an LLM line that T1 excludes and T4 recosts.

7.1 Money per month

Line 50 200 500 Source and basis
Today's spend (Fomo Starter + CoinGecko Basic) about $85 T1 §5
T1 MVP data stack (DexPaprika, CoinGecko, GoPlus free, Helius, Fomo, one EVM key, free Coinalyze) $213 not computed $732 (about $381 with a Solana stream instead of Helius Business, C) T1 §5, scripts/t1_cost_rollup.py
T1 full set (adds X API, Coinglass, DefiLlama API with consent, LunarCrush, Solana Tracker at 500) $2,312 not computed $10,045 T1 §5; X is $1,200 / $7,950 of it
T2 total excluding X (LLM + Solana data + EVM RPC + Fomo; market data at current plans) about $230–330 about $620–720 about $1,330–1,730 T2 §5.3
of which T2's LLM line (extraction + stance) $183 $523 $1,132 T2 §5.2 (C)
T4 LLM recost at 20 human channels (independent of token count) about $25 about $25 about $25 (about $90 at 40 busy channels) T4 §3.5 (C); $3/day busy week
T3 Layer 1 first eight $0 new $0 new $0 new T3 §5; $300 DefiLlama Pro and about $399 Dune only if chosen
T5 narrative tagging under $30 under $30 under $30 T5 §4.1 says "under a dollar a day"; monthly figure is my arithmetic
T9 beta tooling (Tally free, Deepgram, rubric on the subscription route) "a few dollars" T9 §1.7 (C)
Hosting $0 marginal on Bot-PC $0 $0 T2 §5.3; a second box is T2 §8 D8 (not in §8 of this file)

Reading. At the beta design point the realistic monthly bill is T1's MVP data stack plus T4's LLM figure plus a Solana stream: about $213 + $25 + $0–99 = about $240–340 at 50 tokens (my sum of T1 and T4 lines, C). No report computed T1's lines at 200 tokens; T2's $620–720 at 200 is the only 200-token figure and it carries T2's higher LLM assumption. X is the single line that changes the order of magnitude (T1 §5).

7.2 LLM tokens per day

Stage 50 200 500 Source
Extraction, Haiku-class (T2 basis: 6,300 / 18,000 / 39,000 messages a day) 1.5M 4.2M 9.2M T2 §5.2 (C)
Stance, Sonnet-class (T2 basis) 1.4M 4.0M 8.7M T2 §5.2 (C)
T2 total 2.9M 8.2M 17.9M sum of T2 rows
T4 basis at 20 human channels (300 messages to stage 1 at about 2,350 tokens each incl. cached prompt; 45 to stage 2 at about 2,750) about 0.83M same same my arithmetic on T4 §3.5 figures (C)
Narrative tagging under 1M under 1M under 1M T5 §4.1 (C)
Beta rubric scoring about 20K (600K a month) T9 §1.7 (C), subscription route, not metered

7.3 Machine budget (target design)

Metric 50 200 500 Source
CPU cores 0.65 1.1 2.0 T2 §5.2 (C)
RAM GB 1.8 2.7 4.5 T2 §5.2 (C)
Bandwidth down GB/day, stream design 0.4 1.5 3.8 T2 §5.2
Disk GB/month, capture compressed 1.1 4.4 11 T2 §5.2
Requests/day, EVM getLogs (flat) 57,600 57,600 57,600 T2 §5.2
Messages/day ingested 6,300 18,000 39,000 T2 §5.2 (C)

Today's stack for about 100 tokens: 1.3 cores, 4.3 GB RSS, about 3.3 GB/day down (T2 §1.2). The box (28 cores, 64 GB, 567 GB free) is not a constraint at any size (T2 §5.2).


8. Questions for Thomas (twelve, deduplicated)

Each with Vesper's recommendation and the reports that raised it. Q1, Q3, Q6 and Q10 block the build's first week; the rest can wait until the build is running.

# Question Vesper's recommendation Raised in
Q1 Displayed price source. Is our own chain tape the only source of any price, volume or flow a user sees, with DexScreener internal only and DexPaprika Dev / CoinGecko (attributed) for untaped chains? Yes. It is both the rights answer and the data-quality answer; it puts the Solana and BSC tape rebuilds first on the critical path T1 §2.5, §6 rec 1; T2 §8 D7; T8 Q1; T3 §5; C2
Q2 What ships to the beta. Situation room plus the brief as a strip; zones held until Gate 1a and a lawyer's read? Yes. Zones as clocks and rulers only (T6 §5), tab off by default (T8 Q7) T9 §6, §7.4; T8 Q7; T6 §5
Q3 Telegram route and LLM on channel text. Stay on the public web preview and run LLM extraction as inference only (T4 option a), or deterministic parsing only (option b), or a dedicated MTProto account (T1)? Preview plus option (a) with T4's mitigations for wave 0 and wave 1; MTProto only after the lawyer; never train on channel text T4 §6.1; T1 §0.5, §6 rec 7; T2 §1.2; C3
Q4 Caller text and scorecards. No quoted caller text on screen (derived stance plus a link to the public post), and caller scorecards member-only? Yes to both for the beta; public Record is the situations log and the deep-pullback record T1 §2.5; T4 §6.1, §6.4; T8 §6, Q8; T9 §5.1; C4, C5
Q5 Records. Keep the deep-pullback record public as R&D, and add a situations log with its own frozen basis and start date as the companion's record? Yes. Freeze basis and start date before any tester sees it T9 §5.1, §7.5; T8 §8
Q6 Solana capture. Move to a filtered stream now (Sodae free for 50 tokens; Shyft about $99 or Chainstack $49 per stream above), with Helius Developer $49 for gap repair and holdings? Yes; decide the provider by build day 3, when I5 starts T2 §8 D2, §5.4; T1 §6 rec 5; C14
Q7 Robinhood Chain. Keep taping it in the companion? Keep. It is the only chain with complete holdings today (T5 §0.7), #3 by DEX volume (T3 §1.4), and the most distinctive recruiting community (T9 §2.1) T2 §8 D4
Q8 X. Stay dark for the beta, or fund the official pay-per-use API (about $1,200/month at 50 tokens)? Dark for the beta, with the "X dark since 18 Sep" line everywhere; revisit on beta evidence T1 §0.6, §6 rec 8; T4 §6.2; T8 DA-01; C13
Q9 Beta shape and price. 28-day cycles, extend on 8 active days + 3 of 4 check-ins + 2 useful items, cap 50, waves of 15; beta replaces the 7-day trial for the first 50; $69 / $600 with a $49 fallback instead of $39? Yes to all; $69 only if the brief strip and the room ship together, else $49 T9 §7.1–7.3, §4.2
Q10 Staffing and timeline. One builder (wave 0 around mid November) or a second lane on front end and intake (wave 0 around 20 Oct)? Which plan do we publish internally? Two lanes: Bolo on I-lane and review, a second lane on P, G and N against fixture read models. Plan internally on the two-builder column of §4.7 and treat T9's fast columns as out of reach T9 §7.8; T2 §7; §4.7 of this document; C1
Q11 Legal person and lawyer. Who is the legal person, and who engages the lawyer on T9 §3.6's five questions, by when? Name the legal person tomorrow; lawyer engaged before wave 0 strangers; wave 0 friends can run on the drafts T9 §3, §7.6; 09-25 plan decision 4
Q12 Beta scale and caps. Tracked-token design point 200 (stretch 500), LLM cap $5/day for the beta rising to T2's $10 + $15/day at 200 tokens, "unavailable" shown when hit? Yes T2 §8 D5, D6; T4 §3.5; C10

Not asked because the reports agree and Vesper can decide: storage option B (T2 §8 D3), retention classes (T2 §8 D9), the six-metric reliability gate (T2 §8 D10, C16), lights as eight fixed cells (T8 Q2), look-list size and sort (T8 Q4, Q5), one skeleton with three profiles (T8 Q6), declared "in" without wallet link and no money P&L (T8 Q3, Q10), alerts to a Telegram bot first (T8 Q9).


9. Suggested order for tomorrow's clean-up session

Read first (about 40 minutes). 1. This file, §1 and §3 (15 min). 2. T1 §0 and §2.5: the eight findings and the three display rules (10 min). 3. T6 §0 and §4: why every number needs a baseline (5 min). 4. T2 §1.3 and §7: what is broken and the infra sequence (5 min). 5. Open wireframes/situation-day2.html and wireframes/situation-hold.html on a phone (5 min).

Decide first (in this order, because each unblocks build work). 1. Q1 displayed price source (unblocks the tape-first plan). 2. Q10 staffing (decides whether the front end starts in parallel on day one). 3. Q6 Solana stream provider (on the critical path by day 3). 4. Q3 and Q4 Telegram route and caller text (unblock intake N4 onward). 5. Q11 legal person and lawyer (long lead time). 6. Q2 and Q5 beta scope and records. 7. Q7, Q8, Q9, Q12 in one pass.

Do the same day, no decision needed. - Bolo starts I1, I2, I3 (T2 §7 items 1–3) and N1, N2 (T4 §7 steps 1–2). - Vesper sends the five vendor emails (T1 §6 rec 9) and upgrades DexPaprika to Dev and Helius to Developer if Q1 and Q6 are yes. - Vesper applies the reconciliations of §3 to the specs Bolo will read: T5's eight lights and band table into scripts/T8-wireframes.py, T6's win definition into Z4, the no-DexScreener rule into C03. - Label the experimentation stack as such in its READMEs (T2 §8 D1, mandate).

Write after the session. A Bolo brief built from §4 of this file (modules, dependencies, specs by report section), and an updated business-plan note replacing the 2 Oct and 9 Oct dates with the §4.7 timeline (T9 §5).


10. What this synthesis did not do

Status: DONE