Field Value
Case GRC-GONKA-265 - Kimi operator restitution: external attack (e265–e266) + ComputeGroupCap underpayment (e267–e276)
Epochs affected 265–276
Affected participants 53 unique addresses
Estimated compensation 946,509.93 GONKA
Cause and evidence External actor sent malicious requests crashing Kimi vLLM nodes in e265–e266, causing CPoC degradation and mass nonce exclusion; the resulting N-1 weight collapse triggered ComputeGroupCap inversion every epoch through e276. Evidence: weight_fluctuation_analysis.md, per-epoch compensation scripts and JSON in e265/–e276/.
Can it happen again? ComputeGroupCap: No known repeat path — resolved by v0.2.13 WeightScaleFactor adjustment. External vLLM attack: Reduced risk — the specific crash vectors (stop_token_ids+min_tokens OOM, prompt_logprobs OOM, JSON schema recursion bomb, Jinja injection via chat_template_kwargs) are patched at the devshard gateway (PRs #1170, #1171, #1180, merged May 16–18). A proxy geo-module bug (PR #1183) that made all IP-based rate-limiting ineffective during e265 was also fixed May 18. No known remaining crash path, but gateway coverage of unknown vLLM attack surface cannot be guaranteed.
Mitigation / fix Nonce exclusion bug: fixed before e267 (version unconfirmed). ComputeGroupCap: v0.2.13 at block 4,267,300 (WeightScaleFactor for Kimi reduced to 0.78). External vLLM attack: devshard gateway patches merged May 16–18 2026 — PR #1170 (stop_token_ids+min_tokens strip, the confirmed e264/e265 attack vector), PR #1171 (prompt_logprobs OOM strip), PR #1180 (CVE-class defenses: schema recursion, Jinja injection, body depth pre-scan), PR #1183 (proxy geo-module bug fix — rate-limiting was non-functional during the attack).
Compensation overlap Requires investigation — potential address overlap with separate restitution tracks not yet audited.
Current decision GRC voted to exclude this case from proposal #4 (6 exclude, 2 include, 1 abstain). Calculations are published here for community reference if anyone chooses to bring this forward independently.
Review focus Highest risk: e266 nonce reconstruction methodology (18 participants, 183,605 GONKA); confirm reconstructed weights match on-chain nonce commits in poc_commits/.
GRC Position
E265–E266 (external attack): The lost rewards in epochs 265 and 266 were a direct result of a third-party attack on Kimi vLLM nodes. Losses caused by external actors are outside the scope of the Gonka Restitution Committee. These epochs cannot be included in a GRC proposal.
E267–E276 (ComputeGroupCap): The cap breach was triggered by the attack-induced N-1 weight collapse in e266, but it was not caused solely by that event. As noted by a committee reviewer:
"Additionally, the GroupCap issue was not caused solely by recovery from the attack. It was also influenced by the significant weight imbalance between Kimi and Qwen. In fact, update 0.2.13 specifically reduced Kimi's weight for this reason. GroupCap was effectively acting as a balancing mechanism to normalize profitability between the models."
This context further reduces the case for GRC restitution: the cap was partly correcting a structural imbalance, not purely an anomalous malfunction.
GRC vote outcome: The committee voted against including this case in the next proposal.
Vote Count
Include 2
Exclude 6
Abstain 1
The compensation calculations in this repository were prepared and validated by two committee reviewers. They are published here so that the community can use these findings if anyone chooses to bring this proposal forward independently.
Methodology objection (GRC reviewer): One reviewer raised a concern that the current calculation overstates compensation because it does not recompute the total network weight. Their argument: the correct question is how much each participant would have received if GroupCap had never existed — which would require reconstructing the uncapped weight for every group, summing into a new total network weight, and compensating only the shortfall against that recalculated denominator. Under that approach, Kimi participants would receive less than the current calculation produces.
The counter-argument applied here: Qwen rewards have already been paid and cannot be clawed back, so there is no "same pie" to redistribute. The current approach applies the same reward formula the chain used for all other participants — using EpochGroupData.total_weight as denominator — and corrects only the Kimi side. This is consistent with how prior restitution cases were handled in this repo.
Overview
This repository contains the restitution analysis for epochs 265 through 276 of the Gonka network. Two distinct bugs caused underpayment to moonshotai/Kimi-K2.6 operators across these epochs:
Epochs 265–266: external attack — an unknown actor sent malicious requests targeting Kimi vLLM nodes starting in e265 (attack payload confirmed from production logs: stop_token_ids
min_tokens combination that triggers a vLLM CUDA assert, crashing the engine; 16 crashes across 12/14 mlnodes observed on 2026-05-14), causing CPoC confirmation weight degradation, and escalating in e266 to crash most operators and cause mass nonce exclusion; delegation compensation included for e266 delegators; gateway patches merged May 16–18 (PRs #1170, #1171,
1180, #1183)
Epochs 267–276: ComputeGroupCap underpayment — the e266 collapse destroyed the N-1 reference weight used by the cap formula, and when Kimi operators returned in e267 with full accumulated confirmation weight the mismatch caused a severe inversion; the cap remained triggered every epoch until the v0.2.13 upgrade
The cap breach was resolved by the v0.2.13 upgrade at block 4,267,300 (mid-epoch 276), which reduced Kimi's WeightScaleFactor to 0.78. Epoch 277 is the first clean epoch.
Epoch Block Heights
Epoch PoC Start Epoch End Notes
265 4,089,970 4,105,360
266 4,105,361 4,120,751
267 4,120,752 4,136,142
268 4,136,143 4,151,533
269 4,151,534 4,166,924
270 4,166,925 4,182,315
271 4,182,316 4,197,706
272 4,197,707 4,213,097
273 4,213,098 4,228,488
274 4,228,489 4,243,879
275 4,249,774 4,259,270
276 4,264,130 4,279,520 true epoch end; v0.2.13 upgrade at 4,267,300
Reward formula: 323,000 × e^(−0.000475 × (epoch − 1)) GONKA.
Eligibility Criteria
Epoch 265 (external attack — CPoC degradation): Participants whose confirmation weight dropped abnormally mid-epoch at block 4,103,171 due to the attack. Compensation uses weight / total_epoch_weight × epoch_reward − actual_rewards.
Epoch 266 (external attack — nonce exclusion): The attack escalated, crashing most Kimi operators and causing mass nonce exclusion. Participants who submitted valid Kimi nonces but were never registered in the epoch group (or had their Kimi contribution zeroed) are eligible. Delegation compensation is also included for participants whose operator was excluded.
Epochs 267–276 (ComputeGroupCap): A participant is eligible only if they successfully completed the epoch — meaning they both confirmed work (confirmation_weight on-chain) AND received actual rewards. Participants who failed the epoch for any reason are excluded: the bug caused underpayment to healthy participants, not an obligation to pay those who didn't complete the epoch.
Delegation Impact (Epoch 266)
Gonka uses a PoC delegation system where participants without direct MLNodes can delegate their consensus weight to an operator. When an operator fails to enter the epoch (due to the nonce bug), the chain resolves all their delegators into ModeNone instead of ModeDelegate:
Mode Condition Weight effect
ModeDelegate Operator entered epoch −5% transferred to operator
ModeNone Operator did not enter −15% deducted as penalty
The net extra loss for a delegator whose Kimi operator was excluded is 10% of their original consensus weight.
Full mechanics: e266/DELEGATION.md
Epoch 265 — External Attack (CPoC Degradation)
All Kimi operators entered the epoch group, but CPoC confirmation weights dropped abnormally at block 4,103,171 as a result of the external attack on Kimi vLLM nodes. 3 participants were affected.
compensation = max(0, weight / total_epoch_weight × epoch_reward − actual_rewards)
Address weight cw@healthy cw@end drop Owed (GONKA)
gonka1j7x6dv42xehe9e5au4ku3wvzwtqlegfjhlvzj6 66,311 66,311 323 99.5% 20,896.53
gonka17pw6099q758qwzewtrqmqpf5c2lrhr97fwqexu 189,884 186,719 172,607 7.6% 5,444.49
gonka1830lqug50lse998x2lakk4pj5ypfumz5pasz0y 13,490 7,031 0 100% 4,251.09
Epoch 265 Total: 30,592.10 GONKA (3 participants)
Output: e265/compensation_265.csv · e265/compensation_265.json
Epoch 266 — External Attack (Nonce Exclusion)
The attack escalated in e266, crashing most Kimi operators' inference nodes and causing mass exclusion from the epoch. 9 participants submitted Kimi nonces and had their commits recorded on-chain but were never registered in the epoch group. A further 9 participants entered the epoch but with their Kimi contribution partially zeroed out. The on-chain total epoch weight was 335,159 — the correctly reconstructed weight from all nonce commits is 1,079,698, meaning ~69% of the network's true Kimi work was invisible to the reward calculation.
Part 1: Nonce compensation (18 participants)
compensation = max(0, reconstructed_weight / total_reconstructed_weight × epoch_reward − actual_rewards)
Address Reconstructed weight Fair share (GONKA) Actual (GONKA) Owed (GONKA)
gonka17pw6099q758qwzewtrqmqpf5c2lrhr97fwqexu 214,919.1 56,690.25 14,350.37 42,339.89
gonka1q5xt54wncgzk7dxv9x64uln68455g83wu9tugg 113,567.5 29,956.26 0.00 29,956.26
gonka1wthc28t25pg63hzvl07rl8e8r6km6hesl6jhsz 109,684.6 28,932.04 2,148.14 26,783.90
gonka1j7x6dv42xehe9e5au4ku3wvzwtqlegfjhlvzj6 72,610.3 19,152.78 239.63 18,913.15
gonka1qa90tgczc0k5dvk4l5nvlf5y6phgm6mg22sfjv 70,111.4 18,493.62 0.00 18,493.62
gonka10mmdjau4dnj8krs7sh7t7635ttnmq9u3vqgz09 64,214.9 16,938.28 5,869.99 11,068.29
gonka1jrgm47v5eg876udmzg6j6glqcsd5x0vk6crpax 32,390.2 8,543.71 0.00 8,543.71
gonka1ujnc662v6g69jm6fgxnr79a2m7ehzeut059239 28,109.2 7,414.49 0.00 7,414.49
(10 more — see compensation_266_nonces.csv)
Part 1 total: 183,605.74 GONKA
Part 2: Delegation compensation (9 participants)
9 participants delegated their Kimi weight to gonka1q5xt54... (excluded operator). The chain penalised them 15% (ModeNone) instead of transferring 5% (ModeDelegate) — net extra weight loss of 10% per participant.
compensation = (original_weight × 0.10) / total_epoch_weight × epoch_reward
Address Extra weight lost Owed (GONKA)
gonka1tja3g2da45efhe2p83gk3whtussmgmtsdlgprt 2,124.3 1,805.14
gonka1hwvel7n3zuk6wruefuzc356l9myske9stckwnz 1,854.1 1,575.51
gonka12pcu9mcrpa4w4sjd9y3dsksnvu495ss6f9r4ra 1,372.7 1,166.44
gonka1tlvg4kjx7ljd5thgd5fkgh39q6lu8cmxupktgg 207.4 176.25
gonka1fvly5jrewyjmjfgwah3khy9rttq4cqajcesv9p 134.0 113.86
gonka1cuwejs77gectp3n32wg8q27hlsa4m3hqspf4ww 127.5 108.37
gonka1tmk2tzdneht6smu34pkmqdvu7p34qavvmwtwq2 118.3 100.57
gonka1gyk0aahvr3qeju4zx0nplfreej6cy4jjk8svc5 36.5 30.99
gonka14ef2pxjge75gflqftn7m2wy0xv59gq9uc7qnct 18.4 15.60
Part 2 total: 5,092.73 GONKA
Epoch 266 Grand Total: 188,698.47 GONKA (18 nonce + 9 delegation participants)
Output: e266/compensation_266.json
Epoch 267 — ComputeGroupCap (worst epoch)
All Kimi operators entered the epoch, but the Kimi group's raw weight (899,487) far exceeded the ComputeGroupCap (0.75 × epoch 266 total weight 335,159 = 251,369). The e266 total weight was severely depressed by a targeted external attack on Kimi vLLM nodes, which crashed most operators and collapsed network weight to 335k. This destroyed the N-1 reference weight used by the cap formula for e267 — when Kimi operators returned with full accumulated confirmation weight, the mismatch caused the worst inversion of the entire incident (conf/weight ratio 1.75×).
compensation = max(0, confirmation_weight / EpochGroupData.total_weight × epoch_reward − actual_rewards)
Metric Value
Kimi participants in epoch 27
Affected participants 25
Total compensation 246,471.82 GONKA
Output: e267/compensation_267.csv · e267/compensation_267.json
Epoch 268 — ComputeGroupCap
Cap baseline rose (using epoch 267's total weight), but Kimi's conf_weight remained above cap at 85.2% of total network weight. Conf/weight ratio 0.85×.
Metric Value
Kimi participants in epoch 20
Affected participants 11
Total compensation 42,634.68 GONKA
Output: e268/compensation_268.csv · e268/compensation_268.json
Epoch 269 — ComputeGroupCap
Kimi conf_weight at 87.6% of total network weight. Conf/weight ratio ~1.00×.
Metric Value
Kimi participants in epoch 30
Affected participants 19
Total compensation 47,504.58 GONKA
Output: e269/compensation_269.csv · e269/compensation_269.json
Epoch 270 — ComputeGroupCap
Kimi conf_weight at 94.9% of total network weight. Conf/weight ratio 1.23×.
Metric Value
Kimi participants in epoch 23
Affected participants 19
Total compensation 76,870.08 GONKA
Output: e270/compensation_270.csv · e270/compensation_270.json
Epoch 271 — ComputeGroupCap
Kimi conf_weight at 80.5% of total network weight. Conf/weight ratio 0.95×.
Metric Value
Kimi participants in epoch 23
Affected participants 17
Total compensation 28,422.15 GONKA
Output: e271/compensation_271.csv · e271/compensation_271.json
Epoch 272 — ComputeGroupCap
Kimi conf_weight at 80.0% of total network weight. Conf/weight ratio 0.88×.
Metric Value
Kimi participants in epoch 23
Affected participants 11
Total compensation 16,988.15 GONKA
Output: e272/compensation_272.csv · e272/compensation_272.json
Epoch 273 — ComputeGroupCap (second inversion)
Kimi node count jumped to 81 nodes (30 participants). Kimi conf_weight reached 107.5% of total network weight — a second inversion. Conf/weight ratio 1.18×.
Metric Value
Kimi participants in epoch 30
Affected participants 20
Total compensation 86,243.30 GONKA
Output: e273/compensation_273.csv · e273/compensation_273.json
Epoch 274 — ComputeGroupCap
Kimi conf_weight at 88.5% of total network weight. Conf/weight ratio 0.97×.
Metric Value
Kimi participants in epoch 24
Affected participants 9
Total compensation 41,818.44 GONKA
Output: e274/compensation_274.csv · e274/compensation_274.json
Epoch 275 — ComputeGroupCap
Kimi conf_weight at 103.6% of total network weight. Conf/weight ratio 1.28×.
Metric Value
Kimi participants in epoch 24
Affected participants 18
Total compensation 89,984.78 GONKA
Output: e275/compensation_275.csv · e275/compensation_275.json
Epoch 276 — ComputeGroupCap
v0.2.13 activated at block 4,267,300 (mid-epoch), reducing Kimi's WeightScaleFactor to 0.78. Compensation is calculated against the full epoch end state (block 4,279,520) using the post-upgrade weights and rewards as recorded by the chain.
Metric Value
Kimi participants in epoch 23
Affected participants 11
Total compensation 50,281.35 GONKA
Output: e276/compensation_276.csv · e276/compensation_276.json
Grand Total
Epoch Issue Affected Compensation (GONKA)
265 CPoC degradation 3 30,592.10
266 Nonce exclusion + delegation 18 + 9 188,698.47
267 ComputeGroupCap 25 246,471.82
268 ComputeGroupCap 11 42,634.68
269 ComputeGroupCap 19 47,504.58
270 ComputeGroupCap 19 76,870.08
271 ComputeGroupCap 17 28,422.15
272 ComputeGroupCap 11 16,988.15
273 ComputeGroupCap 20 86,243.30
274 ComputeGroupCap 9 41,818.44
275 ComputeGroupCap 18 89,984.78
276 ComputeGroupCap 11 50,281.35
TOTAL
53 unique addresses 946,509.93 GONKA
Per-address breakdown: aggregate_compensation.json · aggregate_compensation.csv
Issue resolved in e277 following v0.2.13 upgrade. No further compensation required.
Repository Structure
.py │ ├── compensation_.csv │ ├── compensation_.json │ ├── epoch_commits.json │ ├── epoch_group_data.json │ └── epoch_performance.json">gonka-e26x-issue/ ├── README.md ← this file ├── weight_fluctuation_analysis.md ← full root cause analysis, Kimi weight history ├── aggregate_compensation.py ← aggregates all epochs into per-address totals ├── aggregate_compensation.json ← per-address totals (machine-readable) ├── aggregate_compensation.csv ← per-address totals (spreadsheet) ├── poc_commits/ ← raw nonce commits fetched from chain ├── e265/ ← CPoC degradation │ ├── calculate_compensation_265.py │ ├── compensation_265.csv │ └── compensation_265.json ├── e266/ ← nonce exclusion + delegation │ ├── calculate_compensation_266.py │ ├── DELEGATION.md │ ├── compensation_266_nonces.csv │ ├── compensation_266_delegation.csv │ ├── compensation_266_combined.csv │ └── compensation_266.json ├── e267/ … e276/ ← ComputeGroupCap (one directory per epoch) │ ├── calculate_compensation_.py │ ├── compensation_.csv │ ├── compensation_.json │ ├── epoch_commits.json │ ├── epoch_group_data.json │ └── epoch_performance.json
Running the Analysis
Requires the inferenced binary and archive node access. Configure via .env:
:26657 INFERENCED_BINARY=/path/to/inferenced">ARCHIVE_NODE_URL=http://:26657 INFERENCED_BINARY=/path/to/inferenced
The .env is loaded from ../gonka-segment-report/.env.
Run any epoch's calculator
python3 e267/calculate_compensation_267.py
Regenerate aggregate per-address totals
python3 aggregate_compensation.py
Chain Query Reference
All queries use --height pinned to the epoch end block to read state before pruning.
\ --node --height -o json
Epoch group data (validation weights, confirmation weights)
inferenced query inference show-epoch-group-data \ --node --height -o json
Per-participant reward summary
inferenced query inference show-epoch-performance-summary-by-participant \ --node -o json
Delegation snapshot (query at poc_start - 500)
inferenced query inference poc-delegation \ --node --height -o json"># PoC nonce commits for an epoch inferenced query inference all-poc-v2-store-commits poc_start> \ --node ARCHIVE_NODE> --height epoch_end> -o json
Epoch group data (validation weights, confirmation weights)
inferenced query inference show-epoch-group-data epoch> \ --node ARCHIVE_NODE> --height epoch_end> -o json
Per-participant reward summary
inferenced query inference show-epoch-performance-summary-by-participant \ epoch> address> --node ARCHIVE_NODE> -o json
Delegation snapshot (query at poc_start - 500)
inferenced query inference poc-delegation address> \ --node ARCHIVE_NODE> --height poc_start - 500> -o json
Epoch poc_start epoch_end
265 4,089,970 4,105,360
266 4,105,361 4,120,751
267 4,120,752 4,136,142
268 4,136,143 4,151,533
269 4,151,534 4,166,924
270 4,166,925 4,182,315
271 4,182,316 4,197,706
272 4,197,707 4,213,097
273 4,213,098 4,228,488
274 4,228,489 4,243,879
275 4,249,774 4,259,270
276 4,264,130 4,279,520