Skip to main content

What reputation costs and earns you

Reputation enters the selection score as inverse square:
score = rate_per_mb × rtt_ms × (1 / max(reputation, 0.1)²)
  • A node with reputation 1.0 is “baseline.”
  • A node with reputation 0.5 is worse in the score — a 50% reputation hit means you need to halve your rate (or be in a 4× closer region) to compete.
  • A node with reputation 0.1 (the floor) is 100× worse — effectively unreachable for competitive selection.
Reputation is the most leveraged variable in the score. Guarding it is more profitable than cutting rates.

How your score is computed

Other nodes’ scores of you are composites:
  • 70% from their own local observations of you.
  • 30% from gossip-propagated reports from other nodes.
You do not have a single global score — each node maintains its own view. But because gossip contributes 30%, consistent bad behavior spreads.

What moves your score

EventDirectionMagnitude
Clean delivery, valid voucher acceptedSmall
Valid probe, then stream deliveredSmall
BLAKE3 mismatch mid-streamLarge (slashable)
Probe has_blob=true then stream failsLarge (slashable)
Rate charged ≠ rate advertisedLarge (slashable)
Channel close dispute lossMedium
Declared has_blob=false (honest miss)Neutral
Failed to respond to probeNeutral or small negative
The honest-miss case is specifically protected: declaring has_blob=false is the correct response when you don’t have the blob. Penalizing it would push nodes to over-claim, which is the opposite of what the protocol wants.

Reporter weighting

A gossip report’s weight is a function of the reporter’s effective settled value:
  • Total settled on-chain settlement value (verifiable).
  • Counterparty diversity — minimum 5 distinct counterparties for full credit.
  • Time decay — half-life ~7 weeks.
  • 3× cap — no single reporter can weight more than 3× a typical reporter.
In practice: reports from a nodes with real revenue and diverse clientele carry meaningful weight. Reports from a Sybil cluster of throwaway nodes carry little.

Clamping

A single report can only move a score by a bounded amount. A coordinated attack of 100 fresh-stake nodes submitting negative reports on you at once gets clamped into a far smaller aggregate impact than the naive sum would suggest.

Cold start

New nodes start with a reputation bonus:
  • One-time per operator Ethereum address (so re-staking doesn’t get you the bonus twice).
  • Enough traffic to let you prove yourself on real deliveries.
  • Decays quickly once you have real interaction data — the bonus is a ramp, not a floor.
This means: register once with your production address. Don’t burn your cold-start bonus on test runs.

Decay

In the absence of fresh data, scores drift toward neutral. If you stop participating, other nodes’ views of you decay toward ~neutral over weeks, not vanish to zero. When you return, you start from roughly neutral — better than a new node, worse than you were.

What protects your reputation

Operational:
  • Keep probe_hold_duration at the default. Never downscale it to squeeze more cache.
  • Monitor decdn_probe_hold_slots_used — saturation above 80% means you’re at risk of evicting under-promise.
  • Monitor decdn_blacklist_sync_lag_seconds — lag here becomes blacklist-violation slashing.
  • Never rate-manipulate. Charge exactly what you advertised.
Structural:
  • Run ≥ 1 region. Serving only one region makes you invisible to other regions’ clients even when your reputation would otherwise be fine.
  • Maintain stable identity. Frequent key rotation or deregistration/re-registration cycles prevent reputation from accumulating.
  • Choose origin-backed carefully. Origin-backed nodes never phantom-announce on their own content, but they can on other nodes’ content if you enable cross-caching. Either serve only your origin’s content, or accept the normal cache-node reputation risks.

When you’re slashed

A successful slash challenge results in:
  • Stake reduced by 5% / 10% / 20% (corruption/phantom/rate, escalating) or 10% / 25% / 50% (blacklist).
  • A reputation penalty propagated via the gossip channel.
  • Reputation score drops — typically significantly, since the event generates reports from every witness.
Recovery is possible but slow. See disputes & slashing for counter-evidence paths.