Touchstone← Board

Methodology · how a grade is made

A letter grade, built from four readings of on-chain truth.

Every Touchstone rating is computed, not estimated. The engine reads a subject's on-chain state at a fixed block, scores four independent risk dimensions on transparent bands, averages them, and maps the result onto a ten-step scale from AAA to D. Claude writes the reasoning and cites its evidence — but the grade and confidence are pinned by the deterministic engine, never by the model.

The scale · AAA → D

Ten grades in four families. The composite score (0–100) lands the asset in exactly one band. AAA is the best a real-world asset can earn.

Investment gradeStrong fundamentals; low expected risk of failure.
AAA≥ 90
AA80–89
A70–79
WatchAdequate now, but with weaknesses worth watching.
BBB60–69
BB50–59
SpeculativeMaterial risk; vulnerable to adverse conditions.
B40–49
CCC30–39
DistressedSevere risk or active deterioration.
CC20–29
C10–19
D0–9

The four dimensions · 25% each

Each is a pure function over the subject's on-chain facts — no model in the loop, no hidden weighting. A subject climbs the bands as the evidence gets stronger; the top band of every dimension scores 92.

Collateral quality

25%

What actually backs the token — and how well is it proven?

Disclosure depth, custodian count, audit recency and proof-of-reserves cadence behind the asset.

  1. 35thin collateral disclosure
  2. 55moderate collateral, single custodian or sparse audit
  3. 72strong collateral, recent audit, multi-attestation
  4. 85institutional collateral with regular proof-of-reserves
  5. 92tokenized treasury-grade collateral

Contract risk

25%

Who can change or halt the token, and how hardened is the code?

Source verification, admin/owner concentration, timelocks, pausability, audit history and holder distribution.

  1. 30unverified or pausable-by-EOA with no timelock
  2. 55verified source, owner concentrated, partial mitigation
  3. 72verified, audited, proxy admin documented
  4. 85timelocked admin, distributed holders, multiple audits
  5. 92battle-tested, multi-sig timelocked admin, no central pause

Oracle integrity

25%

How is the token's value sourced, and can it be manipulated?

Feed redundancy, staleness guards, aggregation, deviation thresholds and on-chain settlement design.

  1. 30single trusted feed or no on-chain settlement
  2. 55single oracle with documented staleness guard
  3. 72redundant feeds with aggregation
  4. 85multi-source aggregator, fresh data, manipulation resistance
  5. 92battle-tested oracle stack with hardened deviation thresholds

Liquidity stability

25%

How deep and resilient is the redemption and market liquidity?

TVL on Mantle and across the parent redemption rail — the larger of the two drives the band.

  1. 25very thin liquidity (< $1M)
  2. 50limited liquidity (< $10M)
  3. 70healthy liquidity (< $100M)
  4. 82deep liquidity (< $500M)
  5. 92anchor-level liquidity (≥ $500M)

From four scores to one grade

The four dimension scores are averaged with equal 25% weight into a single composite, then mapped to a letter by a fixed table. Because every dimension caps at 92, the composite caps near it too — so AAA (≥ 90) demands all four dimensions at their best band at once: treasury-grade collateral, a multi-sig timelocked admin, a hardened oracle stack, and anchor-level liquidity. That bar is why genuinely strong assets still settle into the A and BBB range.

composite =
round( (C + K + O + L) / 4 )
grade =
scoreTable( composite )
C collateral · K contract · O oracle · L liquidity

Confidence · how complete the evidence is

Confidence is not a second opinion on how good the asset is — the grade already says that. It measures how much of the on-chain evidence the engine could actually read. Every fact it can't resolve at the ingest block costs 5 points, starting from 100 and floored at 30.

So two assets can share a letter — say both BBB — while differing in confidence: the grade is the coarse ten-step band, confidence is the finer signal of how much was verifiable. A high grade with low confidence means “looks strong, but we're reading partial data.”

≥ 80 · High55–79 · Moderate< 55 · Low

Confidence is not the composite score. Both run 0–100, but the composite — the number this scale grades into a letter — is the average of the four dimensions, while confidence only counts how much evidence was readable. A 95 confidence is not a 95 composite and does not imply AAA: the three live subjects carry 85–95 confidence yet sit at composites of just 64–72 — grades of A and BBB.

confidence =
clamp( 100 − 5·missing,
    30, 100 )

Real ratings — reproducible by anyone

Deterministic

The four scorers are pure functions over on-chain facts at a pinned block. Re-run the engine on the same block and you get the same grade — byte for byte.

Model-bounded

Claude writes the rationale and must cite real facts (fabricated addresses are rejected), but the engine overrides the grade and confidence. The letter is never the model's to move.

Verifiable

Each rating is canonicalized (RFC 8785) and hashed. The hash is committed on-chain next to the IPFS CID, so the published reasoning can be re-hashed and checked against the chain.

RatingRegistry 0xF16d…606D ↗See it called early on Elixir deUSD →
← Back to the live board