# Methods (pre-registered) This file fixes the questions, data rules, estimators and pass/fail thresholds **before** the final analysis was run. It was committed to git ahead of the analysis code and the results; `git log -- METHODS.md` shows the order. Any later change to this file is listed in section 8, with its reason. **Disclosure.** On 2026-11-12, before this file was written, the author did one exploratory comparison: 30-day mean volume before and after 2026-02-26 for NEAR Intents (−25%), THORChain (−24%), Chainflip (−43%) and Maya (−47%). No other outcome was looked at before pre-registration. ## 1. Question Revenue is only worth a high multiple if it is defensible. NEAR Intents now earns real fees, so this project asks: **does NEAR Intents have a moat?** It is measured three ways, each against a control built from competitors, so that market-wide moves are not mistaken for NEAR-specific ones. | Test | Event | Question | Moat signal | |---|---|---|---| | T1 Pricing power | Fee switch, 2026-02-23 (Mon) | When NEAR started taking a cut, did volume leave? | No significant loss vs control | | T2 Trust shock | Exploit and pause, 2026-10-01 (Thu) | After a $3.8M exploit and pause, did volume come back? | Gap closed by weeks 4–5 | | T3 Price premium | Last 26 complete weeks | Does Intents charge more than rivals and still hold share? | Premium in ≥75% of weeks, share not falling | Event sources: the fee switch announcement by NEAR Intents () and the exploit post-mortem coverage (). ## 2. Data All data is public, keyless and listed in `moat/sources.py`. `refresh` downloads it and records a sha256 per file in `data/manifest.json`. `build` refuses to run on files that do not match their hashes. - **Treated:** NEAR Intents daily volume (DefiLlama `near-intents`). - **Clean donor pool (controls):** THORChain, Chainflip, Maya Protocol (native cross-chain swap venues); SOCKET and 0x bridge aggregators; all-DEX volume (the market). None is known to route its own flow through NEAR Intents. - **Excluded from the donor pool, sensitivity only:** LI.FI, Rango and Bitget Wallet. Each routes part of its flow through NEAR Intents, so a shock to Intents also hits them and would bias the estimated effect toward zero. - **Fees:** DefiLlama daily fees for Intents, THORChain, Chainflip and Maya. - **Revenue:** DefiLlama daily revenue for Intents (the protocol's own take). ## 3. Cleaning rules 1. Snapshot date `S` is the UTC date of the fetch. Only days before `S` are used. 2. A daily value of zero is treated as missing (a reporting gap or a halt). It is never imputed. 3. **Event-weeks.** For event day `E`, week `k` covers days `E+7k … E+7k+6`. A week's value is `7 × mean(observed days)` if at least 4 of its 7 days are observed; otherwise the week is missing. Only complete calendar spans (all 7 days before `S`) are used. 4. All volume comparisons use natural logs of weekly volume. ## 4. Estimator: synthetic control with intercept For an event `E`, a pre-window of `P = 26` weeks (`k = −26 … −1`) and a post window of `H` weeks: - A donor is eligible if it is observed in ≥ 80% of pre-weeks and in every post week. - Weights `w ≥ 0, Σw = 1` minimise the pre-period squared error between the demeaned log volume of Intents and the demeaned weighted donors. The fit uses pre-weeks where every eligible donor is observed. The intercept is `mean(y_pre) − mean(X_pre)·w`, so the control matches trends, not levels. - `gap_k = log y_k − (intercept + X_k·w)`. The **effect** is the mean gap over the chosen post weeks, reported as `exp(effect) − 1`. ### Inference - **Placebo-in-time (primary).** The same estimator is run at fake event dates `E'`, every 7 days, whose whole window (pre and post) ends before `E` and starts on or after the first week Intents is observed. The one-sided p-value for a loss is the share of placebo effects ≤ the real effect. Placebo windows may contain other real events; this widens the placebo spread and so makes the test conservative. - **Placebo-in-space (secondary).** Each eligible donor is treated as if it had the event, using the other donors (not Intents) as controls. Intents' rank by post/pre RMSPE ratio is reported. With few donors this is weak, and the report says so. ### Robustness (reported, not used for verdicts) - Leave-one-donor-out: the range of effects. - Simple difference-in-differences: change in `log y − mean(log donors)`. - Pre-window of 13 weeks instead of 26. - Contaminated donors added to the pool. ## 5. Tests and decision rules **T1 Pricing power.** `E = 2026-02-23`, `H = 8`, effect = mean gap over weeks 0–7. **PASS** unless the effect is below −10% **and** the placebo p-value is below 0.10, which is **FAIL**. A *first stage* is also reported: protocol revenue as a share of fees before and after `E`, to show the switch changed what NEAR earns. User-facing fee rates around `E` are reported too: if users' price did not change, T1 measures whether integrators and their flow left, not user price sensitivity. **T2 Trust shock.** `E = 2026-10-01`, `H = 6` (weeks 0–5, the most the snapshot allows). The shock is the mean gap over weeks 0–1. The recovery statistic is the mean gap over weeks 4–5, with its own placebo distribution. **PASS** unless the recovery gap is below −10% **and** its placebo p-value is below 0.10, which is **FAIL**. **T3 Price premium.** Over the last 26 complete event-weeks before `S`: - Fee rate (bps) = weekly fees / weekly volume, per venue. - Competitor rate = volume-weighted rate across THORChain, Chainflip and Maya for the venues observed that week. - Premium weeks = share of weeks where Intents' rate is above the competitor rate. - Share = Intents volume / (Intents + competitors' volume), log-share trend estimated by OLS; 90% CI by moving-block bootstrap (block 4, 2,000 draws, seed 7). **PASS** if premium weeks ≥ 75% **and** the CI upper bound of the share slope is ≥ 0. **FAIL** otherwise. The elasticity of share to relative fee rate is reported as descriptive only, because fee rates and route mix move together. ## 6. Overall verdict Three passes: **strong moat evidence**. Two: **moderate**. One or none: **weak**. The verdict is printed as computed, whichever way it falls. ## 7. Known limits, stated in advance - DefiLlama is an aggregator. Its adapters can change history; the snapshot and hashes pin the version analysed. - The donor pool is small, so placebo-in-space p-values have a floor around 1/(donors+1). - A synthetic control assumes no other Intents-specific shock in the window. Known overlapping events are listed in the report where found. - Solver-level concentration is out of scope: in a 200-transaction sample of recent `intents.near` activity no `token_diff` swap events appeared, so solver identity can't be measured reliably from public data without the Explorer API. ## 8. Changes after pre-registration Each item says whether it was made before or after the first results were seen. None of them changes a verdict rule or threshold. 1. **Revenue zeros kept (before results).** Rule 2 treats zero as missing. For the revenue series a zero means "no revenue taken", which is the quantity being measured, so zeros are kept there (`parse_chart(keep_zero=True)`). Volume and fee series follow rule 2 unchanged. 2. **T3 share denominator (before results).** THORChain (May 16–Jun 21, 2026) and Maya (May–Sep 2026) report no volume for weeks inside the T3 window. A share whose denominator gains and loses venues would move for reasons unrelated to Intents, so the share uses only competitors observed in every week. That leaves Chainflip. A sensitivity slope against THORChain and Chainflip, over the weeks both are observed, is also reported. 3. **"Conservative" was wrong (after results).** Section 4 says placebo windows containing other events only widen the placebo spread. In this data that is false: most placebo windows fall in Intents' 2025 hypergrowth, which shifts the whole placebo distribution upward rather than widening it, so a normal result in a mature period can look like a loss. The pre-registered p-values are still reported and still decide the verdicts. As a labelled diagnostic, the report adds placebo effects plotted by date and the p-value using 2026 placebos only. 4. **Robustness in percent (after results).** Robustness numbers in `results.json` are expressed as percentage effects (`*_pct`) instead of log points, so they read the same as the headline effects. 5. **On-chain sample and NEAR economics added (after results).** These are corroborating, descriptive analyses. They do not enter any T1–T3 verdict. - *Block sample* (`moat/onchain.py`): a deterministic sample of NEAR blocks read keylessly from neardata.xyz. Week boundaries are block heights found by timestamp search, within 15 minutes. Inside each week, heights are drawn uniformly with seed 20261112: 80 a week from 2026-04-02 to 2026-11-11 and 40 a week from 2025-12-29 to 2026-04-19. Blocks are read every 30 minutes from 2026-09-30 to 2026-10-02, plus two header blocks a year apart for supply. Each block is reduced to a summary with its hash (`data/onchain/blocks.jsonl`), and `verify-onchain` re-fetches a random subset to check it. - *Uses:* (a) the weekly correlation of on-chain executed intents with DefiLlama volume, in levels and in changes; (b) the exploit before/after on both measures; (c) timing of the Oct 1 pause, with the longest run of empty sampled blocks benchmarked against the share of empty blocks before Oct 1 (exact probability of such a run anywhere in the sample); (d) repeat-counterparty concentration in `token_diff` events over the last 8 weeks; (e) exact annual NEAR issuance from `total_supply`, set against annualised Intents revenue at the 30-day average price (CoinGecko, added to the manifest with its own `fetched_at`, leaving the snapshot date unchanged). - Section 7's statement that solver concentration was out of scope was based on a 200-transaction sample from the FastNear transactions API. Full blocks do contain `token_diff` events, so (d) became possible. 6. **Live dashboard (after results).** A daily CI job rebuilds the page on freshly fetched data under the same rules, labelled as a live re-test. The pinned, pre-registered build stays in the repository and at `/pinned/` on the site. 7. **Code package renamed (after results).** The project was renamed Zeat, and the code package `moat/` was renamed `zeat/` (so `python3 -m zeat build`). Paths above that say `moat/` now live under `zeat/`. No logic, data or threshold changed; `results/results.json` is byte-identical.