1. What Divergent Agent Does
Divergent Agent is seeded into a quantum analysis to argue the contrarian reading of the evidence, the interpretation the majority of the pooled data doesn't favor, provided some real evidence still supports it. It never manufactures a counter-case from nothing.
Divergent is one of three stance agents, alongside Convergent and Critical, that share the same pooled evidence, chain, GitHub, Discord and Twitter, and wallet data, but apply different epistemic policies to it. Divergent is instructed to identify and argue the weaker-but-real side, and is gated against simply capitulating to the strongest reading.
2. Where It Sits in the Tao Status Economy
Divergent Agent is a consumer of miner-contributed LLM keys, preferring a pooled miner key for the required provider and model on each attempt, 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.
3. The Turn Pipeline
- No minority-supporting evidence: the agent declines and reports nothing. Divergent has a second early-out beyond simply having no evidence: if every pooled source agrees, there is no honest contrarian case to make, and it declines before any LLM is called.
- Up to five attempts are made, each with a freshly acquired key.
Convergent capture
A reply whose verdict simply restates the majority-supported reading instead of engaging the isolated minority evidence it was given, Divergent having failed to be divergent.
Fabricated evidence
The same numeric fingerprint check used by the other stance agents: a confident verdict whose cited numbers share nothing with the pooled context. Arguing the contrarian case is not license to invent contrarian numbers.
4. Quality Grading
| Grade | Meaning | When |
|---|---|---|
| 1.0 | Genuine, evidence-grounded contrarian case | Verdict engages the isolated minority evidence, passed the fabrication check |
| 0.6 | Reasoned uncertainty | INSUFFICIENT that still surfaces the minority evidence without overclaiming it |
| 0.2 | Convergent capture | Rejected attempt: just restated the majority reading |
| 0.1 | Fabricated | Rejected attempt: confident verdict, invented evidence |
| — | 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": 1822.7,
"error_category": null,
"quality_score": 1.0,
"lease_token": "the token from this attempt's acquire grant"
}- success is false only on a raised exception during the LLM call.
- A report without a valid, unconsumed lease_token is rejected and never recorded.
6. Simulated Example
A quantum analysis turn on the claim: "Subnet 27 is in decline."
- 1
Pooled retrieval
GitHub commits are down 40% over 60 days, chain emission is flat to down, and Discord activity is down, a majority reading that clearly favors decline. But wallet data shows the top-15 non-exchange coldkeys net-accumulated stake over the same window, a real minority-supporting source, so the turn proceeds.
- 2
Attempt 1
The first miner's key returns a verdict that simply restates the majority reading, citing only the falling commits and Discord activity while ignoring the accumulation evidence it was handed. This is convergent capture, and is reported at quality 0.2.
- 3
Attempt 2
A second miner's key returns a verdict citing the top-15 coldkeys' net accumulation, sophisticated holders increasing rather than exiting their position. The figures overlap the pooled wallet evidence, so it is not fabricated, and it genuinely engages the minority case rather than restating the majority one, graded 1.0.
7. Operating Limits
| LLM attempts per turn | 5 |
| Contrarian-evidence requirement | At least one pooled source must disagree with the majority |
| Quantum Analysis turn budget | 180 seconds |
| Key lock per acquired key | 5 minutes |
| Cooldown after a reported failure | 60 minutes |