Skip to main content

Livepeer Delegation Outflows & Retention — On-chain Research (Arbitrum)

This note is meant to ground any “small-delegator incentive” tokenomics proposals in observed onchain behavior, and specifically to test the hypothesis:

“Delegator outflows were driven by sybil farmers cashing out.”

Scope​

  • Chain: Arbitrum One
  • Contract: Livepeer BondingManager proxy 0x35Bcf3c30594191d53231E4FF333E8A770453e40
  • Data source: JSON-RPC eth_getLogs (no Arbiscan/Etherscan key)
  • Window scanned:
    • Start (deployment): block 5,856,381 (2022-02-11 13:25:10 UTC)
    • End (scan time): ~block 422,274,552 (2026-01-17)
  • Events scanned:
    • Bond
    • Unbond
    • WithdrawStake
    • EarningsClaimed

Outputs live under:

  • artifacts/livepeer-delegator-flows/daily.json (UTC day aggregates)
  • artifacts/livepeer-delegator-flows/delegators_state.pkl (per-delegator summary state)
  • Scanner: tools/livepeer/arb_bondingmanager_scan.py

High-level results (from the full window)​

New delegators (first-time bonders)​

  • Total new delegators (addresses with a recorded first Bond): 4,886
  • New delegators by year (first bond timestamp):
    • 2022: 2,213
    • 2023: 1,492
    • 2024: 685
    • 2025: 482
    • 2026: 14 (partial year to 2026-01-17)

Note: there are additional addresses that appear in later events but do not have a recorded first_bond_ts (likely due to TransferBond paths, which are not included in the scan set).

Delegator size distribution (max bonded amount, among first-time bonders)​

Among the 4,886 addresses with first_bond_ts:

  • Median max bonded stake: ~72 LPT
  • 83.5% had max bonded stake ≤ 1,000 LPT
  • Threshold counts:
    • ≤ 10 LPT: 1,097
    • ≤ 100 LPT: 2,760
    • ≤ 1,000 LPT: 4,079

How fast do new delegators exit? (first bond → first withdraw)​

Among the 4,886 first-time bonders:

  • Withdrew within 7 days: 126 (2.6%)
  • Withdrew within 30 days: 419 (8.6%)
  • Withdrew within 60 days: 585 (12.0%)
  • Withdrew within 90 days: 734 (15.0%)
  • Withdrew within 180 days: 1,030 (21.1%)
  • Withdrew within 365 days: 1,472 (30.1%)
  • Never withdrew (as of 2026-01-17): 2,564 (52.5%)

This is not perfect “retention” (delegators can re-enter), but it’s a useful “first exit” signal.

Merkle migration claimers: how fast do they exit?​

Livepeer’s Arbitrum migration uses L2Migrator (Controller key L2Migrator) with a Merkle snapshot contract (Controller key MerkleSnapshot).

  • L2Migrator: 0x148D5b6B4df9530c7C76A810bd1Cdf69EC4c2085
  • MerkleSnapshot: 0x10736ffaCe687658F88a46D042631d182C7757f7
  • Event analyzed: StakeClaimed(address,address,uint256,uint256) (topic0 0xc08c27…)
  • Scanner: tools/livepeer/arb_l2migrator_stake_claims.py
  • Outputs:
    • artifacts/livepeer-delegator-flows/stake_claimed.ndjson
    • artifacts/livepeer-delegator-flows/stake_claimed_summary.json

Claim volume​

  • Unique claimers: 1,575 addresses
  • Total stake claimed: ~9.76M LPT
  • Claim timing (by count + stake):
    • Concentrated in Feb–Mar 2022 (initial migration), but claim activity continues through 2026 (long tail).

Time from claim → exit actions (claimer behavior)​

Quantiles (days from claim):

  • Claim → first unbond: q25 ~16d, median ~261d, q75 ~689d
  • Claim → first withdraw: q25 ~27d, median ~275d, q75 ~724d

Fast-exit rates (count-weighted, among claimers):

  • Unbond within 7d: ~14.9%
  • Withdraw within 30d: ~15.8%

Fast-exit rates (stake-weighted, by claimed stake of those wallets):

  • Withdraw within 30d: ~11.4% of claimed stake

Interpretation: there is a meaningful “claimed → withdraw within ~1 month” cohort, but it is not uniformly distributed across delegates and is amount-concentrated (a few large wallets can dominate the stake-weighted picture).

Outflow spikes: by amount vs by number of withdrawers​

Largest outflow-by-amount days (WithdrawStake)​

Top days by total withdrawn amount (LPT), with number of withdrawers:

  • 2023-02-09: ~1.76M LPT withdrawn by 6 withdrawers
  • 2025-09-17: ~1.07M LPT withdrawn by 17 withdrawers
  • 2024-12-23: ~1.04M LPT withdrawn by 17 withdrawers
  • 2024-12-31: ~1.02M LPT withdrawn by 20 withdrawers
  • 2024-03-15: ~0.82M LPT withdrawn by 31 withdrawers

Interpretation: the biggest outflow-by-amount days are generally dominated by small numbers of withdrawers (whale-like behavior).

Largest outflow-by-count days (WithdrawStake)​

Top days by number of withdrawers:

  • 2025-06-06: 45 withdrawers (total ~404.7k LPT)
  • 2024-12-13: 33 withdrawers (total ~149.3k LPT)
  • 2025-06-05: 31 withdrawers (total ~582.1k LPT)
  • 2024-03-15: 31 withdrawers (total ~824.6k LPT)
  • 2024-03-14: 31 withdrawers (total ~27.0k LPT)

Interpretation: there are days with many withdrawers but relatively small total amounts, which is where “many small accounts exiting” would show up — but these peaks are on the order of ~30–45 withdrawers/day, not thousands.

“First withdraw” clustering is small​

If we only count first-ever withdrawers (proxy for “cashout once and leave”), the maximum is still small:

  • largest day by first-withdrawers: 2024-03-14 with 16 first-withdrawers

Outflows by “team” (delegate/orchestrator address)​

The Unbond event is indexed by delegate (orchestrator), so we can attribute stake outflows by “team” (delegate address).

Scanner: tools/livepeer/arb_unbond_by_delegate_scan.py
Output: artifacts/livepeer-delegator-flows/unbond_by_delegate.json

Top delegates by total unbonded stake (LPT)​

Top 5 delegates by total unbonded stake over the full window:

  • 0x525419ff5707190389bfb5c87c375d710f5fcb0e: ~7.54M LPT unbonded (699 unique unbonders)
  • 0x4416a274f86e1db860b513548b672154d43b81b2: ~4.64M LPT unbonded (8 unique unbonders)
  • 0x9c10672cee058fd658103d90872fe431bb6c0afa: ~3.86M LPT unbonded (57 unique unbonders)
  • 0x4f4758f7167b18e1f5b3c1a7575e3eb584894dbc: ~2.04M LPT unbonded (13 unique unbonders)
  • 0xda43d85b8d419a9c51bbf0089c9bd5169c23f2f9: ~2.00M LPT unbonded (81 unique unbonders)

Interpretation:

  • Some “top outflow” delegates are whale-driven (very few unbonders, very large unbond totals).
  • Others are broad-based (many unique unbonders, moderate average size).

Delegate outflow profile “categories” (by unique unbonders)​

This is a behavioral categorization based on unique_delegators in artifacts/livepeer-delegator-flows/unbond_by_delegate.json (i.e., how many distinct delegators unbonded from a given delegate over the full window).

Note: the unbond_by_delegate sum is 48.431M LPT unbonded, slightly below the full-unbond scan total (48.433M LPT) due to being a separate scan. The category shares below use the unbond_by_delegate total.

1) Broad-based outflows (many delegators leave)​

Definition: ≥100 unique unbonders (5 delegates total).

These few delegates account for ~11.166M LPT unbonded (~23.1% of unbond_by_delegate total), dominated by a single large delegate:

  • 0x525419ff…: ~7.54M LPT unbonded (699 unique unbonders)
  • 0xdac81729…: ~1.44M LPT unbonded (203 unique unbonders)
  • 0x141e6d49…: ~1.64M LPT unbonded (134 unique unbonders)
  • 0x0d509d8b…: ~0.36M LPT unbonded (126 unique unbonders)
  • 0xe9e28427…: ~0.18M LPT unbonded (103 unique unbonders)

Interpretation: this is the closest on-chain proxy to “many delegators churned from this orchestrator” (but it still doesn’t tell you why: performance, commission changes, whales rotating many wallets, etc.).

2) Whale-driven outflows (few delegators leave, huge amounts)​

Definition: ≤10 unique unbonders (129 delegates total).

Together these delegates account for ~12.846M LPT unbonded (~26.5% of unbond_by_delegate total). The top cases are extremely whale-concentrated:

  • 0x4416a274…: ~4.64M LPT unbonded (8 unique unbonders)
  • 0xf5a88945…: ~1.39M LPT unbonded (4 unique unbonders)
  • 0x731808ad…: ~1.11M LPT unbonded (3 unique unbonders)

Interpretation: this is best explained as “a few large delegators exited or rotated”, not “retail churn”.

3) Mid-scale outflows (the majority by amount)​

Definition: 11–99 unique unbonders (67 delegates total).

This is the largest category by total unbonded amount: ~24.419M LPT (~50.4% of unbond_by_delegate total).

Interpretation: most stake outflow occurs in this “middle” regime: not single-whale exits, but not hundreds of exiting delegators either.

How much of delegate outflow is migration-claimer driven?​

For some delegates, a large fraction of unbonded stake plausibly traces back to migration claims (claimer stake claimed to that delegate):

  • 0x9c1067…: claimedStake/unbond ≈ 63%
  • 0xda43d8…: claimedStake/unbond ≈ 52%
  • 0x4ff088…: claimedStake/unbond ≈ 49%

For others, migration-claimer stake is a small fraction of total unbond outflow:

  • 0x525419…: claimedStake/unbond ≈ 10%

Whale-driven outflows: are the whales “orchestrators”?​

There are multiple “whale-driven” delegate outflow cases where a very small number of wallets unbond a very large amount. Two separate questions matter:

  1. Is the delegate an orchestrator? (i.e., a transcoder address)
  2. Are the whale wallets themselves orchestrators, or just delegators?

Delegate attribution (service URIs)​

On Arbitrum, orchestrator “service URIs” are stored in ServiceRegistry (Controller key ServiceRegistry).

  • ServiceRegistry: 0xC92d3A360b8f9e083bA64DE15d95Cf8180897431
  • Query used: getServiceURI(address)(string)

Examples of whale-driven delegates and their service URIs:

  • 0x4416a274… (registered transcoder): "https://livepeer-orchestrator.prod.dcg-labs.co:8935"
  • 0x104a7ca0… (registered transcoder): "https://node.eliteencoder.net:8935"
  • 0x21d1130d… (registered transcoder): "https://93.115.27.44:8935"
  • 0xf5a88945… (registered transcoder): "https://54.144.38.201:8935"

Whale wallet status​

For the top whale wallets driving these delegate outflows:

  • BondingManager.isRegisteredTranscoder(address) returns false (i.e., they are not transcoders/orchestrators).
  • They are delegators (EOAs) that can be bonded to a delegate/orchestrator.

Interpretation: whale-driven outflows are often “whales leaving (or rotating) an orchestrator”, not “orchestrators exiting”.

Funding-source clustering (wallet linkage signals)​

Purpose: look for clusters that suggest a small number of entities controlling many wallets (sybil-like), or internal wallet shuffles.

Tooling:

  • Script: tools/livepeer/arb_funding_source_clustering.py
  • Output: artifacts/livepeer-delegator-flows/funding_cluster.json

Heuristics used (read as signals, not proof):

  • Migration claim sources via L2Migrator.StakeClaimed (delegate + claimed stake)
  • Inbound LPT Transfer events in a backward block window before a wallet’s first observed Bond event

In the whale-driven cohort around the 0x4416a274… orchestrator, there is a direct funding edge:

  • 0x3d6182c5… (migration claimer; claimed ~1.81M LPT on 2022-02-27 to 0x9c10672…) → transferred 1,000,000 LPT to
  • 0xc5519fd1… shortly before 0xc551…’s first observed Bond to 0x4416a274…

Interpretation: this looks like wallet linkage / internal capital movement, not a mass of unrelated small wallets.

Bridge-mint pattern​

Several large wallets’ “funding source” appears as Transfer(from=0x0000… , to=wallet) near bonding time, which is consistent with canonical bridge minting on L2 (not necessarily sybil behavior).

Whale exits often look like bridge-out burns (Transfer to zero)​

Many of the largest withdrawers by amount show large LPT Transfer(from=wallet, to=0x0000…) events, and the underlying txs are highly consistent:

  • tx.to == 0x5288c571… with selector 0x7b3a3c8b (gateway-like)
  • the Transfer(... → 0x0000…) amount matches the value being exited

Interpretation: a large share of whale exits are not “sold on Arbitrum” in-place; they appear to be bridge-out / burn flows that remove LPT from Arbitrum supply (likely bridging to L1 or elsewhere).

Examples (top withdrawers):

  • 0x3d6182c5…: ~1.73M LPT burned across 16 txs (2023-02-09 → 2024-12-23)
  • 0xc5519fd1…: ~1.16M LPT burned across 18 txs (2025-02-18 → 2026-01-14)
  • 0x962b0295…: ~1.05M LPT burned in 1 tx (2025-09-17)

Artifacts:

  • Burn summary: artifacts/livepeer-delegator-flows/top_withdrawers_burn_to_zero_summary.json
  • Burn date ranges: artifacts/livepeer-delegator-flows/top_withdrawers_burn_to_zero_daterange.json
  • Post-withdraw outflow window sample (includes tx.to + selector): artifacts/livepeer-delegator-flows/top_withdrawer_post_withdraw_outflows_window2m.json

“Minter” transfers are often protocol-funded withdrawals (not external funding)​

When clustering “fast-exit small” wallets, a surprisingly common inbound LPT Transfer source is the Livepeer Minter contract (0xc20de371…).

By inspecting the underlying transactions for these Minter → wallet transfers:

  • 41 / 42 are tx.to == BondingManager proxy (0x35Bcf3…) with selector 0x25d5971f (i.e., withdrawStake(...)).
  • The receipt typically contains a single LPT Transfer log from Minter to the withdrawing wallet plus the BondingManager.WithdrawStake event.

Interpretation: in this cohort, “Minter-funded wallets” is primarily a protocol payout pattern at withdrawal time, not “someone funding many new wallets”.

Artifacts:

  • Minter audit output: artifacts/livepeer-delegator-flows/minter_transfer_fast_exit_small_audit.json

Fast-exit small cohort: infrastructure vs sybil signals​

Cohort definition:

  • First withdraw ≤ 30 days from first bond
  • max_bonded_amount ≤ 1,000 LPT
  • Cohort file: artifacts/livepeer-delegator-flows/cohort_fast_exit_small_30d_1000lpt.txt (275 wallets)

Funding clustering output:

  • artifacts/livepeer-delegator-flows/funding_cluster_fast_exit_small.json

Key patterns:

  • Bridge mints: 51 wallets show Transfer(from=0x0000…, to=wallet) where tx.to == 0x6d2457a4… and selector 0x2e567b36 (bridge/gateway-like). In these txs, tx.from is often 0x7253f1c8… or 0x0000… (system/L1→L2 message execution), so tx.from clustering here is not a reliable “EOA funder”.
  • Protocol withdrawals: 42 wallets show Transfer(from=Minter, to=wallet), and as above this is almost entirely explained by withdrawStake(...) calls to BondingManager.
  • Small, sybil-like batch: one EOA (0xee3da44e…) sends 2 LPT (ERC20.transfer, selector 0xa9059cbb) to 11 wallets in a tight block range (~371.75M–371.78M). Those wallets then:
    • first bond on 2025-08-24
    • unbond on 2025-09-12
    • withdraw on 2025-09-21 (earliest-possible timing)
    • bond across many different delegates (not a single orchestrator)

Cash-out behavior (within protocol + post-withdraw consolidation):

  • Total withdrawn from BondingManager across the 11 wallets: ~51.95 LPT
  • Post-withdraw, they consolidate:
    • ~16.44 LPT moved between the 11 wallets (internal consolidation into 1–2 “collector” wallets)
    • ~29.32 LPT moved from the 11-wallet set to a single external EOA (0x1d863e2a…)
  • The external EOA then re-bonds LPT via BondingManager (selector 0x6bd9add4) rather than transferring to DEX routers/pools in this window.

Artifacts:

  • Batch post-withdraw internal vs external flow summary: artifacts/livepeer-delegator-flows/sybil_batch_ee3da44e_post_withdraw_internal_external.json
  • External collector wallet outgoing LPT transfers: artifacts/livepeer-delegator-flows/ee3da44e_aggregator_1d863_lpt_outgoing.json

Interpretation: there is at least some “farm addresses for delegator-count/eligibility” behavior in the long tail; in this instance it looks like a coordinated batch that withdraws and then consolidates into a collector that re-stakes.

Evidence for/against “sybil farmers cashing out”​

What the data supports​

  • The largest outflow-by-amount days look whale-driven (single-digit to tens of withdrawers).
  • Periods with higher “many-withdrawer” days exist, but maxima are ~30–45 withdrawers/day.
  • Migration-claimer exits exist (non-trivial “claim → withdraw within ~30d”), but stake-weighted fast exits appear concentrated in a small number of large wallets, not “mass thousands of small accounts” in aggregates.
  • In a large “fast exit small” cohort (withdrew within 30 days and max stake ≤ 1,000 LPT), exits are spread across many delegates (no obvious single-orchestrator funnel).
  • Within that cohort, there are some clear automation-looking clusters (e.g., a single EOA funding many wallets that bond/unbond/withdraw in lockstep), but they appear small by stake.
  • There are “dust”/automation-looking cohorts (e.g., many tiny new bonders on a single day), but they do not obviously map to later large withdrawals in the aggregates.

What we cannot yet conclusively rule out​

  • A coordinated actor could still operate many wallets and exit over many days with small amounts.
  • Without funding-source clustering (e.g., “all these wallets were funded by the same 1–3 addresses”), we can’t definitively label/quantify sybil activity.

Next research step (if we want a stronger answer)​

Do funding-source clustering on suspect cohorts:

  • Select cohorts such as:
    • “fast exit small” addresses
    • top “many-withdrawer” weeks
    • specific “spike days” by withdrawer count
  • For each address, fetch inbound LPT Transfer logs around their first bond window to identify common funders.
  • Optionally include TransferBond scanning to capture “entry” via transfers (so we don’t miss “new delegators” that were actually farm wallets receiving stake).

Why this matters for tokenomics design​

If outflows are mostly whale-driven (by amount), then “small-delegator count” problems are likely more about:

  • product/UX friction
  • unclear expected yield
  • lack of liquidity (no LST path)
  • lack of retention incentives

If we want to aggressively boost small delegators, any “per-address progressive rewards” design is sybilable unless we add:

  • identity attestation (optional or required for the highest boosts), and/or
  • proof-of-usage/proof-of-spend gates that make sybils economically expensive.

Addendum (2026-01-17): “Dormant vs active” unbonders (tx-gap proxy)​

The earlier “dormant” proxy (based on EarningsClaimed) turned out to be misleading because Bond/Unbond transactions often emit EarningsClaimed in the same tx.

So we re-scanned the full window and classified “dormancy” by time since the delegator’s previous distinct BondingManager transaction (any of: Bond, Unbond, Rebond, WithdrawStake, EarningsClaimed, TransferBond).

Artifacts:

  • Full scan output: artifacts/livepeer-bm-scan-arbitrum-v2/daily.json
  • Per-Unbond NDJSON: artifacts/livepeer-bm-scan-arbitrum-v2/unbond_events.ndjson
  • Summary report (fast, uses embedded timestamps): artifacts/livepeer-bm-scan-arbitrum-v2/unbond_report.embedded.md
  • Scanner: tools/livepeer/arb_bondingmanager_scan.py
  • Reporter: tools/livepeer/arb_unbond_events_report.py

Key results (full window, through 2026-01-17):

  • Unbond events: 24,059
  • Unique unbonders: 3,290
  • Total unbonded: ~48.43M LPT
  • “Exit <30d” (first bond → unbond): 1,548 events, ~8.56M LPT

Prev-tx gap buckets (event-level, amount-weighted):

  • <=30d: 21,011 events, ~34.33M LPT
  • >=90d: 1,939 events, ~8.76M LPT
  • mid (30–90d): 1,109 events, ~5.34M LPT

First unbond per unbonder (unique-unbonder classification):

  • <=30d: 1,527 unbonders (46.4%)
  • >=90d: 1,350 unbonders (41.0%)
  • mid: 413 unbonders (12.6%)

Day-by-day “who unbonds” highlights:

  • Largest unbond-by-amount days are still whale-driven (few wallets dominate the day’s amount), e.g.:
    • 2024-03-07: ~812.5k LPT unbonded by a single wallet with a >=90d gap (dormant-exit shaped).
    • 2024-12-10: 1.0M LPT unbonded by a single wallet.
  • Largest day by unique unbonders is only 122 wallets (2025-05-30) — not thousands — which is inconsistent with “mass sybil cashout wave” in the raw daily counts (but does not rule out smaller trickle-style sybil behavior).

Addendum (2026-01-18): Cumulative staking rewards + “cashout” bounds (Arbitrum)​

To answer “who extracted value systematically” and “did rewards stay staked or get cashed out?”, we ranked all addresses by cumulative EarningsClaimed.rewards and compared against on-chain WithdrawStake.amount.

Artifacts:

  • Report: artifacts/livepeer-bm-scan-arbitrum-v2/earnings_report.md
  • Data: artifacts/livepeer-bm-scan-arbitrum-v2/earnings_report.json
  • Tool: tools/livepeer/arb_earnings_report.py

Key results (through 2026-01-17, spot conversion uses $3.29/LPT passed to the report tool):

  • Total staking rewards claimed: 17,495,206.038769 LPT (≈ $57.56M at $3.29).
  • Reward concentration (by rewards): top 10 = 33.6863%, top 100 = 75.5056%.
  • Rewards claimed by wallets with any withdraw: 15,150,251.289056 LPT (86.6% of all rewards).
  • “Rewards cashed out” (observable only as stake leaving BondingManager, not as fiat conversion):
    • Upper bound (rewards-first attribution): 11,291,848.749311 LPT (≈ $37.15M).
    • Proxy (withdraw beyond bond_additional + stake_claimed, capped by rewards): 7,197,879.226981 LPT (≈ $23.68M).
  • GPU equivalence example (if you assume $5,000 capex per GPU):
    • Total rewards: ≈ 11,511.85 GPUs
    • Upper bound withdrawn rewards: ≈ 7,430.04 GPUs
    • Proxy withdrawn rewards: ≈ 4,736.20 GPUs

Important caveat: This does not prove “selling” (DEX/CEX). It measures stake exiting the staking contract. Full sell-tracing would require following LPT Transfer logs after withdraw and classifying destinations (routers, CEX deposit addrs, etc.).

Addendum (2026-01-18): What top “cashout” wallets did after withdrawing (post-withdraw LPT transfers)​

We additionally traced LPT ERC20 Transfer events after each wallet’s first WithdrawStake to see whether liquid LPT tended to:

  • re-enter BondingManager (likely re-staking), vs
  • leave to other addresses (could be selling/bridging/self-custody moves)

Artifacts:

  • Report: artifacts/livepeer-bm-scan-arbitrum-v2/post_withdraw_lpt_transfers_top50.md
  • Data: artifacts/livepeer-bm-scan-arbitrum-v2/post_withdraw_lpt_transfers_top50.json
  • Tool: tools/livepeer/arb_post_withdraw_lpt_transfers.py

Top 50 wallets by proxy rewards withdrawn (spot uses $3.29/LPT):

  • Total WithdrawStake (liquid LPT leaving BondingManager): 14,914,258.407986 LPT (≈ $49.07M).
  • Total post-withdraw outgoing LPT to non-BondingManager destinations: 18,108,291.978155 LPT (≈ $59.58M).
  • Of that, outgoing “burn” transfers to 0x000…000 (often consistent with bridge-out style burns): 7,933,388.951882 LPT (≈ $26.10M).
  • Many wallets show near-1:1 behavior where withdraw ≈ outgoing_non_bm, suggesting a strong tendency to move withdrawn stake out of the wallet rather than keep it liquid locally or re-stake.
  • GPU equivalence example (if you assume $5,000 capex per GPU):
    • Total withdraw: ≈ 9,813.58 GPUs
    • Total outgoing non-BM: ≈ 11,915.26 GPUs
    • Proxy withdrawn rewards (top-50 cohort only): ≈ 3,506.66 GPUs

Note: Outgoing transfers can exceed withdraw totals because we count all LPT transfers from the wallet after its first withdraw (it may have had LPT from other sources too). This is still strong evidence of “value leaving the staking contract”, but not definitive proof of “sold for USD” without destination labeling.

Addendum (2026-01-18): Delegator retention curves (first bond → first unbond/withdraw)​

To move beyond anecdotes, we generated event-based retention curves using the full BondingManager scan:

  • “New delegator” = first Bond event seen for an address
  • Churn proxies:
    • first Unbond (any partial unbond counts)
    • first WithdrawStake (stake leaves BondingManager)

Artifacts:

  • Report: artifacts/livepeer-bm-scan-arbitrum-v2/retention_report.md
  • Data: artifacts/livepeer-bm-scan-arbitrum-v2/retention_report.json
  • Tool: tools/livepeer/arb_delegation_retention_report.py

Key results (eligible-only to avoid right-censoring, through 2026-01-17):

  • New delegators (first bond): 4,887
  • Ever unbonded: 2,684
  • Ever withdrew: 2,322
  • Overall churn within N days:
    • <=30d: 10.78% unbonded, 6.32% withdrew
    • <=90d: 17.24% unbonded, 12.56% withdrew
    • <=180d: 23.29% unbonded, 18.16% withdrew
    • <=365d: 32.87% unbonded, 26.39% withdrew

Interpretation: there is real early churn, but it’s not “everyone instantly farms and leaves”; much of churn happens over months.

Addendum (2026-01-18): Outflow destination labeling (bridge vs DEX vs transfers)​

We added a classifier that labels post-withdraw LPT outflows for “big cashout” wallets using:

  • Transfer(to=0x0) as a strong bridge/burn signal on Arbitrum
  • receipt-based “DEX sale likely” when the same tx shows incoming USDC/USDT/DAI/WETH to the wallet
  • 4byte selector hints as a weak signal (collision-prone; used only as a fallback)

Artifacts:

  • Top-50 report: artifacts/livepeer-bm-scan-arbitrum-v2/outflow_destination_classification_top50.md
  • Data: artifacts/livepeer-bm-scan-arbitrum-v2/outflow_destination_classification_top50.json
  • Tool: tools/livepeer/arb_outflow_destination_classify.py

Key result (top 50 by proxy rewards withdrawn; limited to top 20 tx per wallet by LPT out for speed):

  • Total outgoing LPT analyzed: 17,666,111.461738 LPT
  • Bridge/burn-like (to 0x0): 7,890,569.262153 LPT
  • Direct transfers (tx.to is LPT token; mostly EOAs): 6,582,511.577837 LPT
  • DEX swap signals (receipt-based): ~31k LPT “likely” (+ ~21k LPT “possible” via selector hints)

Important caveat: the DEX heuristic will miss sales where the wallet receives native ETH (not WETH) or routes proceeds to a different address. So “DEX sold” here is a lower bound. Even with that caveat, bridge/burn + EOA consolidation dominate the observed flows in the biggest-wallet cohort.