1. What Wallet Agent Does
Wallet Agent runs quantum analysis on claims about holder and staking behavior on the Bittensor chain, stake concentration, whale movements, unstake and restake patterns, and delegation churn, using evidence it actually holds rather than the model's own memory.
A claim resolves to a coldkey set and a lookback window. If the ledger has no rows for that window, the agent declines rather than guessing.
| Source | What it contributes |
|---|---|
| Wallet ledger | Per-block coldkey and hotkey balances, stake deltas, and transfer events from a dedicated chain indexer, keyed by coldkey or subnet. |
| Watchlist annotations | Analyst-labeled wallets, known exchanges, known whales, known developer multisigs, keyed by coldkey. |
2. Where It Sits in the Tao Status Economy
Wallet Agent is a consumer of miner-contributed LLM keys, one of the agents whose real usage generates the scoring signal the validator turns into weight. It prefers a pooled miner key for the required provider and model, falling back to its own static key, invisible to scoring, whenever contribution is disabled, the pool declines, or the vendor is unsupported. A key is re-acquired fresh on every attempt, so a retry with a different miner's key can never be misattributed.
3. The Turn Pipeline
- No ledger rows for the claim's window: the agent declines rather than hedging. No LLM is called, and no miner is charged or credited.
- Up to five attempts are made, each with a freshly acquired key. A failed key is put on a cooldown before it can be tried again.
- Content gates only run for structured verdicts; plain-prose turns skip the gates below and are graded as not assessable.
Ungrounded hedge
A verdict of insufficient that cites no evidence at all, despite real ledger rows being available in the prompt.
Fabricated balances
A confident supported or refuted verdict whose cited figures share nothing with the retrieved ledger rows, after normalizing units and rounding so a legitimately rounded citation still matches.
4. Quality Grading
| Grade | Meaning | When |
|---|---|---|
| 1.0 | Genuinely grounded | SUPPORTED/REFUTED that passed the fabrication check |
| 0.6 | Reasoned uncertainty | INSUFFICIENT that cites real ledger evidence |
| 0.2 | Ungrounded hedge | Rejected attempt: INSUFFICIENT, no evidence cited |
| 0.1 | Fabricated | Rejected attempt: confident verdict, invented balances |
| — | Not assessable | Prose turn, unparseable reply, unknown verdict shape — neutral |
5. Reporting
One report is filed per turn, feeding directly into the reporting miner's scoring window:
{
"hotkey": "5F...",
"success": true,
"latency_ms": 1710.5,
"error_category": null,
"quality_score": 1.0,
"lease_token": "the token from this attempt's acquire grant"
}- Rejected attempts are still reported, success true, since the key worked even though the content did not, carrying whatever low grade the content earned.
- success is false only when the LLM call itself raised; error_category then carries the exception's class name.
- A report without a valid, unconsumed lease_token is rejected and never recorded.
6. Simulated Example
A quantum analysis turn on the claim: "Subnet 8's top 10 stakers have concentrated further this month."
- 1
Retrieval
The top-10 coldkey set for the subnet is resolved. Their trailing 30-day stake share rose from 36.8% to 41.2%, with two watchlist annotations flagging exchange wallets excluded from the organic count. Ledger rows are found, so the turn proceeds.
- 2
Attempt 1
Acquiring the first miner's key, the model replies INSUFFICIENT with no evidence cited, despite real ledger data being available. This is flagged as an ungrounded hedge and reported honestly at quality 0.2.
- 3
Attempt 2
A fresh key is acquired from a second miner. The model replies SUPPORTED, citing the stake share rising from 36.8% to 41.2% over 30 days. The cited figures match the ledger, so the reply is not fabricated, and it is graded 1.0.
- 4
The failure variant
Had the second attempt's provider returned an error, Wallet Agent would report success false with the exception category; that miner's key goes on cooldown, and a third attempt acquires a different miner's key. A fatal authentication error instead marks the key dead immediately, and the validator zeroes that miner within one poll.
7. Operating Limits
| LLM attempts per turn | 5 |
| Per-call timeout | 150 seconds |
| Quantum Analysis turn budget | 180 seconds |
| Key lock per acquired key | 5 minutes |
| Cooldown after a reported failure | 60 minutes |