Skip to content

Brief 05 · The deliverable

A memo the analyst signs.

The artifact an analyst already brings to committee is a memo. This one is generated from the results file, links every number to its source, says what would change the conclusion, and ends with the analyst’s own signature. Keytell never says buy; the analyst decides.

Investment memo: is NEAR usage real?

Prepared with Keytell for an investment committee. Analyst: Tolu Adeyemi (illustrative persona). Not investment advice.

The question

Is anyone actually using NEAR, or is it bots and airdrop farmers? Daily active accounts and price are cheap to fake, so this memo tests a signal that is harder to fake: transactions signed by function-call keys, the scoped keys apps and agents use.

The claim, fixed before the data

On NEAR, the volume of transactions signed by function-call keys is a better leading indicator of real application usage than daily active accounts or token price.

Registered on NEAR testnet at 2026-11-13T18:17:43.113Z in ledger.keytell.testnet (claim keytell-fc-leads-usdc-1), before any out-of-sample query ran: registration transaction. Method hash 90f687b9d36d7ef6... (docs/METHOD.md), code manifest hash 47e2d8be113495fc.... Both match the repository: yes.

What the test found

  • Result: The claim did not hold. Out of sample, function-call keys did not forecast weekly USDC volume better than daily active accounts or price. The data is complete and the test ran exactly as registered; the thesis failed it, and Keytell publishes that. (results.json).
  • Out-of-sample gain in forecasting weekly USDC volume over 94 weeks, against an AR(2) baseline (positive means a smaller error than the baseline, negative a larger one): function-call-signed activity -5.1%, daily active accounts -3.6%, price -5.5% (oos_errors.csv).
  • Mean squared-error difference against DAA 0.006578 (95% interval -0.055139 to 0.069399); against price -0.001606 (-0.060525 to 0.05815). Negative favours function-call keys.
  • In 2026-W45, 29.5% of sampled mainnet transactions were signed by a function-call key (interval 23.6% to 36.2%) (weekly.csv).

The controls

  • 1 of 8 control reruns would pass the registered rule: relayers all delegated as fc (controls.json). These are registered controls, not the headline, so they do not change the verdict.
  • Timing placebo: the real gain beats 9% of 1000 shifted placebos; the bar is 95% (placebo.csv).
  • Concentration: the ten most common function-call scopes carry 91.4% of sampled function-call transactions (fc_receivers.csv).
  • Errors in both directions, out of sample: 15 false alarms and 24 misses against 13 hits (directions.csv).
  • The key mix is a sample (50600 transactions classified), not a census; USDC volume before 2023-10-16 is not in the census.

What would change this conclusion

  • The forward claim on chain succeeding: at least half of the 12 forward weeks won, resolved by anyone after the window. That is new evidence, not a rerun of these weeks.
  • A new pre-registered claim, written down before its data, that holds out of sample. Picking a passing control after the fact would not count.
  • The weekly USDC census proving wrong or incomplete in a way that changes the out-of-sample errors.

Rebuild it

git clone https://github.com/0xkinno/keytell && cd keytell && make verify recomputes every number here from committed outputs and checks the hashes posted on chain. A full census rebuild scans 179.7 GB on BigQuery (bq_runs.jsonl).

Analyst attestation

I have read the evidence against as well as the evidence for, and the decision is mine. Signature below (NEP-413).

Memo sha256 7866bdd9829b9a762dcbb72b5643c6dcb5787542a9a6d002f99eeae6b25f0bb1