LLMSkillHub
mirroredCopied from PlamenTSV/plamen under MIT. PlamenTSV has not claimed this page, and is not affiliated with LLM SkillHub.

plamentsv / peer-scoring-correctness

L1 trigger - audits peer reputation and scoring logic for symmetry, farming resistance, and penalty coverage.

0
SKILL.md
3.9 KB

Injectable Skill: Peer Scoring Correctness

L1 trigger: P2P flag AND (score_peer, peer_score, reputation, misbehavior, ban_peer, peer_scoring detected)
Inject Into: depth-network-surface
Language: Go and Rust
Finding prefix: [PSC-N]

1. Reward / Penalty Symmetry

Enumerate all score increments and all score decrements. For every reward path,
identify the matching penalty path and ask whether a malicious peer can gain
trust faster than it can lose it.

Tag: [PEER-SCORE:SYMMETRY]

Required full-leg matrix: Do not stop after the first scoring bug. Build a
complete table of every score mutation and every peer-performance observation:

LegEvent / endpointReward?Penalty?Error classes penalizedCan attacker farm?
block validationvalid / invalid block
block pool / orphan handlingunknown parent, invalid data, timeout
health checkHTTP 200, non-200, timeout, parse error
data/chunk requestdelivery success, delivery fail, no response
gossipvalid message, invalid message, duplicate, re-gossip
bootstrap / peer listuseful peer, bogus peer, stale address

Every blank Penalty? for a harmful event is a candidate finding. Every blank
Can attacker farm? for a reward event is incomplete analysis.

2. Penalty Coverage

Build a failure-mode table and verify every harmful action has a score impact:

Failure modePenalized?Immediate?Resettable?
Invalid block / tx
Timeout / stall
Malformed gossip
Health-check failure
Excessive requests
Non-response / timeout
Non-200 HTTP response
Delivery failure after accepted request
Invalid block held in cache / orphan pool
Stale peer address update

Tag: [PEER-SCORE:COVERAGE]

3. Farming / Free-Riding

Check whether peers can cheaply farm score via ping loops, get_data style
requests, acknowledgements without contribution, or reconnect-based resets.
Quantify attacker cost versus defender work.

For each positive score path, identify the cost the peer actually pays. A peer
must not gain score for:

  • returning any HTTP 200 when the requested data was not delivered;
  • responding to health checks without serving blocks/chunks;

  • causing the victim to fetch data from the attacker while the attacker withholds
    useful data;

  • reconnecting to clear a negative history and immediately earning fresh rewards.

Tag: [PEER-SCORE:FARMING]

3b. Direction and Sign Sanity

For every increase_score, decrease_score, reward, penalize, or raw
score +=/-= site, verify the sign matches the semantic event. Common miss:
error branches log failure but do not call the penalty path, or non-200
responses are treated as success because transport completed.

Tag: [PEER-SCORE:DIRECTION:{file}:{line}]

4. Reset and Persistence Semantics

  • Does disconnect/reconnect reset score?
  • Is score keyed by peer ID, IP, ENR, or connection object?
  • Can bans be bypassed by cheap identity rotation?
  • Are penalties persisted across restart?

Tag: [PEER-SCORE:RESET]

5. Selection Stability

Trace where score affects sync-source selection, top-peer preference, or
bandwidth allocation. A manipulable score that changes any of those choices is
an enabler finding even if the scoring bug alone is not terminal.

Tag: [PEER-SCORE:SELECTION]

6. Output Requirement

Always include the full-leg matrix in the output, even when every row is SAFE.
If a row is not applicable, write N/A and cite the file/line proving the
protocol does not implement that leg. Missing rows are treated as incomplete
peer-scoring coverage.