From Price to Flow: How Bittensor Re-Plumbed Its Emission Engine

Bittensor mints 3,600 TAO per day — roughly $770k at today’s ~$214 — and splits it across 128 “subnets,” each one a competitive market for some AI task: serverless inference (Chutes), GPU rental (Targon), even decentralized pretraining (τemplar, the largest subnet token by market cap and a cousin of the systems we covered in our decentralized-training teardown). The question of which subnet deserves what share of that block reward is the protocol’s central resource-allocation problem, and in the last sixteen months Bittensor has answered it three different ways: validator voting, then AMM prices, and — since November 2025 — exponentially smoothed net staking flows.

That second pivot is the interesting one. It’s a rare, well-documented case of a live protocol discovering that the market signal it wired into its consensus was cheaply manipulable, and replacing it under fire with a signal that costs real, locked capital to fake. If you build incentive systems — token emissions, validator rewards, or for that matter any ML reward function — the failure mode and the fix both generalize.

Regime one: validators vote (and why that died)

Until February 2025, root-network validators decided emission splits by weighted vote. The Dynamic TAO (dTAO) whitepaper is blunt about why that couldn’t scale: validators can’t meaningfully evaluate dozens of heterogeneous subnets (apathy), they over-weight subnets they themselves validate (self-dealing), and a subnet owner who controls or bribes a few large validators controls his own budget (capture). The fix, in classic crypto fashion, was to fire the committee and hire a market.

Regime two: the AMM price oracle

dTAO gave every subnet its own token — alpha — and a constant-product AMM pool pairing that alpha against TAO. Staking into subnet i swaps TAO into the pool and alpha out; unstaking swaps back. The pool quotes a price

p_i = τ_i / α_i        # TAO reserves over alpha reserves

and the protocol read that price as a valuation. Each block, freshly minted TAO was injected into pools proportionally to price:

Δτ_i = ( p_i / Σ_j p_j ) · Δτ̄        # subnet i's slice of the block reward

with a matching alpha injection capped by each subnet’s own halving schedule, so injections were price-neutral. Emission share was relative alpha price. The whitepaper anticipated shenanigans and smoothed the signal with a moving average of prices, arguing that broad market participation makes valuations “resistant to manipulation.”

It wasn’t resistant enough, because of a property every DeFi engineer knows in their bones: in a constant-product pool, price moves quadratically with a one-sided buy, and the move is refundable. Push X TAO into a pool holding τ TAO of reserves and the quoted price multiplies by (1 + X/τ)² — buy into a thin pool with 10% of its reserves and you’ve moved its price 21%. Hold, collect the inflated emission share, then sell back and recover your principal minus slippage and fees. The manipulation isn’t just cheap; it’s a loan.

The ecosystem ran the experiment within weeks. The best-documented case is subnet 281, which rebranded as the “TAO Accumulation Corporation” (the community called it the LOL subnet): miners ran no model and did no work — validators scored miners by how much of the subnet’s own alpha they held. Buying alpha raised the price; the price raised the subnet’s TAO emissions; the emissions paid the holders, which attracted more buying. A textbook reflexive loop, fully compliant with the emission formula, contributing nothing to the “decentralized intelligence” the formula was supposed to fund. Speculative subnets multiplied behind it, and the foundation found itself doing exactly what dTAO was built to avoid: manually intervening in emission dynamics.

Regime three: TaoFlow, or emissions as flow accounting

The November 2025 fix changed what the protocol measures. Emission share no longer reads the AMM price at all. It reads net TAO flow: TAO staked into a subnet minus TAO unstaked out of it, tallied per block at the source — the staking extrinsics themselves (record_tao_inflow / record_tao_outflow in subtensor’s stake operations) — and smoothed with a per-block exponential moving average:

S_i ← (1 − a) · S_i + a · net_flow_i      # a ≈ 0.000003209 per 12s block
z_i = max(S_i − L, 0)                     # clip: sustained outflows → zero
share_i = z_i / Σ_j z_j                   # linear normalization (power p = 1)

That smoothing constant is the load-bearing number: it gives the EMA a 30-day half-life (an effective window of ~87 days). A one-shot deposit, however large, enters the score scaled by a ≈ 3×10⁻⁶ and then decays — moving the EMA halfway to any new level takes a month of sustained flow. Three details do real security work:

• Protocol flows don’t count. The accounting deducts TAO the protocol itself moves — per-block pool injections, automated chain buys, root-staker claims — and ignores miner registration burns. Only discretionary user staking registers.
• Swaps don’t count. The old attack surface — trading against the pool — no longer touches the emission signal. You must stake, which means principal parked in the subnet, not a round trip through its order book.
• Exit erases you. Unstaking records an equal-and-opposite outflow. The capital that built your score tears it back down on the way out, and subnets with sustained negative net flow clip to zero emissions.

The rest of the pipeline is unchanged and worth knowing. Each block, run_coinbase distributes 0.5 TAO (halved from 1.0 on December 14, 2025) across subnets per the flow shares, injecting TAO and alpha into pools price-neutrally — with a subsidy branch that down-scales both sides when an alpha price sits below its emission proportion. Per-subnet rewards accumulate in PendingEmission, then drain on a staggered schedule — (block + netuid + 1) % (tempo + 1) == 0, one tempo ≈ 360 blocks ≈ 72 minutes — through Yuma Consensus, which splits alpha 41% to miners by validator-scored performance, 41% to validators and their stakers, and 18% to the subnet owner. TaoFlow only changed the first stage: which subnet gets funded. Who inside the subnet earns it is still Yuma’s problem.

What manipulation costs now

Compare the attacks side by side. Under the price regime, capturing emission share required a refundable swap whose price impact registered quadratically and (modulo the price EMA) near-instantly. Under TaoFlow, the equivalent move — wash-staking — requires you to lock principal for weeks against a 30-day half-life, forfeit the score on exit, and compete against every honest staker’s flow in the denominator. The manipulation budget went from briefly borrow X to park X for a month and hope emissions outrun your opportunity cost. That’s the design lesson in one line: prefer signals that cost sunk, time-integrated capital over signals that cost a refundable balance-sheet entry.

The calculator below runs both regimes with the real protocol constants — push on the attack size, pool depth, and network flow and watch where the asymmetry comes from.

⬢ loading artifact…

The honest caveats

TaoFlow is a better proxy, not ground truth — Goodhart’s law doesn’t repeal that easily.

• Flow measures capital sentiment, not AI quality. A subnet shipping great models with flat staking earns less than a well-marketed one attracting fresh TAO. The signal is “where do stakers move money,” which correlates with value but is not value — the same gap that solver auctions close with explicit objective functions, and Bittensor deliberately doesn’t.
• Whales still steer. A fund that stakes eight figures of TAO moves the EMA honestly and durably. TaoFlow prices out cheap manipulation; expensive, patient manipulation is indistinguishable from conviction — arguably by design.
• The zero-clip cuts both ways. Mature subnets with stable stake and slow churn can drift to near-zero net flow and lose emissions despite doing real work — the community’s “dead weight” debate is genuinely unresolved, and emissions-zeroed subnets now need a comeback story denominated in fresh inflows.
• A 30-day half-life is a 30-day lag. The same smoothing that blunts attacks also slows the network’s response to genuinely new, genuinely good subnets. Responsiveness and manipulation resistance trade off through a single constant, and a is a governance knob, not a law of nature.

The meta-lesson travels well beyond Bittensor. The protocol’s allocation history — committee → spot price → smoothed flows — is a compressed tour of mechanism-design failure modes: the committee got captured, the price got pumped, and the flow metric now gets Goodharted slowly instead of instantly. Each iteration didn’t find an ungameable signal; it raised the price and the latency of gaming. For systems that pay out real money against a proxy metric — token emissions, oracle rewards, even RLHF reward models — that may be the only kind of progress on offer.