# Signal methodology (English translation)

> Translation of METHODOLOGY.md for convenience. The Hungarian original (served at /methodology) is the authoritative
> text: its SHA-256 is in every daily fingerprint and is timestamped on Bitcoin. This translation is not
> fingerprinted. If the two differ, the Hungarian text applies.
>
> translation-of-sha256: 4a5a59db095ef8e8ad7ba53c2aadee9f12c3fd376c9a9a9676dc2957321b917d

Version: 2.4.2 (risk flag: r1)
Fixed: 2026-11-15, with the owner's approval (in force: 2026-11-16, from that day's daily run, together with 2.4.0 and 2.4.1)

This file fixes the exact rules of the signal. The daily fingerprint (data/fingerprints) contains the SHA-256 of
this file every day, and the Bitcoin timestamp proves that too. This makes it provable that a rule was fixed before
the result, not adjusted afterwards. The content does not have to be public for the fingerprint to prove this:
anyone can later check whether the published text is identical to the fixed one.

## Versioning rules

- **Major version (X.0.0):** what the signal means changes (new component, threshold, the rule of the tiers). The old
  and the new track record must not be combined; effectiveness is shown per version.
- **Minor version (1.X.0):** coverage grows, the meaning of the signal is unchanged (new protocol, new measured holder
  path, new field).
- **Patch (1.0.X):** bug fix. If it changes a previously published value, the change log says so openly.
- Changes only point forward: a rule must be written here BEFORE it goes live, with the new version number and the
  date it takes effect. Adjusting to past results is forbidden.
- A new component may only make an effectiveness claim after its own live period; its backtested result is labelled
  "past test".
- With a major version the old version runs in parallel for a transition period (after the x402 launch).
- **Separate output (from 2.2.0):** a new output separate from the existing signal (e.g. a risk flag) gets its own
  sequence number (r1, r2, …) and its own track record; it does not affect the version or the track record of the
  existing signal. If a new element changes the selection or the tier of the existing signal, that is still a major
  version.
- **A price increase (from 2.2.0)** is only POSSIBLE if BOTH conditions are met – meeting them allows an increase, it
  does not require one:
  (1) **Demand:** at least **100 returning paying addresses** in the previous 30 days. Returning paying address = an
      address from which payments arrived at the receiving address on at least two different UTC days in the previous
      30 days (x402: the sender of the USDC transfer), measured on-chain. Not counted: own and test addresses. An
      address is not necessarily one agent; the measurable quantity is the address.
  (2) **Evidence:** a B or A verdict on the LIVE track record for the tier concerned, OR materially wider coverage (a
      new minor version that raises the daily median of priced tokens by at least 10%, or a new chain).
  An increase is announced at least 30 days in advance (`price_change_notice` field in every response), with the
  reason; the list of counted addresses is public at the announcement. Price decreases at any time, without notice.
  Layering remains the alternative (the basic data stays cheap).

## Rules of the current version (2.4.2)

### Measurement
- Source: the Ethereum chain only (own Alchemy endpoint, public endpoints as fallback).
- Once a day, at a fixed block (~03:40 UTC), we measure the token QUANTITIES of 11 protocols (Aave v3, SparkLend,
  Compound v3, Morpho Blue, Uniswap v4, Lido, Rocket Pool, ether.fi, Ethena, Sky, Maple).
- Prices (2.1.0): on-chain measured price, per window (6 hours, 2 hours, 1 hour, 30 minutes; the longest one that at
  least half of the candidate liquidity can provide). Sources:
  (a) Uniswap v3 built-in time average; Uniswap v2, SushiSwap and (from 2.2.0) ShibaSwap aggregated price (ShibaSwap
      factory 0x115934131916C8b277DD010Ee02de363c09d037c, verified on-chain 2026-11-02);
  (a2) (from 2.2.0) Aero concentrated-liquidity (CL) pools: factory 0xf4fd2e6c5b60c5b365cd1b96dd6245e6d60414c2
      (verified on-chain 2026-11-02); as with v3, pools are queried from the factory per token pair
      (`getPool(a, b, tickSpacing)`), for every tickSpacing value enabled by the factory at the measurement block
      (`tickSpacings()`); a pool is identified by the token pair and the tickSpacing; price from the built-in time
      average (`observe`, identical to v3, with the same sampled completion), depth from the active liquidity,
      crossing tick boundaries (liquidity staked in the gauge is part of it); fee filter: the pool's current fee
      (`fee()`) ≤ 1%;
  (b) Uniswap v4: only pools WITHOUT a hook (a hook can implement its own pricing); pools come from a registry built
      from the PoolManager's creation (Initialize) events (data/pool_registry.csv), only those created at least
      1,800 blocks (6 hours) before the measurement, and (stated from 2.4.1) with a fee of at most 1%;
  (c) sampled time average where there is no built-in one (v4), or it does not reach back for all four windows (v3,
      few stored observations): the last 6 hours are split into 48 equal slices, one block is drawn per slice from
      the hash of the measurement block, and per window the MEDIAN of the sampled prices is the price. (Exactly, from
      2.4.1, so anyone can recompute it:) B = the measurement block, i = 0…47; slice i: [B − 1,800 + 1 + ⌊i · 1,800 / 48⌋,
      B − 1,800 + ⌊(i + 1) · 1,800 / 48⌋]; the block within the slice: the lower bound + (the first 8 bytes of
      SHA-256(the 32-byte hash of block B ‖ i as 4 bytes, big-endian) as a big-endian integer) mod the slice length
      (the number of blocks between the two bounds). Samples per window: 6 hours: 48, 2 hours: 16, 1 hour: 8,
      30 minutes: 4. A window is only valid if at
      least 75% of its samples are present. The sample does not override the built-in time average.
  Common filters for every pool: ±2% active depth of at least $500 per pool and $1,000 per token; the price of a pool
  last traded more than 12 hours ago cannot be used (in v4: if its price has not changed for 12 hours); a pool that
  deviates by more than 3% from the liquidity-weighted median drops out; if the remaining pools do not provide 2/3 of
  the liquidity, there is no price.
  Liquidity grade: strong (≥ $50k ±2% depth) / medium (≥ $10k) / weak. (From 2.4.0) The grade and the published ±2%
  capacity come from the token's TOTAL depth: the depth of its price pools + the contribution of its supplementary pools.
  Supplementary pool: a pool of the token on the DEXes above (Uniswap v3, Aero CL, Uniswap v2, SushiSwap, ShibaSwap, v4
  without hooks) whose other side is also a token on our list; both tokens are priced that day from our own pools (time
  average or sampling – not a rate, vault, 1:1, Curve or definition); neither side is a quote (USDC, USDT, WETH / native
  ETH, DAI, USDS – their pairs are the other token's price pools); and the pool is not used in the token's price (a token
  priced in round 2 has its cross-quote pool in its price, so for that token it is not supplementary). Filters: fee ≤ 1%;
  v4: without hooks, existing for at least 1,800 blocks; active – its price (sqrtPrice) changed within 12 hours (v2: the
  reserves were updated within 12 hours); the pool's spot price deviates by at most 3% from the ratio of the two tokens'
  own prices; a contribution of at least $500. Contribution: the same ±2% calculation as for a price pool (crossing tick
  boundaries), the SMALLER of the two directions, each side at its OWN price of the day in USD; the same number for both
  tokens. So no token gets more than its own side, and the sum given to the two tokens is at most the pool's ±2% value
  (the quality gate checks this per pool). An unreadable pool does not count (so the depth never overstates); if the
  whole pass fails, the supplement is "not measured" that day, depth and grade come from the price pools, and the
  quality gate warns. Supplementary pools do not affect the price, the price's $1,000 minimum or cross-quote eligibility.
  Per-pool table (with reasons for excluded pools): data/depth_extra_latest.csv.
  Quotes: round 1 – USDC (= $1), USDT, WETH (in v4 also native ETH, 1:1), DAI, USDS. Round 2 (cross quote) – tokens
  graded "strong" or "medium" in round 1 (from the depth of their price pools); their price stays the round-1 price, the other tokens get a price from the
  direct and cross pools together, with the quote's price for the same window. One step (a token priced via a cross
  price is not a quote). (From 2.3.3) A pool counts at most once in a token's price and depth (by pool address), even
  if it is found both as a direct and as a cross-quote pool. (Stated from 2.4.1) The quotes' own prices come from pools
  with a built-in time average (Uniswap v3, Aero CL, v2, SushiSwap, ShibaSwap), without v4 and sampling: USDT against
  USDC, WETH and DAI against USDC and USDT; as a quote, USDS gets DAI's price for the same window (1:1 by protocol).
  Fallback when there is no market price: protocol rate (LSTs; from 2.1.0 also Lido wstETH: stEthPerToken × ETH
  price), Rocket Pool oDAO price, Curve oracle ((stated from 2.4.1) two-coin pools only; its depth is 10% of the USD
  value of the smaller side), ERC-4626 vault rate, 1:1 conversion defined by the protocol. (Made precise
  in 2.3.3) A fallback price may be used only if the token has no market pool after the filters (missing or too thin
  liquidity). If there are market pools but their prices disagree (the pools agreeing with the median do not reach 2/3
  of the liquidity: DISAGREEMENT), the token gets NO fallback price that day: the price is "not measured" and the quality
  gate warns. A fallback replaces a missing market, it may not override a contradicting one. A token with a protocol
  price or a weak grade is not a member of the index.
  (Stated from 2.4.2) A network error does not change the measurement. If every endpoint fails for a required query
  (also after retries), the affected token is "not measured" that day, together with every token whose measurement
  builds on it (quote, cross-quote, the underlying of a derivation, Curve quote): no price is given from fewer pools,
  fewer samples, without the quote or by a fallback method (a fallback replaces a MISSING market; a network error is not
  that). If a round-1 quote (USDT, WETH, DAI) cannot be measured, the day's pricing is not produced (the daily run is to
  be repeated). A sample may be missing only as the chain's answer (the pool did not exist yet at that block), never
  because of a network error. Next to a token that failed on a network error, the supplementary depth is "not
  measured" that day. A back-test may not contain a day not measured because of a network error: we measure it again.
  Without errors the result is byte-identical, whichever endpoint answered.
  Coverage limit (not watched): Balancer, Fraxswap and other DEXes, v4 pools with hooks, (stated from 2.2.0)
  oracle-based AMMs (the pool price follows an external price reference, e.g. Metric), pairs whose other side is not
  on our own asset list.

### Components
1. **Capital inflow (price-neutral):** the change in quantity valued at a constant (today's) price, in 1/7/14-day
   windows, per protocol. Subset vaults (Ethena sUSDe, Sky sUSDS) are not counted in the protocol total.
2. **Own altcoin index:** equal-weight, daily chained index of the watchlist tokens with role "signal" and "index";
   a token with weak liquidity is not a member.
3. **Relative strength (RS):** the token's N-day return minus the index's N-day return, in percentage points.
4. **Flow–price gap:** gap_pp = 7-day net inflow % − 7-day RS pp.
5. **Net holder value (USD/year):** burn + buyback + payouts to holders, annualised from a 30-day window, minus
   dilution. Dilution has two parts:
   (a) emission: only from verified emission (VERIFIED_EMISSIONS, currently empty); supply did not grow → 0; supply
       grew for an unknown reason → not measured;
   (b) treasury / vesting outflow: tokens sent from the protocol's on-chain verified treasury addresses
       (TREASURY_HOLDERS) to another address (not between themselves, not to a burn) in the last 12 months, without
       scaling, at today's price. (From 2.3.0) A **verified protocol-controlled vault** also counts as part of the
       treasury (e.g. a delegation contract owned by the treasury; `data/protocol_vaults.csv`, with evidence and the
       date of on-chain verification): tokens going there from the treasury are not dilution, tokens going from
       there into circulation are. Until a vault is verified, tokens going there from the treasury are dilution.
       (From 2.3.4) Tokens going from the treasury to a **verified holder-reward distributor** are not dilution
       (`data/holder_reward_distributors.csv`: token, distributor address, holder reward contract address, evidence,
       date of on-chain verification; the distributor is named by the protocol's own on-chain registry, and it is
       verified on-chain that it forwards the tokens to the reward contract). Two limits, so that issued or vesting
       tokens cannot be booked as "rewards": (1) at most the amount that reached the treasury from outside (market
       purchases) in the same 12 months; (2) whatever the distributor sends anywhere other than the registered reward
       contract is dilution. Reason: the value of tokens bought back and then distributed to holders already counts as
       holder value at the buyback; counting their distribution as dilution would deduct it twice. Without an entry,
       tokens going there from the treasury are dilution; with a faulty registry, dilution is not measured.
       Without a verified treasury list, or on a measurement error, dilution is not measured → net holder value is
       not measured.
   **Holder paths (from 2.3.0).** Only a registered, verified path counts as holder value (`data/holder_paths.csv`:
   token, type – burn / buyback / holder payout –, addresses, source of the evidence, date of on-chain verification,
   state, expected rhythm; a row without evidence cannot be added). Every watched token with a documented, on-chain
   traceable path is added – regardless of whether this improves or worsens its tier.
   (1) A purchase counts when the token first arrives FROM OUTSIDE at a verified protocol address of the path
       (executor, buyback Safe, treasury); further movement between known protocol addresses does not count again
       (no double counting), a token arriving from 0x0 (minted) is not a purchase. 30 days, annualised, at today's
       price. The end point of the path must verifiably be holder value: burn, holding in the protocol treasury, or
       a payout to the token's holders.
   (2) Not counted: if the bought token goes back to users (cashback, incentive, reward to non-holders), or the end
       point cannot be traced (another chain); then the token's holder value is "not measured", with the reason.
   (3) Burn: only a protocol-linked burn (from the path's address to 0x0 / 0x…dead); an unexplained decrease in supply
       is "not measured". (From 2.3.3) UNI: only burns of the fee switch count – (a) the mainnet Firepit burn (recognised
       by the mainnet TokenJar releasing assets in the same transaction), (b) a canonical bridge withdrawal to 0x…dead
       whose L2-side sender is a registered L2 releaser (`data/uni_burn_routes.csv`, from Uniswap's official address
       list, with evidence and the date of on-chain verification). UNI reaching 0x…dead in any other way is not holder
       value: it is shown separately for information (`other_burn_30d`, token count, not annualised) and does not count
       for the tier. If the sum of the burns does not equal the increase of the dead balance, UNI holder value is "not
       measured".
   (4) Holder payout in another asset (e.g. ETH to stakers): at the price measured on the day of the payout,
       annualised from a 30-day window, only on a registered, verified path.
   (5) Silence: if a path has been silent for longer than twice its expected rhythm, and the daily path finder finds a
       linked candidate for the same token → "not measured (possible move)" + "INTERVENTION REQUIRED!"; without a
       linked candidate the 30-day measured amount counts, and if that is 0 → verified 0 ("paused"). On a path that is
       not registered, a zero found is still not measured (not 0). (From 2.3.2) If the path finder did not run that day,
       it is unknown whether a linked candidate exists: the silent path is "not measured (the path finder did not run
       today)" that day, not verified 0 – unknown is not "none".
   (6) Linked candidate (on-chain): bought from a DEX on at least 2 different UTC days, for at least $25,000, sold at
       most 10% of what it bought, and its means of payment comes 1 step back from a known protocol address, OR the
       bought token goes 1 step forward to an address of the path / to a burn; supporting sign: contract / Safe. A
       regular buyer without a link is information only. (From 2.3.3) The filter is applied to every buyer meeting the
       quantity condition (≥ 2 UTC days, ≥ $25,000), at most 60 buyers per token; transfers are paged up to 20,000 per
       source. If the check is incomplete (more than 60 candidates, or the paging limit reached), it is unknown that day
       whether a linked candidate exists: the silent path is "not measured (the path finder's check was incomplete)",
       not verified 0.
   For an executor purchase the price may deviate from the market by at most ±20%; the price is computed from the
   executor's paired trades (the i-th token receipt with the i-th subsequent stETH payment, with a 60-day look-back).
   If the check fails or cannot be done, holder value is not measured that day (warning), not an error.
   A measurement error in the components of holder value is a warning (not measured); it does not stop publication.

### Candidates and tiers
- (Stated explicitly from 2.3.3) Eligibility: only a token with role "signal", a market price and a non-weak liquidity
  grade (≥ $10k within ±2%) can get a tier and r1. The other tokens' rows remain (the data is measured), but with
  `flagged = 0`, `risk_flag = 0` and the reason in `tier_reason` ("not eligible: weak liquidity" / "not eligible:
  protocol rate price"). The random selection of the backtest and of the live track record draws from the same
  eligible set.
- Signal: 7-day net inflow ≥ 2% AND gap_pp ≥ 5. (From 2.4.1) The thresholds are compared with the published values
  rounded to two decimals (flow_7d_pct, gap_pp) – a published gap of "5.00" always reaches the threshold. Within this,
  three tiers:
- **3 – strong:** net holder value > 0 (holders receive more value than the dilution that hits them).
- **2 – candidate (gross value creator):** gross holder value > 0, but net ≤ 0 or not measured.
- **1 – watch:** no measured positive holder value (only capital movement and price divergence).
- Every signal comes with a text reason (`tier_reason`): the gross, the outflow and the net value with numbers.
- The backtest evaluates every tier SEPARATELY; for a tier that does not prove itself (does not beat the index and the
  random selection) we do not claim it is a signal, it stays informational data only.
- Supplementary (non-filter) fields: positive days out of 7, share of the largest contributing asset, 7-day change in
  supply (> 1% = dilution warning), capacity in USD for a ±2% price impact.

### Risk flag: outflow–price gap (r1, separate output, from 2.2.0)
- Descriptive, measured state: "capital is flowing out of the protocol, the token price has not yet followed relative
  to the index". **Not a sell signal, not advice.** Forbidden words in the output: sell, short, exit, target, buy.
  Field: `outflow_price_gap`.
- Trigger (the same inputs as the signal, (from 2.4.1) likewise with the published values rounded to two decimals):
  7-day net inflow ≤ −2% AND gap_pp ≤ −5. Same eligibility as the signal
  (only tokens with role "signal", not weak liquidity, with a market price).
- **R2 – aggravated:** the trigger AND (net holder value < 0 OR 7-day change in supply > 1%).
  **R1 – basic:** the trigger, without an aggravating factor. "Not measured" does not aggravate.
- Accompanying fields (not filters): `risk_reason` with numbers; `concentrated` = in a multi-asset protocol a single
  asset provides ≥ 70% of the outflow; positive days out of 7; capacity for a ±2% price impact.
- Proof: an effectiveness claim only after its own live period (at least 60 days AND at least 20 flags per tier).
  Mirrored test: whether the 7/14/28-day relative return of the flagged tokens stays further BELOW the index than
  that of 2,000 random selections; the A/B/C thresholds are the same as for the signal (with the opposite direction).
  A past test only with the "past test" label, it must not be combined with the live track record. With a C verdict it
  stays informational data.
- r1 does NOT change the selection or the tiers of the signal (2.2.0).

### Publication
- Only data that passed the quality gate (ok / warn) is published; on fail, yesterday's data stays.
- An extraordinary move confirmed by an independent source may be published at warn level, with the
  `extraordinary_move` flag.

## The rule of the backtest (proof)

Fixed on 2026-10-19, BEFORE the result on the back-filled data was seen (code: collector/backtest.py).
- 12 months: the first 8 are training, the last 4 the test segment; no parameter is tuned.
- For every day the signal is computed by the live code, only from the data known by then. Horizons: 7, 14 (main) and
  28 days.
- Prices (2.1.0): for the whole period uniformly recomputed with the current rules, in a separate file
  (data/reprice/prices.csv, each day at that day's measurement block); the live, timestamped data/prices.csv does not
  change.
- Dropped signals (2.1.0): if there was a price on the flagged day but none at the end of the horizon, the
  observation is left out (we do not estimate), but their number is printed per tier and horizon as "X / Y", with
  the most frequent tokens.
- Benchmark: relative return (the token minus our own altcoin index) and absolute return (cash = 0); the random
  selection = the average of all signal tokens measurable that day.
- **Placebo test:** in the test segment the signal is compared with 2,000 "blindfolded" selections (on every flagged
  day the same number of randomly chosen tokens). We measure four tiers, so on the main (14-day) horizon the signal
  must beat at least 1,976 of the 2,000 (at most 24 random selections may reach or exceed it; 1.25%), on the 7- and
  28-day horizons at least 1,901 (95%). Computation: p = (number of random selections reaching/exceeding + 1) / 2001;
  the threshold is p < 1.25% and p < 5% respectively.
- **(A) works:** on every horizon the mean relative return is > 0 and > that of the random selection, in the majority
  of months as well; the signal beats at least 1,976 of the 2,000 random selections on the main horizon and at least
  1,901 on the 7- and 28-day horizons; and over the full 12 months the relative return is >= 0 in both market regimes
  (index above / below its 60-day average); at least 20 observations.
- **(B) partly:** the condition is met on at least one horizon, and there the signal beats at least 1,976 of the 2,000.
  A narrower claim; to be sold only after confirmation by the live track record.
- **(C):** otherwise. C means: with this much data it cannot be distinguished from luck – NOT that the signal is
  certainly worthless (4 months is little to detect a smaller edge); the live track record keeps measuring.
- Test of the rule on synthetic data (tests/test_backtest.py): a signal that knows nothing and picks by coin toss got
  A in 0% and B in 1.5–5% of 200 trials (with the old, pre-tightening rule 39–52% B).

## The verdict on the live track record (from 2.3.1)

Fixed on 2026-11-05, BEFORE the first live verdict was computed (code: collector/track_record.py).
- The live track record: for every day the signal as we published it that day (only a file byte-identical to the
  daily fingerprint, timestamped on Bitcoin), the outcome from the live measured prices (data/prices.csv) and our own
  altcoin index (data/altindex.csv). A signal or price recomputed afterwards does not count.
- Only days of the current major version count (today 2.x, from 2026-10-18); days of minor versions and patches are
  added together (the meaning of the signal is the same), days of an earlier major version are not.
- The whole live period is the test segment (no training part). The tests and thresholds are the same as in the
  backtest: at least 20 observations per horizon, mean relative return > 0 and > that of the random selection, in the
  majority of months as well; placebo test with 2,000 random selections (at least 1,976 beaten on the main 14-day
  horizon, at least 1,901 on the 7- and 28-day horizons); for (A) the relative return is >= 0 in both market regimes
  over the whole live period. (A), (B), (C) mean the same.
- A verdict only after at least 60 live days (published, byte-identical days in the current major version); until
  then "not evaluable yet", with descriptive numbers only.
- Recomputed daily for the whole live period, so the verdict can change (B can become C); we always publish the
  current verdict (/performance and the `track_record` block of the responses).
- The evidence condition of a price increase ("a B or A verdict on the LIVE track record") refers to this verdict.

## Change log

| Version | Fixed | In force | Change |
| --- | --- | --- | --- |
| 2.4.2 | 2026-11-15 | 2026-11-16 | Fix (the code aligned with the fixed text; the measurement rules, the meaning of the signal, the thresholds and the tiers are unchanged; in force together with 2.4.0 and 2.4.1): a network (RPC) error does not change the measurement (see the "stated from 2.4.2" paragraph). When every endpoint failed for a query, the old code did not stop but gave a price from fewer pools or samples, without the quote, or by a fallback method (Curve, vault, rate) – a price that depended on the endpoint and could not be recomputed later; this contradicted the fallback rule ("only if the token has no market pool after the filters") and the recomputability of sampling. Discovery: the first stage of the 2.4.0 back-test (123 days, heavy parallel load) gave a price different from 2.3.4 on 2,449 token-days (e.g. 2025-10-27: the USDT measurement failed, WETH was priced without the USDT pools, from 5 pools instead of 10, 4,170.8 $ instead of 4,200.9 $); that stage was deleted and the back-test runs again. Openly: in the live, published data (43 days) and in the earlier back-tests there were 0 quote failures and 0 network-error rows; a silently dropped sample leaves no trace in the data – the corrected repricing compares all 386 days with the 2.3.4 prices. On-chain, the corrected code reproduced the 2.3.4 prices byte for byte on 12 days (also under load); under load 0 silently different days (a day with a network error is loudly marked for re-measurement). Endpoints (measured 2026-11-15): flashbots removed (403 on eth_call everywhere); backfill starts with the public archive endpoint, the own endpoint is only a fallback, up to the safe remainder of the monthly quota. Plan, failure modes (K1–K15), on-chain diagnostics: terv/rpc-kieses-terv.md. |
| 2.4.1 | 2026-11-15 | 2026-11-16 | Fix (the meaning of the signal, the price, the thresholds and the tier rules are unchanged; in force together with 2.4.0): (B12) the thresholds are compared with the published values rounded to two decimals – until now the unrounded floating-point value decided, so a gap published as "5.00" gave no signal because of a floating-point remainder (e.g. 8.04 − 3.04 = 4.999…), and the published number contradicted the flagged field; the same for r1 (−2%, −5 pp). (B13) the `concentrated` companion field measures, as the text says, the concentration of the OUTFLOW (the largest outflowing asset's share of the outflow); until now it used the share of the total – in- and outflowing – movement (in the 2.3.4 back-test 18 of 324 r1 rows change, all 0 -> 1; the signal and the r1 level do not change). (A5) the text made precise to match the code (the numbers do not change): v4 pool fee at most 1%; quote prices without v4 and sampling; the exact sample-block formula; the Curve oracle uses two-coin pools only, with a depth of 10% of the smaller side. (B15) the code aligned with the 2.3.0 rule: without a registered holder path a zero found is "not measured" – MORPHO's holder value was "0" until now ("no fee recipient"), from now on it is "not measured", and the reason keeps the fact verified on-chain (the fee switch is off); the tier does not change. Openly: previously published days show "0" (fingerprinted files are unchanged). |
| 2.4.0 | 2026-11-15 | 2026-11-16 | Minor version (coverage grows; the meaning of the signal, the price, the thresholds and the tier rules are unchanged): supplementary depth – pools between two listed tokens priced from our own pools, neither of them a quote, that the token's price does not use (e.g. stable–stable, BTC–BTC, ETH–ETH, two protocol tokens), now count in the token's ±2% depth (and so in its liquidity grade and published capacity). A pool gives both tokens the smaller direction, each side at its own price of the day: no token gets more than its own side, and the sum given to the two tokens is at most the pool's ±2% value (quality-gate guard). The price, its $1,000 minimum and cross-quote eligibility are unchanged. The owner's decision (2026-11-15): the "cautious variant" – instead of the draft of 2026-11-08 (a pool gives only its own side, with any counterpart), because finding off-list pairs would need a new registry and one-sided liquidity would inflate the depth. New prices.csv columns: depth_price_usd (the depth_usd before 2.4.0), depth_extra_usd, extra_pools. Openly: the depth and grade of previously published days do not change (fingerprinted files are unchanged); the back-test ("past test") with the 2.4.0 rule runs in a separate folder (reprice-2.4.0). |
| 2.3.4 | 2026-11-12 | 2026-11-13 | Fix (item T6 of the code review of 2026-11-08; the meaning of the signal, the thresholds and the tier rules are unchanged): tokens going from the treasury to a verified holder-reward distributor are not dilution, up to the amount that reached the treasury from outside in 12 months, and only if the distributor forwards to the registered reward contract (data/holder_reward_distributors.csv; first entry: SKY REWARDS_DIST_LSSKY_SKY -> REWARDS_LSSKY_SKY, from the Sky CHAINLOG, verified on-chain). The old calculation counted SKY that was bought back and then distributed to stakers twice (as holder value and as dilution): on 2026-11-12, 30,783,294 SKY / 30 days. Openly: the fingerprinted files of days published up to 2.3.3 do not change (SKY's net is stricter there), the live track record counts the published version; the major version does not change. Then recomputation of fundamentals and a new backtest ("past test"). |
| 2.3.3 | 2026-11-09 | 2026-11-10 | Fix (the code aligned with the fixed text; the meaning of the signal, the thresholds and the tier rule are unchanged; items T1–T5 of the code review of 2026-11-08): (T1) a pool counts once per token – the old code could count a pool both as a direct and as a cross-quote pool in round 2; (T2) no fallback price on a market DISAGREEMENT – the old code never switched this protection on, because the error text started with the exception name; (T3) the path finder checks every quantity candidate (at most 60, up to 20,000 transfers per source), an incomplete check is "not measured" – the old code checked only the first 6 buyers by quantity and could derive "verified 0" from that (AAVE Nov 5, 7, 8: 6 candidates, all traders); (T4) only eligible tokens (market price, non-weak liquidity) get a tier and r1, and the random selection draws from the same set – the old code also gave a tier to the weak-liquidity SPK (Nov 2–4: 1, Nov 5: 2); (T5) UNI: only fee-switch burns are holder value (mainnet Firepit + canonical bridge with a registered L2 releaser), UNI burned in other ways is `other_burn_30d` – the old code counted the whole increase of the dead balance (on Nov 9: 1,114,000 UNI / 30 days, of which verified 1,040,000, via Wormhole 74,000, foreign 0). Openly: the fingerprinted files of days published up to 2.3.2 and affected by this (SPK tier Nov 2–5; the AAVE "verified 0" days; the UNI burn) do not change; the live track record counts the published (first) version; the major version does not change (the live track record count continues). Then repricing and a new backtest ("historical test"). |
| 2.3.2 | 2026-11-06 | 2026-11-07 | Fix: if the path finder did not run that day, a silent holder path is "not measured (the path finder did not run today)", not verified 0 (until now the code treated missing path-finder data as "no linked candidate"). Openly: on 2026-11-06 the path finder did not run in the daily run because of a transient error, and AAVE's paused path was published as "verified 0"; rerun the same day at 07:35 UTC, the path finder found 0 linked candidates, so the published 0 turned out to be correct. The published file does not change. On back-filled days (without path-finder data) the documented limitation of 2.3.0 remains. |
| 2.3.1 | 2026-11-05 | 2026-11-06 | Addition: the rule of the verdict on the live track record fixed, before the first live verdict – the same tests and thresholds as the backtest, the whole live period is the test segment, only days of the current major version, at least 60 live days, recomputed daily. The computation of the signal does not change. |
| 2.3.0 | 2026-11-04 | 2026-11-05 | Minor version (coverage grows, the rule of the tiers is the same): holder value from the registered, verified paths (data/holder_paths.csv) – first arrival at the protocol's address, without double counting; silent path + linked candidate = not measured (possible move); paused, verified path = verified 0; distributed token = not measured. For dilution a verified protocol-controlled vault is part of the treasury (data/protocol_vaults.csv; empty today). Then full repricing and a new backtest ("past test"). Limitation: for back-filled days the registry of 2026-11-04 applies, and there is no old linked-candidate data; the backtest prints this. |
| 2.2.0 | 2026-11-02 | 2026-11-03 | (sequence number of the risk flag: r1, from the same day) Minor version (coverage grows, the meaning of the signal is the same): Aero CL pools (factory registry, built-in time average) and ShibaSwap (v2 factory); oracle-based AMMs stated in the limitation list. The versioning rule is extended: a separate output gets its own sequence number (r1); new output: outflow–price gap risk flag (r1, R2/R1), with its own track record. The condition for a price increase is extended with demand (100 returning paying addresses, with an exact definition) – the conditions only allow an increase. Then full repricing and a new backtest ("past test"); the 2.1.0 backtest stays published unchanged. |
| 2.1.0 | 2026-10-20 | 2026-10-21 | Minor version (coverage grows, the meaning of the signal is the same): Uniswap v4 pools without hooks (registry derived from Initialize events); sampled time average (median of 48 blocks drawn from the block hash) for pools without a built-in time average or with a short history; cross quote (tokens strong/medium in round 1, one step); wstETH protocol rate. The backtest runs on uniformly repriced prices (data/reprice/prices.csv) and prints the number of dropped signals. The live, timestamped data/prices.csv does not change. |
| 2.0.1 | 2026-10-19 | 2026-10-19 | Addition: the rule of the backtest (with a placebo test) fixed before the result was seen. The computation of the signal does not change. |
| 2.0.0 | 2026-10-17 | 2026-10-18 | Major version (the rule of the tiers changes): three tiers – 3 strong (net > 0), 2 candidate (gross > 0, net ≤ 0 or not measured), 1 watch; a reason with every signal; the backtest separately per tier. Contains the measurement rules of 1.1.0 unchanged; 1.1.0 had no live day (it would have taken effect on the same day). The back-fill and the backtest of Nov 2 measure 2.0.0. |
| 1.1.0 | 2026-10-17 | 2026-10-18 | Minor version (coverage grows): treasury/vesting outflow is added to dilution (12 months, on-chain verified addresses: Uniswap Timelock, Sky MCD_PAUSE_PROXY, Lido Agent); without a treasury list net holder value is not measured. New watched path: ether.fi buyback wallet. An error in the components of holder value is a warning. The back-fill and the backtest of Nov 2 measure this version (1.0.x had no live track record). |
| 1.0.1 | 2026-10-16 | 2026-10-17 | Fix: the executor price check from paired trades, with a 60-day look-back (payment follows receipt by days; on Oct 16 a delayed payout gave a false −53%). A failed check = not measured + warning, not a stop of the whole publication. A previously published value does not change. |
| 1.0.0 | 2026-10-16 | 2026-10-16 | First fixed version. The backtest of Nov 2 measures this version. |
