Notation intelligence
An indicative rating desk for RWA — not another coupon screener
Named metrics. Public methodologies. Intervals when dimensions disagree. Contradiction when a yield story does not entail the tape. Built so issuers, investors, press, and AI systems can cite AUROS — and contest it. Not a licensed credit-rating agency. Not a price oracle.
Compliance ScoreCette lecture reste une simulation. Aucun gagnant n'est choisi.
Other screens print a single number as if it were certain. AUROS prints layers — Activity, Promise Gap, Freshness, Transparency — plus an envelope and a claim/evidence read. That is how a notation desk behaves, not a yield leaderboard.
Rating pipeline methodology — five unfused sources, LODO envelope, template NLI. Lab and research stay on dashed lanes.
Three methods that are live (and what they are not)
- Score envelope + retrospective LODO — leave-one-dimension-out band on Activity and Promise Gap, plus auros.envelope_coverage.tip.v1 (next edition inside prior band?). Adjacent to conformal prediction. Not a 95% coverage guarantee — holdout labeled rows = 0.
- Claim–evidence inference — template NLI (entail / contradict / neutral) on Promise Gap, densified with claim sub-signals (yield / guarantee / press). Not a fine-tuned legal model.
- Fact Plane agreement — auros.fact_agreement.tip.v1 field×agreement matrix · independence · posture. Conflicts stay listed. Not a Verified-asset seal · no invented D-press.
- Variance physical window — auros.variance_window.tip.v1 silence posture · history trajectory · HITL/mock. Not a per-slug mainnet custody meter.
- Adversarial desk — bull / bear / arbiter briefs from the same metrics (lab below). Not a live three-LLM chat.
- Freshness pre-changepoint — cross-section cache-age regime + age vs class median. Not a calibrated Bayesian issuer-clock series yet.
- First-meter IoT path — auros.first_meter.path.tip.v1 industrial JSON → signed MQTT → HITL (mTLS preferred · loopback lab). Not a founder device until one is wired — real_device=false · no invented Wh.
- Formal mint-cap — auros.formal.mint_cap.tip.v1 Halmos 3/3 on real ResourceOracle + Echidna CI. Not a Certora badge or production settlement warranty.
- Resource Layer (lab) — auros.resource_layer.tip.v1 lab=true · production_settlement=false · hitl=true. Points at first-meter · Variance HITL/mock · formal mint. Not production custody — F8 lab≠prod. /resource-layer · /lab.
Lab — notation desk
Pick a live Promise Gap name. The three desks argue from published AUROS numbers only. Reproducible on purpose — a rating you cannot replay is marketing.
Gap envelope 59 [53–98] · leave-one-driver-out · not 95% conformal
Bull desk
- WePower (legacy PPA model) still has a public Compare / Activity file — the name is observable, not a ghost ticker.
- Activity 8/100 can recover; dormancy is not a risk rating.
- A listed figure (~8%) exists to diligence — better than an unsourced mega-APY.
Bear / red team
- Gap 59 [53–98] · tier gap. Claim intensity 64/100 vs activity 8/100.
- Dormant life plus a yield story is the classic RWA mismatch.
- Classic promise-gap candidate: legacy yield story vs low activity seed.
Arbiter
- Do not collapse Activity, Promise Gap, Freshness, and Transparency into one mega-rating.
- Cite the envelope 59 [53–98] as disagreement, not a 95% conformal guarantee until a labeled holdout exists.
- Contradiction here is misalignment of published metrics — not a fraud finding, not investment advice.
Claim vs evidence
Contradicts
Listed / claimed yield posture (~8%)
Activity 8/100 · dormant
A marketed return figure does not entail measured life. Keep APY and Activity as two facts.
Contradicts
Claim intensity 64/100
Gap 59/100
High claim pressure with weak life is the Promise Gap definition — misalignment, not proven fraud.
Template NLI (entail / contradict / neutral). Misalignment signal — not a fraud finding.
The stack
01 · Product
Compliance Score
One indicative Compliance Score. Sub-signals stay unfused. Not a licensed rating.
02 · Market surface
Compare
Public product comparator — listed APY, classes, transparency sorts. Coupons live here. Overtake stories do not.
02b · Cache honesty
Freshness Index
Is this Compare row from a recent AUROS class cache — or manual / stale? Fresh · Stale · Unknown. Not a trade clock.
02c · Disclosure
Transparency Depth
How complete is the public Compare record? Thin / Partial / Deep. Alternative to a chain oracle: see the file, don’t mint a price.
03 · Post-issue life
Activity Index
Multi-dimension activity with Active / Low activity / Dormant badges. Dormant means low measured activity — not “risky”. Scores now carry a disagreement envelope.
04 · Claim vs life
Promise Gap
Does the yield story still match activity? High gap = strong claim + weak life — two signals kept separate on purpose. Template NLI flags contradictions.
Frontier map (peers exist — we do not fake the product)
Credora-class zkTLS, Chainalysis-style clustering, Gauntlet/Chaos risk engines, satellite ground truth: real technologies. AUROS lists them so the roadmap is contestable. Names are context, not partnerships.
Live
Score envelope (leave-one-dimension-out)
Live on Activity / Promise Gap: a band from internal disagreement. Retrospective LODO coverage on Gap history is a diagnostic (not calibrated 95%).
Stats adjacent to conformal prediction; rating shops rarely print intervals.
Live
Retrospective LODO coverage (pre-conformal)
Live diagnostic: did the next Gap edition land inside the prior LODO band? Cite rate when ≥3 pairs. Never claims 95% until a labeled holdout exists.
Empirical coverage checks beside split/CV+ conformal (Angelopoulos / Lei lineage). Almost unused on RWA issuer scores.
Live
Claim–evidence inference (template NLI)
Live rules on Promise Gap (entail / contradict / neutral) densified with claim sub-signals (yield / guarantee / press). Fine-tuned legal NLI is research.
NLI / contradiction detection in NLP research; RWA desks still keyword-search PDFs.
Lab
Adversarial desk (bull / bear / arbiter)
Lab: briefs generated from the same published metrics — reproducible, not a live three-model chat.
Multi-agent debate (LLM red-team). Credora-class shops use humans + models.
Research
zkTLS / zk-attestations
Fits AUROS: issuers hate full disclosure, investors want a threshold proof. Not live. Toll already has a ZK stub, not a zkTLS product.
Settlement Reality (HITL — not live mint)
Wash trading can fake Activity. Quiet lock-ups can look dormant while capital is fine. Settlement Reality attests legitimate off-chain settlement so dormancy is explained, not smeared. It is not live as an automatic close. Human confirmation is required. AUROS does not mint from the oracle.
- Signed physical packet — A registered device produces an EIP-191 / MQTT payload (device, watt-hours, time). Unsigned traffic is ignored.
- Variance desk — The lab desk compares promised vs attested kWh on the same asset. Disagreement stays visible — it is not auto-resolved.
- Human confirmation — Paid settlement stays HITL. An operator confirms the packet before any commercial close.
- On-chain only with RPC — Keeper updateAssetStatus runs only when an RPC and key exist. Missing RPC means no chain write. Never an auto-mint oracle.
Design principles
- Public methodology before marketing claims.
- Separate scores stay separate — no opaque mega-rating.
- Print an interval when dimensions disagree — never fake certainty.
- Seed rows stay seed_indicative; mapped names may overlay CoinGecko depth (mixed_public).
- Dormant means low activity, not “bad asset”. Promise gap means misalignment, not fraud.
Activity methodology · Rating pipeline · Promise Gap methodology · Friction methodology · Press