Every published ceiling carries this model id. A parameter set is frozen once a ceiling has been published under it — changing a number means a new version, so that a ceiling from last week stays recomputable this week.
CreditWhat an agent may owe, and what it pays to owe it.
MODEL_ID
tab-v3
APR_BP
A 6.00% · B 8.50% · C 12.00% · Unrated 12.00%
TIER_MULTIPLE
A 3x · B 2x · C 1x · Unrated 0x
STARTER_CEILING
0.250000
INTEREST_ROUNDING
down (never in the agent's favour)
RampAsymmetric on purpose: trust is slower to earn than to lose.
RAMP_START
25.00%
RAMP_STEP_CLEAN
+15.00%
RAMP_STEP_MISSED
−30.00%
RAMP_CLAMP
0.00% – 100.00%
RAMP_ROUNDING
down (compounds)
IndependenceEarnings only count as credit if they came from someone the agent does not control.
MAX_FUNDING_HOPS
3
CONCENTRATION_CAP
40.00%
UNATTESTED_DISCOUNT
60.00%
AGE_FULL_DAYS
7
tuned for testnet
Tuned down for testnet. Every testnet account is days old, so a mainnet-realistic age factor rejects every agent — including the honest one.
Independence discountsApplied to a counterparty’s weight. They MULTIPLY and truncate down.
RECIPROCAL_FLOW
50.00%
RECIPROCAL_THRESHOLD
25.00%
SHARED_FUNDING_ROOT
70.00%
YOUNG_ACCOUNT
60.00%
CONCENTRATED
80.00%
UNVERIFIED_FUNDING
50.00%
tuned for testnet
New in v3. A counterparty whose funding provenance was never observed and never published used to count in FULL, because absent ancestry read as absent relationship — the unsafe direction, and how the loop attacker went uncaught on its first full run. A discount rather than a block, because blocking would turn routine Mirror Node lag into a refusal for every counterparty at once. It is self-healing: once provenance is observed it is published, and a published fact is never forgotten.
WindowsHow often the tab settles, and the most one call may cost.