# Evidence behind the README

What the [README](README.md) leaves out, kept word for word as it stood before the README was shortened on 13 November, under the same headings. Every marked figure is checked against [`data/numbers.json`](data/numbers.json) by `python scripts/rebuild.py`.

## Why Hold

The table with every figure:

| Question | Answer | The number |
|---|---|---|
| How much of the new NEAR do the buybacks offset? | A small part | <!--n:n1.coverage_buybacks-->10.8%<!--/n--> of it, counted in NEAR (95% interval <!--n:ci.coverage_buybacks.low-->9.5%<!--/n-->–<!--n:ci.coverage_buybacks.high-->12.4%<!--/n-->): <!--n:iv.buybacks_kept_usd_per_year-->$17.1M<!--/n--> a year kept, against <!--n:iv.issuance_usd_per_year-->$157.5M<!--/n--> of NEAR issued, both at the window's mean price of <!--n:n2.mean_price_usd-->$4.88<!--/n-->. At <!--n:peer.near.price_usd-->$3.71<!--/n--> (6 November) both dollar figures fall by the same share; the <!--n:n1.coverage_buybacks-->10.8%<!--/n--> does not change |
| Does a rule guarantee it? | No: it is a choice | nothing locks or burns it, and <!--n:c.buybacks_confirmations-->2 of 4<!--/n--> signers can move it |
| Is the demand real? | Yes, but most fees stay with the apps that route swaps | users paid <!--n:n2.fees_usd_per_day-->$206,049<!--/n--> a day in fees (DefiLlama), and successful relay calls into NEAR Intents (one can carry several intents) went from <!--n:bq.paths.october.relay.ok_a_day-->45,918<!--/n--> a day in October to <!--n:bq.paths.november.relay.ok_a_day-->103,267<!--/n--> on 1–11 November (BigQuery) |
| Are holders paid more elsewhere? | Yes | on DefiLlama's count of holders' revenue for both (for NEAR all three revenue wallets, so more than the buybacks keep), NEAR holders get <!--n:peer.near.holders_yield_on_mcap-->0.40%<!--/n--> of NEAR's market value a year; Hyperliquid's is <!--n:peer.hyperliquid.holders_yield_on_mcap-->4.09%<!--/n--> of its circulating value (<!--n:peer.hyperliquid.holders_yield_on_fdv-->1.03%<!--/n--> fully diluted), and the tokens it buys sit at an address with no key ([The Block](https://www.theblock.co/news/defi/2025-12-17-hyper-foundation-proposes-removing-1-billion-hype-tokens-from-circulating-supply-by-zeroing-out-hyperliquids-assistance-fund-383091)) |
| Is the chain's own use growing? | No, except NEAR Intents | daily transactions not relayed <!--n:u.txns_not_meta.change-->-58%<!--/n--> in a year (<!--n:u.txns.change-->-83%<!--/n--> with relayed ones, which fell in one step in April, cause not identified); all its gas paid <!--n:u.fee_usd_year-->$1.2M<!--/n--> a year |

**What would change the rating, in detail**

- **Buy**, all of: (1) a lock declared in `daily.LOCKS` (the author adds it by hand once a published rule names one) that the chain shows cannot send NEAR out; the code checks the lock, not how much NEAR is in it; (2) `treasury.near` opens no new multisig request in the window (a request opened earlier and confirmed inside it would not be seen); (3) counted coverage of 100%, where the buybacks, DefiLlama's revenue and DefiLlama's fees must each reach issuance (the buybacks need <!--n:iv.wallets_multiple_buybacks-->×9.2<!--/n--> today's pace, all fees <!--n:iv.fees_multiple-->×2.1<!--/n-->), or at the yearly check each 1 November growth of <!--n:lr.required_growth.buybacks.A.2031-->+75%<!--/n--> a year (<!--n:lr.required_growth.buybacks.B.2031-->+56%<!--/n--> once HSP-029 is in force, a switch the author also sets by hand).
- **Avoid, at once:** the amount of NEAR the buybacks hold (an amount, not a price; NEAR they delegate to a staking pool also counts as gone, although they would still own it) falls more than 10% below its peak since 1 November. This is the drain test; the dashboard runs it live from the chain. [All the rules](#the-rating-as-rules).

## Also measured

- **It did not stop when swaps went private.** From 27 October to 8 November (13 days, running past the window's end), when swaps began to settle privately, the NEAR kept a day by the three revenue wallets (the two buybacks wallets and the front-end fund) changed by <!--n:ps.change.kept_near_per_day-->-3%<!--/n-->. That cannot be told from no change (<!--n:ci.shift.kept_near.low-->-28%<!--/n--> to <!--n:ci.shift.kept_near.high-->+28%<!--/n-->), but a stop would show as about -100% ([claim D](#d-it-did-not-stop-when-the-flow-went-private)).
- **The front-end fund**, a third wallet that keeps the fees of NEAR Intents' own front end, raises coverage to <!--n:n1.coverage_with_front_end-->16.5%<!--/n-->; the Buy rule counts the two buybacks wallets' <!--n:n1.coverage_buybacks-->10.8%<!--/n-->.
- **People decide issuance too**, but only through a release that block producers with more than <!--n:c.issuance_upgrade_stake_threshold-->4/5<!--/n--> of their stake run.

## What to watch from Monday, 16 November

- The bought NEAR, every day, on the [dashboard](https://nikolaskahris-png.github.io/near-thesis-data/dashboard/): a drain turns the rating to Avoid.
- The House of Stake draft of 8 November: does it get a number and a vote?
- HSP-029, the issuance cut, leaves its timelock at <!--n:f8.timelock_end_utc-->2026-11-16 16:48 UTC<!--/n-->; until then House of Stake's Security Council can still veto it ([gov.near.org](https://gov.near.org/t/hsp-002-hsp-003-next-steps/41791): after a vote the council may veto on security grounds). It cuts issuance only once validators run a release.

## The thesis in five claims

Every marked figure comes from [`data/numbers.json`](data/numbers.json); a check fails if one differs. Claims A–D with every source: [long version](docs/DETAILS.md#v5-the-thesis-in-four-claims); claim E is complete here.

### A. The demand is real

NEAR Intents moved <!--n:d.ours_30d_usd-->$4.02B<!--/n--> in DefiLlama's latest 30 days, to <!--n:d.window_end-->7 November<!--/n-->, <!--n:d.parent_rank-->15th<!--/n--> of <!--n:d.parent_groups_with_volume-->885<!--/n--> protocols by DEX volume on DefiLlama. Users paid <!--n:n2.fees_usd_per_day-->$206,049<!--/n--> a day in fees in the window, most of it to the apps that route swaps.

**Against.** Volume and fees are DefiLlama's; NearBlocks, indexing the chain, sees <!--n:u.visible.before_private-->66%<!--/n--> of that volume before the private path ([claim E](#e-the-chain-is-used-less-and-its-use-pays-holders-almost-nothing)). Coverage and the rating use fees and wallet balances, not volume.

**Check.** `python scripts/demand.py --verify` → `OK`.

### B. It is small

Each day the buybacks wallets keep <!--n:n1.buybacks_near_per_day-->9,570<!--/n--> NEAR, against <!--n:n1.new_near_per_day-->88,392<!--/n--> new NEAR issued: coverage <!--n:n1.coverage_buybacks-->10.8%<!--/n-->. Resampled by week, it lies between <!--n:ci.coverage_buybacks.low-->9.5%<!--/n--> and <!--n:ci.coverage_buybacks.high-->12.4%<!--/n--> (95%); by month since April, <!--n:f1.coverage_buybacks.2026-04-->5.3%<!--/n--> to <!--n:f1.coverage_buybacks.2026-09-->15.3%<!--/n-->. If NEAR kept every fee, coverage would be <!--n:n2.max_coverage-->47.9%<!--/n-->; fees alone would cover issuance only at a NEAR price of <!--n:n2.equilibrium_price_usd-->$2.33<!--/n--> or less. New NEAR adds <!--n:iv.issuance_rate_per_year-->2.46%<!--/n--> to supply a year.

**Against.** Coverage is in NEAR and revenue in dollars, so coverage falls when the price rises; in dollars the buybacks kept <!--n:lr.history.2026-04_2026-10.buybacks_kept_usd_multiple-->×7.5<!--/n--> as much a day in October as in April. And most new NEAR goes to stakers: <!--n:c.treasury_share_of_new_supply-->10.1%<!--/n--> goes to `treasury.near`, and the buybacks keep <!--n:iv.buybacks_over_treasury_issuance-->107%<!--/n--> of that slice. So a staked holder is diluted only by that slice and the validator's commission; coverage measures what the buybacks give back against all new NEAR.

**Check.** `python scripts/daily.py --verify`; `python scripts/coverage.py --verify`; `python scripts/intervals.py --verify`. Each ends with `OK`.

### C. It is a policy, not a rule

Nothing in code burns, locks or requires the buybacks. The three revenue wallets' <!--n:n1.wallets_held_near-->4,109,089<!--/n--> NEAR moves with <!--n:c.buybacks_confirmations-->2 of 4<!--/n--> signers for each buybacks wallet and <!--n:c.front_end_transfer-->2 of 5<!--/n--> for the front-end fund, where <!--n:c.front_end_rules-->1 of 5<!--/n--> can rewrite its rules. A passed vote is not yet a rule: HSP-027 passed on 3 July 2026 and reached the chain <!--n:pr.v87.days_from_hsp027_passed-->124<!--/n--> days later, and HSP-029, the issuance cut, waits for the same step. The protocol treasury and NEAR AI's links work the same way ([control map](docs/DETAILS.md#the-control-map-other-links)).

**Against.** The policy has been kept: April to October the wallets held <!--n:f2.wallets_kept_share_of_revenue.2026-04_2026-10-->96.9%<!--/n--> of DefiLlama's revenue for NEAR Intents. A policy kept for years is worth almost a rule, so the missing rule blocks Buy, not Hold.

**Check.** `python scripts/control_map.py --verify`; `python scripts/control_history.py --verify`. Each ends with `OK`.

### D. It did not stop when the flow went private

From 27 October swaps began to settle privately on FAR, unseen by public indexers; from 2 to 8 November it carried 85% of the dollars counted on chain ([Tideline at commit 90a78e6](https://github.com/ervhint/tideline-thesis/blob/90a78e6a358315331dbb0c68548edcb736244a3e/THESIS.md), another public repository). DefiLlama's revenue counts the same three wallets we read daily:

| A day, in NEAR | 2–26 October | 27 October – 8 November | Change |
|---|---|---|---|
| Kept by all three wallets (claim B counts two) | <!--n:ps.before.kept_near_per_day-->14,261<!--/n--> | <!--n:ps.after.kept_near_per_day-->13,800<!--/n--> | <!--n:ps.change.kept_near_per_day-->-3%<!--/n--> |
| DefiLlama's revenue | <!--n:ps.before.revenue_near_per_day-->14,323<!--/n--> | <!--n:ps.after.revenue_near_per_day-->13,514<!--/n--> | <!--n:ps.change.revenue_near_per_day-->-6%<!--/n--> |

In dollars the wallets kept <!--n:ps.change.kept_usd_per_day-->-19%<!--/n--> a day, mostly the price. The fall in NEAR is inside the noise: resampling the days gives <!--n:ci.shift.kept_near.low-->-28%<!--/n--> to <!--n:ci.shift.kept_near.high-->+28%<!--/n--> for the wallets (95%); no block length tried goes below <!--n:ci.shift.kept_near.lowest_any_block-->-30%<!--/n-->, and shuffling the 38 days gives p = <!--n:ci.shift.kept_near.permutation_p-->0.83<!--/n-->. A stop would be near -100%. The private path's account on mainnet held <!--n:ct.priced_usd.2026-11-08-->$123.5M<!--/n--> on 8 November, <!--n:ct.stablecoins_usd.2026-11-08-->$37.4M<!--/n--> in stablecoins.

**Against.** Thirteen days show capture did not stop, not that it held level; of five start dates tried, Tideline's gives the smallest fall, the worst <!--n:ps.cut.kept_near_lowest-->-13%<!--/n-->. [PREREG-3](PREREG-3.md) tests the next 30 days.

**Check.** `python scripts/private_shift.py --verify`; `python scripts/intervals.py --verify`. Each ends with `OK`.

### E. The chain is used less, and its use pays holders almost nothing

The track's own prompt example asks for daily transactions and active accounts, and whether NEAR Intents' volume leads network usage better than the price does ([nearai/near-hackathon-prompt-examples](https://github.com/nearai/near-hackathon-prompt-examples/blob/main/tracks/02-thesis.md)). From NearBlocks' daily statistics, November 2025 against October 2026:

| A day, 30-day means | November 2025 | October 2026 | Change |
|---|---|---|---|
| Transactions | <!--n:u.txns.first-->4.27M<!--/n--> | <!--n:u.txns.last-->716k<!--/n--> | <!--n:u.txns.change-->-83%<!--/n--> |
| Active accounts | <!--n:u.active_accounts.first-->2.77M<!--/n--> | <!--n:u.active_accounts.last-->132k<!--/n--> | <!--n:u.active_accounts.change-->-95%<!--/n--> |
| Transactions, not relayed (meta) | <!--n:u.txns_not_meta.first-->1.14M<!--/n--> | <!--n:u.txns_not_meta.last-->476k<!--/n--> | <!--n:u.txns_not_meta.change-->-58%<!--/n--> |

Most of the fall was relayed transactions, and most came in one step: <!--n:u.txns.month.2026-03-->3.48M<!--/n--> a day in March, <!--n:u.txns.month.2026-04-->947k<!--/n--> in April. All gas paid on NEAR in those twelve months was <!--n:u.fee_usd_year-->$1.2M<!--/n-->, <!--n:u.fee_share_of_issuance-->0.8%<!--/n--> of a year of new NEAR. The organizers' question: over <!--n:u.lead.weeks-->52<!--/n--> weeks, neither Intents' volume nor the price leads transactions or active accounts by 0–4 weeks that we can detect (largest correlation <!--n:u.lead.max_abs_corr-->0.22<!--/n-->, smallest p <!--n:u.lead.min_p-->0.44<!--/n-->).

NearBlocks also indexes NEAR Intents on mainnet, a check on DefiLlama: it sees <!--n:u.visible.before_private-->66%<!--/n--> of DefiLlama's volume in the 30 days before 27 October, <!--n:u.visible.after_private-->28%<!--/n--> from 27 October to 8 November and <!--n:u.visible.week_85-->12%<!--/n--> from 2 to 8 November: an independent sign of the move to the private path (claim D).

**From blocks, through BigQuery.** The track's prompt example names NEAR's public BigQuery dataset, so we ran five queries on it ([`sql/`](sql/); results in `data/bq/` with their SHA-256; about 492 GB, inside BigQuery's free monthly terabyte). Its daily transactions equal NearBlocks' on <!--n:bq.agree.txns.days_equal-->365<!--/n--> of <!--n:bq.agree.days-->365<!--/n--> days, and its relayed actions on <!--n:bq.agree.meta.days_equal-->304<!--/n-->, never more than <!--n:bq.agree.meta.max_rel_diff-->0.004%<!--/n--> apart: the fall above is not one indexer's artefact. Both read NEAR's own block data, so this checks the counting, not the chain.

The same dataset shows who calls NEAR Intents' contract, and whether the calls work. From 2 to 31 October 2026 there were <!--n:bq.paths.october.calls-->3,515,141<!--/n--> calls to `intents.near`:

| Path | Share of calls | Failed |
|---|---|---|
| Relay calls: `execute_intents` transactions that `intents.near` signs with its own function-call keys, carrying users' and solvers' signed intents | <!--n:bq.paths.october.relay.share-->40%<!--/n--> | <!--n:bq.paths.october.relay.failed-->28,731<!--/n--> |
| The contract's own callbacks (withdrawals, storage deposits) | <!--n:bq.paths.october.callback.share-->23%<!--/n--> | <!--n:bq.paths.october.callback.failed-->4<!--/n--> |
| Other contracts, mostly token deposits (`ft_on_transfer` from bridged `*.omft.near` tokens, native USDC, `wrap.near` and others) | <!--n:bq.paths.october.other_contract.share-->21%<!--/n--> | <!--n:bq.paths.october.other_contract.failed-->48<!--/n--> |
| `escrow.ai.near`, the NEAR AI Agent Market's escrow, calling `execute_intents` itself | <!--n:bq.paths.october.near_ai_escrow.share-->14%<!--/n--> | <!--n:bq.paths.october.near_ai_escrow.failed-->496,805<!--/n--> of <!--n:bq.paths.october.near_ai_escrow.calls-->496,807<!--/n--> |
| Transactions other accounts sign to it | <!--n:bq.paths.october.other_signed_tx.share-->2%<!--/n--> | <!--n:bq.paths.october.other_signed_tx.failed-->212<!--/n--> |

Leaving out the contract's own callbacks and the failures, <!--n:bq.paths.october.ok_not_callback_a_day-->72,380<!--/n--> calls a day succeeded in October and <!--n:bq.paths.november.ok_not_callback_a_day-->133,169<!--/n--> on 1–11 November. Successful relay calls, the path swaps settle on (one call can carry several intents), were <!--n:bq.paths.october.relay.ok_a_day-->45,918<!--/n--> and <!--n:bq.paths.november.relay.ok_a_day-->103,267<!--/n--> a day (+<!--n:bq.paths.relay_ok_a_day_change-->125%<!--/n-->): the swap path is in use on chain, but these are calls, not dollars, and the rating does not use them. The escrow's own calls fail almost every time, cause not examined (<!--n:bq.paths.november.near_ai_escrow.failed-->179,667<!--/n--> of <!--n:bq.paths.november.near_ai_escrow.calls-->179,667<!--/n--> on 1–11 November); whether the Agent Market's payments settle inside relayed intents is not separated here. Users cannot be counted from these senders: every relayed intent has `intents.near` as its sender, and every deposit the token contract; that needs the contract's event logs, which we did not parse. The organizers' question again, with this on-chain activity in place of DefiLlama's volume: still no lead we can detect (largest absolute correlation <!--n:bq.lead.max_abs_corr-->0.32<!--/n-->, smallest p <!--n:bq.lead.min_p-->0.17<!--/n-->).

**Against.** NearBlocks' figures are its own aggregates, and what counts as an active or relayed account is its definition. BigQuery's dataset is NEAR's own (NEAR Public Lakehouse), not a third party's. A call is not a dollar: the table counts calls, not value. A relayed intent is found by its signer: every access key of `intents.near` is a function-call key for `execute_intents` only (`data/raw/control/keys_intents.near.json`). `escrow.ai.near` is named from the Agent Market's job pages, which list payments to it as "secure deposit"; why its own calls fail is not in the data. Relay calls rose in November while NearBlocks saw less of DefiLlama's volume (claim D): counts are not dollars, and we do not split one from the other. The lead test counts failed calls too. Fifty-two weeks give little power: "not detected" is not "no effect".

**Check.** `python scripts/usage.py --verify` and `python scripts/bq_usage.py --verify`, each → `OK`. To rerun the queries: any Google account, the five files in `sql/`.

## What a holder earns to 2031

How your share of all NEAR changes; only staking moves it. HSP-029 assumed in force from <!--n:lr.activation_utc-->2027-02-14 16:48 UTC<!--/n-->.

| | A year, today | To 1 Jan 2031, today's rule | To 1 Jan 2031, with HSP-029 |
|---|---|---|---|
| Staked, before commission | +<!--n:iv.staked_share_growth-->2.44%<!--/n--> | +<!--n:iv.to2031.staked_share_change-->10.6%<!--/n--> | +<!--n:iv.to2031.staked_share_change.hsp029-->7.9%<!--/n--> |
| Staked, after a 10% commission | +<!--n:iv.staked_share_growth.commission_10-->1.95%<!--/n--> | +<!--n:iv.to2031.staked_share_change.commission_10-->8.4%<!--/n--> | **+<!--n:iv.to2031.staked_share_change.hsp029_commission_10-->6.3%<!--/n-->** |
| Not staked | <!--n:iv.unstaked_share_change-->-2.40%<!--/n--> | <!--n:iv.to2031.unstaked_share_change-->-9.6%<!--/n--> | <!--n:iv.to2031.unstaked_share_change.hsp029-->-7.4%<!--/n--> |

In dollars: that share times the value of all NEAR. No price target; three stated paths: [paths to 2031](docs/DETAILS.md#v4-what-a-holder-earns-to-2031).

**What the price implies.** NEAR is valued at <!--n:peer.near.mcap_to_revenue-->248×<!--/n--> the revenue DefiLlama counts for NEAR Intents, which flows to the three revenue wallets (<!--n:peer.near_all.mcap_to_revenue-->234×<!--/n--> adding L1 gas); Hyperliquid, Jupiter and Aave at <!--n:peer.jupiter.mcap_to_revenue-->24×<!--/n-->–<!--n:peer.aave.mcap_to_revenue-->39×<!--/n--> theirs. At their multiples, that revenue carries <!--n:iv.fee_supported_share.low-->10%<!--/n-->–<!--n:iv.fee_supported_share.high-->16%<!--/n--> of NEAR's value; carrying all of it needs <!--n:iv.breakeven_multiple.low-->×6.3<!--/n-->–<!--n:iv.breakeven_multiple.high-->×10.4<!--/n--> the revenue (<!--n:iv.breakeven_multiple.fdv.low-->×2.6<!--/n-->–<!--n:iv.breakeven_multiple.fdv.high-->×6.1<!--/n--> on fully diluted values).

**The rest of the value.** Against other layer-one chains the picture flips: counting Intents, NEAR is valued at <!--n:peer.near_all.mcap_to_revenue-->234×<!--/n--> a year of revenue, Solana at <!--n:peer.solana.mcap_to_revenue-->1,995×<!--/n-->, Ethereum at <!--n:peer.ethereum.mcap_to_revenue-->3,394×<!--/n-->; on its own gas NEAR is <!--n:peer.near_l1.mcap_to_revenue-->4,091×<!--/n-->. Layer-one tokens are priced on more than fees: staking, money, block space. That is most of NEAR's value, and this thesis does not rate it, but the use behind it is falling: <!--n:u.txns_not_meta.change-->-58%<!--/n--> daily transactions not relayed, in a year ([claim E](#e-the-chain-is-used-less-and-its-use-pays-holders-almost-nothing)). It rates the part a holder could be paid by rule, and finds none of it fixed.

## The rating as rules

The bullets as they were:

- **Why Hold today.** Buy needs a lock that does not exist. Across <!--n:sens.combinations-->54<!--/n--> threshold settings, <!--n:sens.hold-->378<!--/n--> of <!--n:sens.month_ratings-->378<!--/n--> monthly ratings from April to October are Hold; a monthly rating reaches Buy only through 100% coverage, so the growth thresholds among the 54 do not change it. So for now the rules can only move it to Avoid; why not to add is in [Why Hold](README.md#why-hold).
- **Price moves the rating one way only.** At this window's NEAR kept a day, test 2 fails at a price of <!--n:iv.avoid_test2_price_usd-->$0.655<!--/n--> or less, so a crash alone can give Avoid; no price gives Buy.

Every note and the backtest: [the rating in full](docs/DETAILS.md#the-rating-in-full).

### The same rules on four other protocols

**The same rules on four other protocols.** `scripts/peers_rating.py` runs the same gates on them, from their docs and DefiLlama, not their chains (no drain test, no growth leg). For the peers, bought = DefiLlama's holders' revenue ÷ price; NEAR's row uses the buybacks wallets read on chain (10.8%), and on DefiLlama's basis would be <!--n:n2.defillama_revenue_coverage-->16.5%<!--/n-->. A basis is a way of counting new tokens:

| | Lock | Coverage, new tokens the protocol issues | Rating |
|---|---|---|---|
| Hyperliquid | <!--n:sr.hyperliquid.lock-->yes<!--/n--> | <!--n:sr.hyperliquid.basis1.coverage-->103%<!--/n--> at the docs' stake, <!--n:sr.hyperliquid.basis2.coverage-->98%<!--/n--> at today's; <!--n:sr.hyperliquid.basis4.coverage-->8%<!--/n--> with team unlocks | Buy at the docs' stake only, Hold on the other three |
| Pump.fun | <!--n:sr.pump.lock-->no<!--/n-->: a contract we have not reviewed | none issued; <!--n:sr.pump.basis2.coverage-->82%<!--/n--> of insider unlocks | <!--n:sr.pump.rating-->Hold<!--/n--> |
| Aave | <!--n:sr.aave.lock-->no<!--/n-->: a reserve governance can spend | — | <!--n:sr.aave.rating-->Hold<!--/n--> |
| Jupiter | <!--n:sr.jupiter.lock-->no<!--/n-->: burns go to a vote | none issued since February | <!--n:sr.jupiter.rating-->Hold<!--/n--> |

The rules can give Buy; NEAR fails the lock first ([inputs and quotes](docs/DETAILS.md#the-same-rules-on-four-other-protocols)).

## What would change our mind

- **Any day:** a drain gives Avoid (the stock rose every day from 1 to 12 November); **from 11 December at the earliest** (the windows ending 1–10 November passed and are final), so can the two window tests.
- **9 December 2026:** if from 9 November the buybacks keep less than half of <!--n:iv.buybacks_kept_usd_per_day-->$46,692<!--/n--> a day, PREREG-3's P6 (0.90) is scored wrong, and claim D with it.
- **Any day:** a sealed lock, a quiet treasury and 100% counted coverage (best month so far <!--n:bt.highest_coverage_counted-->15.3%<!--/n-->) gives Buy.
- **Any three months running:** coverage of 25% or more ends "small" ([the full list](THESIS.md#what-would-change-our-mind)).
- **1 November 2027, then yearly:** a sealed lock, a quiet treasury and growth of <!--n:lr.required_growth.buybacks.B.2031-->+56%<!--/n--> a year (with HSP-029) gives Buy.

**0.25** that the buybacks go to such a lock by 2031: our judgment, from three peers and NEAR's proposals so far ([sources](docs/DETAILS.md#how-likely-a-rule-is)).

## Objections we expect

- **"Hold is built into the rules."** Upward, yes, until a lock exists; downward, no: a drain gives Avoid any day. The same rules give Hyperliquid, which has a lock, Buy on one of four bases ([the table](#the-rating-as-rules)). NEAR fails twice: no lock, and <!--n:n1.coverage_buybacks-->10.8%<!--/n--> coverage. Even with a lock assumed, every monthly rating since April is Hold, because only 100% coverage gives Buy month to month.
- **"HSP-029 changes the arithmetic."** It does, once validators run it: about <!--n:iv.hsp029_cents_per_new_dollar-->17 cents<!--/n--> per $1 of new NEAR, and full coverage at <!--n:iv.hsp029_wallets_multiple_buybacks-->×5.9<!--/n--> instead of <!--n:iv.wallets_multiple_buybacks-->×9.2<!--/n-->. The lock is still missing.
- **"Other repositories say 12–19%, not 10.8%."** They add burned gas, or count revenue, not balances. With all three wallets, ours is <!--n:n1.coverage_with_front_end-->16.5%<!--/n--> (<!--n:ci.coverage_with_front_end.low-->14.1%<!--/n-->–<!--n:ci.coverage_with_front_end.high-->18.7%<!--/n-->) ([definitions side by side](docs/DETAILS.md#why-public-coverage-figures-differ)).

More, including who decides issuance: [all objections](docs/DETAILS.md#v5-objections-we-expect).

## Predictions, anchored before they are judged

Each file's SHA-256 is anchored in SocialDB on mainnet before it is judged; none amended ([details](docs/DETAILS.md#pre-registration)).

| File | Prediction | Probability | Result |
|---|---|---|---|
| [PREREG](PREREG.md) #1–3 | NEAR AI farm; buybacks inflows 1–15 Nov | — | judged 16 Nov; #2 by the daily Action ([`data/prereg_judged.json`](data/prereg_judged.json)) |
| [PREREG-3](PREREG-3.md) #6–8 | buybacks keep ≥ half / all of the window's dollars, 9 Nov – 9 Dec; DefiLlama complete | 0.90 / 0.75 / 0.90 | judged 9 and 11 Dec |

Two earlier process checks ([PREREG-2](PREREG-2.md) #4–5: HSP-027 in effect, protocol version 88) test the pipeline only: #4 <!--n:pr.4.result-->holds<!--/n--> (Brier <!--n:pr.4.brier-->0.0225<!--/n-->), #5 <!--n:pr.5.result-->holds<!--/n--> (Brier <!--n:pr.5.brier-->0.0625<!--/n-->). PREREG-3 was anchored at block <!--n:pr.anchor3.tx_block-->223,613,293<!--/n-->, before its window opened.

## Check it with one command

Python 3.10 or later (`python3` on some systems, `py` on Windows), standard library only, no network, no keys, under a minute on a laptop. From the repository root:

```
python scripts/rebuild.py
```

It runs the tests, among them `tests/test_independent.py`, a second implementation that imports nothing from `scripts/` and recomputes <!--n:n1.coverage_buybacks-->10.8%<!--/n-->, <!--n:ps.change.kept_near_per_day-->-3%<!--/n--> and <!--n:c.buybacks_confirmations-->2 of 4<!--/n--> from the raw responses; it recomputes every number from the saved raw responses, checks each marked number here against [`data/numbers.json`](data/numbers.json), each quote against its saved page and the chart. Each step: [AGENTS.md](AGENTS.md). Against the chain itself: [Verify live](docs/DETAILS.md#verify-live).

## What we can't measure

The fees (NEAR's datasets on Dune; only revenue is checked on chain), volume and fees on FAR, why NEAR moved (purposes are not on chain), the NEAR AI credits' dollar value (off-chain), and peers (from their docs, DefiLlama and CoinGecko).

## Terms

- **Buybacks wallets:** `buybacks.multisignature.near` and `1csfundsadmin.sputnik-dao.near`, where NEAR bought with NEAR Intents' fees is kept.
- **Front-end fund:** `fefundsadmin.sputnik-dao.near`, a separate DAO that holds the fees of NEAR Intents' own front end. With the two above, the **three revenue wallets**.
- **Coverage:** NEAR the buybacks wallets keep (inflows less outflows) ÷ new NEAR issued (minted less gas burnt), over the same 30 days.
- **FAR:** a private instance of the NEAR protocol, run by a few validators ([NEAR Intents' docs](https://docs.near-intents.org/integration/market-makers/confidential-intents.md)); NEAR Intents' contract there is `intents.far`.
- More, including HSP-027 and HSP-029: [Terms](docs/DETAILS.md#terms).