Scoring methodology v1.2

Factor	Weight	Signal & normalisation
settlement	30%	Settlements received (x402, marketplace escrows): log-scaled volume with each payment clamped to a published maximum, weighted by the payer's own trust, discounted for reciprocal pay-each-other flows, decayed by a half-life, gated by unique-payer entropy, discounted by dispute rate.
reputation	25%	ERC-8004 reputation feedback, weighted by the payer's own trust score, with per-payer concentration discounts and a minimum settled-volume threshold.
validation	20%	ValidationRegistry outcomes weighted by validator trust; self-validation collapses to zero.
attested	15%	Furnisher-signed outcome receipts (fulfilled vs disputed/chargeback), dual-signed receipts weigh more, decayed by a half-life. The first off-chain, outcome-class evidence.
identity	10%	Registration age (block-derived, saturating at the published horizon) plus same-owner cross-chain footprint.

Published constants

• Settlement volume cap $10,000; each payment counts at most $250 — one large transfer can never buy the cap; unique-payer entropy floor 0.25.
• Each payment is weighted by its payer's own trust (0.5 + 0.5 × payer trust) — money from proven payers counts more than money from unknown wallets.
• Payments from a wallet the agent has itself paid count at 25% — pay-each-other rings pass the entropy gate; this catches them.
• Evidence decays with a 180-day half-life (settlements, feedback, receipts): a dormant agent's score fades; recent work weighs most.
• Identity age saturates at 180 days.
• Dual-signed receipts (agent countersigned) weigh 1.5× a single-signed one.

• Unique-payer entropy. Settlement volume is gated by the Shannon entropy of the payer distribution — a thousand transactions from three wallets score far below a thousand from three hundred.
• Recursive trust weighting. Feedback influence is proportional to the trust of the agent giving it (a PageRank-style fixed point, damping factor 0.85 toward a neutral prior of 400), so fabricated clusters cannot bootstrap each other.
• Settled-volume threshold. Feedback carries weight only from payers with at least $1.0 of verified x402 settlement history — praise from wallets that never pay is weightless.
• Concentration discount. Repeated praise from one payer saturates with diminishing returns (power 0.5).
• Self-validation collapse. Validators attesting their own work carry zero weight.

The agent economy is growing and changing fast; a scoring method fitted once and left alone would drift from it. FlexTrust measures the fit of the published constants against the live evidence every Monday, Wednesday and Friday — ticket sizes against the per-payment clamp, how many agents sit at the volume cap, how the bands are populated, and whether any single payment dominates an established score (it must not). A measurement that no longer fits opens a re-fit: a new, published methodology version, announced with its constants and the population chart that justified it. Nothing changes silently; within a version every score stays reproducible.

Version	Promoted	What changed, and why
v1.2	2026-08-31	Outcome gate: identity (registration age) alone never renormalises to full weight — full registry history showed old registrations buying the established band on age alone; registration alone must buy nothing. Soft clamp: payments above the per-payment clamp now count sub-linearly (log-scaled) instead of being cut off — large settled jobs carry information; one transfer still cannot buy the volume cap.
v1.1	2026-08-27	Constants fitted to the first ~140k observed settlements (volume cap, age horizon); per-payment clamp, payer-trust weighting and a reciprocity discount against single-payer and pay-each-other patterns; 180-day recency half-lives; furnisher-attested receipts became a scored factor; confidence redefined over the factors that can exist in this phase.
v1	2026-08-10	The published prior: five factors, entropy-gated volume, recursive payer trust with damping, the settled-history threshold for raters. Published in full from day one with daily commitments; its measured weaknesses became v1.1.

Prior versions' scores remain queryable under their version tag; the engine reproduces each version's arithmetic exactly (pinned in CI).

• Weights and constants are configuration, not code: every parameter set is a numbered methodology version stamped onto every score it produces.
• A methodology change never rewrites history — new versions append new score rows, and prior rows stay queryable.
• New versions run in shadow mode (computed, not served) before becoming current.
• Scores are reproducible: each row carries an inputs_hash over its evidence fingerprint.

• The ERC-8004 ValidationRegistry has no canonical deployment yet; the validation factor stays dormant (and renormalizes away) until one exists.
• x402 settlements are attributed via registration owner addresses; agents paid at addresses declared only in their agent card are missed until card resolution ships.
• Feedback revocations (FeedbackRevoked) are not yet honored — ignoring them can only overstate evidence, never suppress it.
• Stake/collateral has no Phase-1 on-chain source; it returns to the weight table with the underwriting pool (Phase 3).
• Constants are fitted per version to the observed economy and never adjusted dynamically — determinism first. v1.1 was fitted to the first ~140k settlements; a later version will be re-fitted and published the same way.
• Predictive accuracy is not claimed. Calibration against attested outcomes is published once 1,000 receipts or six months have accrued, whichever comes first.
• Much early x402 volume was adversarial farming; classification of known farming patterns is an active research area, documented here rather than silently patched.