Sybil / counterparty-independence model

Status: v1 IMPLEMENTED in backend/src/scoring/independence.ts (per-payer w_i = age · diversity · not_funded, concentration/HHI penalty, coarse temporal organicity → R_eff). Proven against the §1.3 attack catalog by npm run test:independence (honest passes; A1–A4 caught/discounted; A7 collusion is the documented gap, §5). Supersedes the MVP min-counterparty gate. This is the defensible core of TrustLine — see the root README "Risk model and economics". zkTLS and on-chain indexing prove revenue is real; this model is what argues revenue is independent. The two are different problems and independence is the open, fundable one.

0. Why this is the moat

Anyone can index getEvents and read an agent's USDC inflows. That is not a business — it is a SELECT statement. The hard, defensible question is:

Is this revenue from genuine, independent counterparties, or is the operator paying themselves to manufacture a credit score?

If the answer can be faked cheaply, the protocol is insolvent by construction: an attacker fabricates revenue, draws an uncollateralized credit line against it, and walks. Every uncollateralized-credit protocol lives or dies on this. The minimum-counterparty rule (MIN_COUNTERPARTIES = 3) shipped in backend/src/scoring is a floor, not a moat — it is defeated by spinning up three wallets and paying yourself. This document replaces that binary gate with a continuous, economically-grounded independence model.

1. Threat model

1.1 Attacker goal

Maximize creditLimit(agent) per dollar of real economic cost, then draw and default. The protocol is safe iff faking the revenue that unlocks a credit line costs the attacker more than the line is worth (see §2).

1.2 Attacker capabilities (assume all of these)

• Can create unlimited Stellar accounts at negligible cost.
• Can move their own USDC freely between their own accounts.
• Can stand up a real Stripe account and charge their own cards.
• Can wait (age accounts), and can script organic-looking activity.
• Cannot (cheaply) obtain other people's genuine, independent economic activity.

That last line is the only asymmetry we have, and the entire model is built to price it.

1.3 Attack catalog

2. The reframe: independence is an economic-security property

You cannot prove independence. You can only make faking it expensive. State the protocol invariant we are underwriting toward:

cost_to_fake(R_independent)  >  value_unlocked(R_independent)

where value_unlocked = limitMultiple(tier) × R minus expected recovery. Because limitMultiple is ≤3× (revenue_math), the model must ensure that manufacturing $1 of counted independent revenue costs the attacker more than ~$3 of at-risk capital they could otherwise abscond with. Everything below is in service of pricing cost_to_fake upward until the inequality holds. This framing — not a particular heuristic — is the thing to defend in a fundraise: it is a clear, falsifiable security target with a roadmap of levers.

3. Independence-v1 scoring model (on-chain)

Replace the binary onchainCounts gate with a continuous effective independent revenue R_eff and an independence score in [0,1]. All inputs are derivable from the payment graph the indexer already builds (from → agent, amount, ledger), plus cheap per-account lookups.

3.1 Per-payer independence weight w_i ∈ [0,1]

For each distinct payer i contributing revenue r_i:

w_i = age_factor_i · diversity_factor_i · not_funded_factor_i

• age_factor — min(1, account_age_days / AGE_FULL). A wallet younger than AGE_FULL (e.g. 30 days) is discounted; one born this week counts ~0. Defeats A2.
• diversity_factor — function of the payer's own out-degree: how many other, distinct parties it transacts with besides this agent. A payer that pays only this agent looks like a puppet → ~0; a payer embedded in real economic activity → 1. Defeats A1/A2.
• not_funded_factor — 0 if payer i was funded (directly, or within K hops) by the agent or a known agent-controlled address; else 1. Defeats A3.

3.2 Fund-flow loop detection

Build the directed value graph over the lookback window. Run a bounded K-hop (default K=3) reverse search from each payer back toward the agent's own funding sources. Any payer whose inbound funds originate from the agent is circular and its r_i is excluded (its not_funded_factor = 0). This is the primary defense against the cheapest realistic attack (A3).

3.3 Concentration cap

Cap how much any single payer can contribute to counted revenue, and penalize concentration via a normalized Herfindahl index:

r_i_capped = min(r_i, MAX_PAYER_SHARE · Σ r)        # e.g. MAX_PAYER_SHARE = 0.4
HHI        = Σ (r_i_capped / Σ r_i_capped)²
concentration_factor = 1 − clamp((HHI − HHI_FLOOR)/(1 − HHI_FLOOR), 0, 1)

A long tail of independent payers → low HHI → factor ≈ 1. Two payers → high HHI → factor → 0. Defeats A4.

3.4 Temporal organicity (lightweight v1)

Real revenue is irregular; scripted loops are periodic. v1 uses a coarse signal: coefficient of variation of inter-payment intervals and amount entropy, combined into organicity_factor ∈ [0.5, 1] (never below 0.5 — this is a soft signal, not a gate). Flags A5 for review without hard-failing legitimate steady revenue.

3.5 Composition

R_eff = concentration_factor · organicity_factor · Σ_i ( w_i · r_i_capped )
independence_score = clamp( R_eff / Σ r_i , 0, 1 )

R_eff (not raw on-chain revenue) feeds computeScore exactly where onchainUsdc does today. The existing MIN_COUNTERPARTIES rule becomes a cheap pre-filter, not the decision. zkTLS off-chain revenue keeps its higher weight but is itself subject to §4.

4. Off-chain independence (Stripe) — partial in v1

Self-charged Stripe (A6) is harder because we don't see cardholder identities. v1 ships signals, not a solution, and weights them into the off-chain figure:

• Stripe account age and payout history (provable via additional zkTLS fetches against Stripe's balance/payout endpoints).
• Refund / dispute / chargeback rate — a self-charge ring tends to have an anomalous (often zero) dispute profile.
• Payment-method diversity as exposed by Stripe's own aggregate metrics (count of distinct cards / customers), proven via zkTLS without revealing PII.

These are explicitly partial. Robust off-chain independence is a v2 research line (see §5).

5. What v1 does NOT solve (be honest in the pitch)

• Collusion rings (A7): partially solved. Two signals now catch the common cases: (1) net-flow reciprocity — the agent paying a "payer" back nets the revenue toward zero, killing mutual cross-pay rings; (2) co-payer-excluded out-degree — a ring member whose only counterparties are the agent + fellow members scores ~0 diversity. The remaining gap is a sophisticated, non-reciprocal ring: distinct aged operators who each pay the agent one-way (no reciprocated flow) and launder genuine external diversity so co-payer exclusion still leaves real counterparties. That is indistinguishable from real customers at the graph level and needs staking / skin-in-the-game or global community detection over a persistent graph (Track C dependency).
• Patient capital: an attacker who ages wallets and scripts diverse activity for months can still pass — but now at real time/capital cost, which is the point (§2). We raise the cost; we don't make it infinite.
• Sophisticated off-chain self-dealing with diverse real cards.

Path beyond v1: payer staking / skin-in-the-game, cross-protocol identity and reputation (reusing the stellar8004_identity interface), ZK reputation proofs, and per-agent credit ramps (limits grow with repayment history, so a cold attacker can never jump straight to a large draw — this caps value_unlocked directly and is the cheapest, highest-leverage lever to add next).

6. Implementation map

• Data: extend the indexer to retain the payment graph (from, to, amount, ledger) and expose cheap per-account lookups (age via first-activity ledger; out-degree via counterparties). This is the same persistent-indexer→DB upgrade the on-demand scanner needs anyway — independence and scale share one dependency.
• Engine: new backend/src/scoring/independence.ts computing R_eff and independence_score; computeScoreResult consumes R_eff in place of raw onchainUsdc. Surface every factor in the API so the borrower dashboard's "Score Breakdown" can show why (turns a black box into a trust feature).
• Contracts: unchanged. Independence is an off-chain underwriting concern; the on-chain score_registry still just stores the signed result.

7. Parameters (initial, all tunable)

All defaults are starting points to be calibrated against real + adversarial data; the calibration methodology is itself part of the v1 deliverable.