Methodology behind the StableLens Score

1. Mission

Stablecoins now settle more daily volume than Visa. They underpin DeFi, on-ramp/off-ramp infrastructure, and an increasing share of institutional treasury operations. But there is no neutral, third- party rating system for them — only issuer marketing, sporadic journalism, and DeFi dashboards focused on TVL.

StableLens exists to fill that gap. We score every tracked stablecoin on a fixed six-dimension compliance framework, monitor pegs continuously, and surface yield opportunities with an honest risk-adjusted view. We differ from generalist on-chain analytics platforms (Glassnode, Nansen, DeFiLlama) on three fronts: stablecoin-specialized, multi-jurisdictional regulatory coverage, and methodology-versioned with an audit trail per score (with the multi-agent audit trail described in Section 2 shipping incrementally; current scoring is deterministic and reproducible).

2. Compliance scoring framework — six dimensions

The compliance scoring framework below is a separate layer from the venue-aware yield-risk model surfaced in section 2.5 below. Compliance scores rate the asset (issuer transparency, regulatory posture, treasury composition); yield-risk scores rate a position in a pool (smart-contract surface, exit cost, chain finality). They share inputs in places but compose independently.

Composite scores are 0–100, mapped to a 10-step letter grade ladder — full thresholds in section 3 below. The mapping is deliberately tighter at the extremes: AAA requires 95+; D triggers below 20.

Per-dimension detail

Honest about proxies: dimensions marked proxy use a heuristic input until the proper data pipeline lands. For example, Audit & Code Security currently aggregates audit-firm tier and exploit history but doesn't yet read individual audit reports for finding-severity. Each proxy dimension carries its own roadmap milestone in the engineering backlog.

How StableLens rates DeFi vs CeFi yield

A yield product's risks depend on where it lives. A DeFi pool exposes you to smart-contract code, oracle dependencies, bridge surfaces, and AMM liquidity. A CeFi platform exposes you to custody arrangements, withdrawal gating, the platform's solvency, and the integrity of its proof-of-reserves. Grading them with the same dimension set forces inapplicable signals to average out to neutral — which is dishonest about the actual risks.

The applicability principle: a dimension that doesn't apply to a venue is removed from the model for that row, not neutral-filled. The remaining dimensions are re-normalized so the weights still sum to 1.00. Two consequences:

• A CeFi row carries zero contribution from smart-contract, oracle-bridge, settlement-chain, AMM-liquidity, or emission-yield dimensions — those are not in its profile at all.
• A DeFi row carries zero contribution from custody, PoR cadence, withdrawal-gating, or platform-licensing dimensions — same reason.

DeFi risk profile

17 dimensions. The 12 universal dimensions (asset-level) + 5 DeFi-specific dimensions (code, oracle/bridge, chain, AMM liquidity, real-yield sustainability). Weights are preserved from the v1.1 model (Code Session U Phase U-2 calibration against the 12 historical-failure backtest).

CeFi risk profile

17 dimensions. The same 12 universal dimensions + 5 CeFi-specific dimensions (custody, proof-of-reserves cadence, platform solvency, withdrawal gating, platform regulatory standing). The 5 DeFi-only dimensions are removed; the 0.34 weight they carried is redistributed across the new CeFi-only dimensions.

Strict boundary: the platform regulatory dimension grades licensing + registration standing from public registries only (NYDFS BitLicense, FinCEN MSB, NMLS state-MSB, FCA, MAS, MFSA). It is NOT a CLARITY / GENIUS securities classification — that work is counsel-gated and lives separately in the regulatory-scoring module. The two never share inputs or weights.

Category → profile routing

The unified yield_sources universe carries a category discriminator on every row. The grader maps each category to its profile:

Confidence, coverage, and “limited data”

Every grade ships with three numbers:

• Composite (0–100) — the weighted average of present sub-scores after re-normalization.
• Confidence (0–1) — coverage-weighted average of per-dimension confidences, scaled by the coverage ratio.
• Coverage (0–1) — the fraction of the profile's applicable weight that's actually present (not missing) on this row.

A row is flagged limited data — surfaced as “—” across the UI rather than a numeric grade — when:

• composite confidence < 0.50, or
• coverage < 0.70 (less than 70% of the profile's weight is supported by real data).

When a row is limited-data, its cross-venue sort tier becomes ungraded, so a poorly-supported high score never outranks a well-supported A in the safest-first sort. This is the no-fabrication contract operationalized end-to-end.

Versioning

Profile weights carry an explicit version stamp:PROFILE_WEIGHTS_VERSION = “2.0.0”. Every persisted grade records the profile it was scored against and the weights version, so a future calibration bump (2.0.0 → 2.1.0, etc.) re-grades cleanly and the audit trail stays intact.

How StableLens benchmarks stablecoin yield against risk-free cash

A stablecoin yield is only interesting in comparison to what a US-dollar holder can earn risk-free. StableLens shows every stablecoin yield as a spread over the 3-month Treasury bill, paired with the venue's independent risk tier. The investor judges the trade-off: +170 bps for an A-tier custody risk vs +40 bps for the same are very different decisions.

The risk-free anchor

The primary anchor is the 3-month Treasury bill yield (FRED series DGS3MO), pulled live by the existing macro pipeline. Additional context rates surface on /macro: the Federal Funds rate (DFF), 2- and 10-year Treasury yields (DGS2, DGS10), and the 10Y-2Y curve (T10Y2Y).

Two reference numbers anchor the “cash” band of the spectrum and are reviewed weekly from public sources:

• FDIC national-average savings rate — the “doing nothing” baseline, published weekly by the FDIC.
• Top insured-cash reference — a single curated public reference for the upper bound of FDIC-insured cash yields available to a US saver. One number, not a maintained directory. StableLens does NOT publish a per-institution rate list, does NOT show “open account” links, and does NOT accept compensation from any deposit institution.

Excess yield (spread over risk-free)

For any stablecoin yield product:

excess_yield(row) = row.yield_apy − risk_free_rate

UI surfaces show excess yield in basis points alongside the row's StableLens grade and venue tier. The two are never collapsed into a single number — that would obscure the trade-off the product exists to expose. Cross-venue sorting uses excess yield (not raw APY) as the within-tier tie-breaker, so a 4.2% yield when T-bills pay 3.7% (+50 bps) ranks below a 5.4% yield (+170 bps) at the same tier.

The dollar-yield spectrum

The spectrum strip on /yields and /optimizer lays out the cash-to-DeFi ladder visually: risk-free cash → bank-issued stablecoins → tokenized treasuries → CeFi rates → DeFi pools. Each rung shows its top yield, the StableLens grade for that venue, and the spread over the 3-month Treasury.

Rungs with no graded venue yet render an honest “—”. Independence and provenance constraints apply uniformly: every rung's data is sourced from public / on-chain / regulatory records only, and TradFi rungs carry zero affiliate relationships of any kind.

3. Composite → letter grade & risk-adjusted APY

Each yield product's sub-scores from the venue profile above are weighted-averaged into a 0–100 composite, which maps to a letter grade with deliberately-tighter bands at the extremes. Source: lib/risk/grade.ts.

Risk-adjusted APY

Risk-adjusted APY is computed as radjAPY = APY × confidence × (1 − damping(grade)). Damping factors are tier-graded: AAA = 0 (no damping), AA = 0.05, A = 0.15, BBB = 0.30, BB = 0.50, B = 0.65, CCC = 0.75, CC = 0.80, C = 0.83, D = 0.85. A 10% APY pool with a BBB grade and 0.9 confidence becomes 6.3% risk-adjusted.

Source of truth lives in-repo at lib/risk/ with full unit-test coverage. Weights are sanity-checked at module load — the test suite fails immediately if they don't sum to 1.0.

4. Peg deviation calculation

Peg status is one of three values:

• STABLE — deviation from anchor < 0.5% over the most recent 10-minute window. Outlier ticks (single-block flash deviations) are excluded if they don't persist for two consecutive 10-minute windows.
• WARNING — deviation between 0.5% and 2.0% sustained across two consecutive windows.
• DEPEG — deviation > 2.0% sustained across two consecutive windows. Triggers depeg-event archival and alerts on Pro / Team / Institutional.

Anchor selection: USD-pegged stablecoins anchor to $1.00 (no FX adjustment; we report the deviation, the user interprets it). Algorithmic stables that drift by design (rebasing, collateral ratios) are reported with their actual peg-target if disclosed by the issuer.

Data sources: Alchemy RPC for on-chain price feeds (Chainlink + Pyth where available; falls back to DEX TWAPs), CoinGecko spot for cross-check. Disagreement > 0.3% between sources is flagged for human review and the peg status holds at the previous state until resolved.

5. Data sources

Every input to the scoring system, with refresh cadence and maturity status. We are deliberately transparent about which sources are live vs. planned — overclaiming data coverage is a trust-killer.

6. Update frequency by tier

7. Methodology versioning

Methodology changes are versioned (e.g., v1.0 → v2.0). When the framework changes, the historical scores produced by previous versions are preserved at the version that produced them. The score page shows which version produced which grade.

Material methodology updates (new dimensions, weight changes > 5%, threshold adjustments) trigger a public changelog entry and a 30-day notice before the new version takes effect on production scores. Issuers and customers can comment during the notice period.

The model version is bumped in lib/risk/version.ts whenever per-dimension logic, weights, or grade thresholds change. Current methodology version: v2.0 (locked 2026-05-26); current scoring-engine version: v1.1.0. See methodology changelog (placeholder; populated as the methodology evolves).

8. Confidence and human review

Every sub-score carries a confidence value between 0 and 1, computed from input freshness, source agreement, and historical specialist accuracy. The composite confidence is a weighted average across dimensions. A dimension with no relevant fields on the pool returns confidence 0.2–0.3 with a neutral score — this lets the UI distinguish "we know this is risky" from "we don't have enough data to say".

Look for the LOW CONF tag in the breakdown card on each yield's detail page — that's the UI surface for low-confidence composites.

Publication thresholds

• Confidence ≥ 0.90 — the score auto-publishes.
• Confidence 0.70 – 0.89 — auto-publishes but flagged for QA-auditor review within 7 days.
• Confidence 0.50 – 0.69 — held in queue. A human reviewer must approve before the score updates publicly.
• Confidence < 0.50 — auto-rejected. The previous score holds until enough evidence accumulates. The dispute process is the path forward.

Public score pages show the agent reasoning with citation links — the entire chain of evidence that led to the grade is visible to any reader, on every score, always.

9. Issuer dispute process

Issuers can dispute any score on their stablecoin. The process:

1. Issuer submits evidence at /methodology/dispute (placeholder for the Phase 2 form).
2. A human reviewer evaluates the evidence within 5 business days.
3. If the score is revised, an explicit changelog entry on the score page records: the prior grade, the new grade, the evidence accepted, the reviewer, the timestamp.
4. If the score is not revised, the issuer receives a written rationale and can re-submit with new evidence.

We do not silently revise scores. Every revision is documented publicly. This is a pre-requisite for the page being trustworthy.

10. Limitations and caveats

What StableLens is honest about not covering:

• We are not auditors. We aggregate third-party audit reports; we don't perform smart-contract code review ourselves.
• We don't make subjective business-quality judgments about issuer teams beyond the operational-health signal.
• We don't cover items not in the documented dimensions. If a dimension matters and isn't listed in section 2, it isn't in our score.
• Scoring is informational, not investment advice. A AAA grade is not a recommendation to buy. A D grade is not a recommendation to short. Stablecoin and DeFi markets carry material risk including total loss of capital.
• Methodology is opinionated. We make explicit choices about weights and thresholds. Reasonable people will disagree. Our defense is transparency: every choice is documented above.

11. The multi-agent compliance system — future roadmap

Once shipped, every StableLens score will be produced by a multi-agent system, not a single LLM call. The architecture exists to make scores reproducible, auditable, and resistant to single-point bias.

Continuous watchers will ingest evidence from primary sources — regulator filings, issuer attestations, on-chain price feeds, audit reports, exploit databases, news flow. Specialist analyst agents will interpret each dimension (one specialist per dimension; see the framework in Section 2). Each specialist will produce a sub-score with explicit reasoning and inline citations.

Two adversarial review agents — Devil's Advocate and Bull Advocate — will actively challenge proposed scores before they reach a final grade. A synthesis agent will compose the sub-scores into a composite letter grade. An editorial agent will generate the public-facing reasoning report.

A citation verification agent will confirm every source URL resolves and the quoted excerpt matches the live source. A QA auditor agent will sample 5% of scores weekly to catch systematic drift. A methodology agent will propose framework updates (new dimensions, weight changes); these require explicit human approval before deployment.

Once the agent layer ships, every score will have a complete audit trail stored append-only in public.agent_runs: which agents ran, which model was used, which inputs they saw, what reasoning produced the output, which citations supported the claims, what confidence value was assigned. Disputes can be reproduced from this trail.

Agent roles (roadmap)

12. Methodology committee

Methodology changes require explicit human approval. The committee that signs methodology updates:

• Dom (founder, methodology lead) — solo signatory until Q4 2026.
• Fractional advisors — onboarding planned at $200K ARR. Target composition: one stablecoin-issuer veteran, one regulator-facing lawyer, one quant. Names will be published here when they sign on.

13. Citation and republication policy

Journalists, researchers, and integrators are welcome to cite the StableLens Score. Recommended format:

StableLens Score: USDC, 92/100 (A · Highly Compliant), retrieved 2026-05-26 from stablelens.com/stablecoins/USDC, methodology v2.0.

Bulk republication of scores or data feeds requires a written agreement (contact partnerships@stablelens.com). Individual scores in journalism, research, or non-competing products are fair use with attribution.