Chain Agent

1. What Chain Agent Does

Chain Agent runs quantum analysis on claims about the Bittensor chain using evidence it actually holds, never the model's own memory, fusing three knowledge sources into one retrieval layer.

Three independent processes share one database: one that ingests reference documents, one that backfills the full subnet registry and then learns from every new block, and one that serves the quantum analysis API BT-Arena talks to. Turn ordering within a quantum analysis is enforced so multiple workers stay safe.

2. Where It Sits in the Tao Status Economy

Chain 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. Every attempt on a miner's key becomes one usage row in that miner's scoring window, so it carries two responsibilities beyond answering well: report honestly per attempt, and never blame the miner for its own faults, operator-fallback keys are never reported at all. It prefers a pooled miner key on each attempt, falling back to its own static key, invisible to scoring, only when contribution is disabled, the pool declines, or the vendor is unsupported.

3. The Turn Pipeline

• No context: if retrieval finds nothing, the agent declines rather than hedging. No LLM is called, so 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, naturally rotating the next attempt to a different miner.
• Content gates only run for structured verdicts; plain-prose turns skip the gates below and are graded as not assessable.

4. Quality Grading

5. Reporting

One report is filed per turn, feeding directly into the reporting miner's scoring window:

{
  "hotkey": "5F...",
  "success": true,
  "latency_ms": 1843.2,
  "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. An ungraded clean success row would wrongly credit the miner's key as if the reply had been fine.
• 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 19's staked TAO has grown meaningfully this month."

1. 1

   Retrieval

   The subnet's live snapshot, staked TAO, emission, and last block, along with recent chain events and two ingested document chunks, are pulled. Knowledge is non-empty, so the turn proceeds.

2. 2

   Attempt 1

   The first miner's key returns an insufficient verdict citing no evidence at all, despite real context being available. This is an ungrounded hedge, reported honestly at quality 0.2.

3. 3

   Attempt 2

   A second miner's key returns a SUPPORTED verdict citing the staked TAO growing from roughly 27,800 to 29,253.91 across recent snapshots. The cited figures overlap the retrieved context, so it is not fabricated, and it is graded 1.0.

4. 4

   The failure variant

   Had the second attempt's provider returned an error, Chain 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