The Centralization Fingerprint: Methodology and First-Cycle Findings
Two atomic numbers per asset (Trust Assumptions, Override Points), 21 assets scored in a day, zero LLM-inferred counts. What the data shows about which crypto protocols are layered, which are concentrated, and where the differentiation actually lives.
Why this exists
The trigger was a competitor review. Risklayer publishes per-protocol risk scores using an automated pipeline that includes static analysis plus LLM inference. The methodology page is genuinely thoughtful in places: their PSL (Probability of Significant Loss) framework uses jump-diffusion models, GARCH volatility, and Monte Carlo simulation. Their commitment in writing is to flag staleness and confidence, never to display a cached value as authoritative.
The product UX violates that commitment. Their Aave page shows a 2.5/10 LOW dial paired with a Score History indicating the same protocol scored 10.0 CRITICAL three weeks earlier on the same data, under a different "recipe." Their Top Dependencies for Aave lists Chainlink six different ways. Their Trust Assumptions count for Aave is 1, a number that requires ignoring Aave Governance, the Protocol Guardian, Risk Stewards, Aave Labs, OneSig multisigs, and LayerZero DVN dependencies, each of which TI's own research page documents directly.
We are not faster than Risklayer. We are not better resourced. The question was whether there is a structural gap we can defend: honest, sourced enumeration over LLM-inferred speed. The Centralization Fingerprint is the answer.
What is being measured
Each scored asset gets two atomic counts and one analytical synthesis. The counts are deliberately simple:
Trust Assumptions (TA)
The distinct entities or contracts that must behave honestly for the protocol to function as intended. Governance bodies, multisigs, oracle providers, custodians, foundation-controlled keys, validator sets. A higher count is not automatically worse; mature systems often distribute trust across more parties on purpose. The number reveals architecture, not quality.
Override Points (OP)
The privileged functions that can bypass normal protocol flow. Pause switches, upgrade authorities, parameter changes outside timelock, hard-fork capabilities. A protocol with many override points has more control levers; whether those levers are dangerous depends on who can pull them, how fast, and with what oversight.
The note
One paragraph of analytical synthesis explaining the structural shape: which trust dependency dominates, which override is the sharpest single risk, what the comparable peer protocols look like. The note is where the interpretation lives. The counts are atomic; the note is analytical.
Framework adaptation by asset type. The same two field names apply across DeFi protocols, Layer-1 chains, and infrastructure networks. For DeFi, trust assumptions are governance + admin keys + oracle providers + multisigs. For L1 chains, they are validators + foundations + client teams + custodial parties. For infrastructure (like Chainlink's oracle network), they are node operators + admin keys + data sources. Override Points become hard-fork coordination + parameter changes + emission discretion for L1s; protocol pause + upgrade authority + parameter change for DeFi.
The schema stays constant so a single renderer handles every asset type. The semantics shift in the backing list and the note, which is where the per-asset detail belongs anyway.
The four-quadrant view
The single most useful read of the full dataset is the scatter plot. Each asset is one dot, positioned by its TA count (horizontal) and OP count (vertical). The quadrant lines split the chart at the midpoint of each axis. The four quadrants have different structural meanings.
Reading the quadrants:
- Top-left (concentrated power): few trusted parties holding many override capabilities. Pump.fun (3 trust assumptions, 4 override points) is the textbook case: a single authority keypair, a single upgrade authority, a single withdraw authority. The May 2024 insider exploit drained $1.9M using one of those keys.
- Top-right (layered governance): many trusted parties, many override capabilities. The structural-positive profile of mature DeFi. Aave (7 TA, 7 OP) sits highest in TA in this quadrant because of its explicit three-tier authority (Guardian for emergencies, Risk Stewards for parameters, full Governance for listings). The April 2026 rsETH event exercised all three tiers as designed.
- Bottom-left (minimal infrastructure): few trusted parties, few override capabilities. Tends to be either purpose-built minimal systems or assets where the protocol has not yet been instrumented for the metrics we measure. Tron and Pump.fun both fall here on the TA axis but high on the OP axis, putting them in the concentrated-power quadrant.
- Bottom-right (distributed but constrained): many trusted parties, few override capabilities. The structurally most defensible shape. Bitcoin sits closest to this quadrant in our dataset, with 5 trust dependencies but only 4 override points, none of which can execute without years of social coordination.
The five sharpest findings
Almost every asset scored, regardless of category, has between 4 and 7 override points. The count alone is not the signal. The signal is what those override points are (pause vs upgrade vs treasury vs parameter) and how fast they can execute (Sky's 18-hour governance timelock vs Aerodrome's zero-timelock factory upgrade vs Bitcoin's multi-year coordination requirement). The number tells you the surface area; the backing list tells you the actual risk.
Aave at 7 TA / 7 OP has the highest trust-assumption count of any asset and uses a deliberate three-tier authority structure (Guardian, Risk Stewards, Governance). The April 2026 rsETH event proved the structure works under stress, with each tier acting at its appropriate authority level. Pump.fun at 3 TA / 4 OP is the inverse: three private keys with unconstrained authority over a closed-source program. The May 2024 insider exploit demonstrated empirically that one key is enough to drain the system. Same OP count, opposite structural meaning.
Ethereum scores 6 TA / 4 OP, which looks healthy at first read. The structural risk inside that 6-count is Lido, which holds approximately 28% of staked ETH. TI's research page flags this directly; competitor risk dashboards often treat ETH as monolithically decentralized. The Lido concentration is the most consequential trust dependency on Ethereum today and the Fingerprint surfaces it explicitly in the backing list.
Aerodrome's Administrative Architecture section on TI's research page documents this directly: "Pool factories can be upgraded with no timelock, creating upgrade risk. Factory changes can be executed immediately without governance delay." Five trust assumptions and seven override points sounds moderate. One of those override points can execute in zero seconds. That changes the risk shape entirely.
Three trust assumptions (27 Super Representatives, Tron Foundation, Justin Sun personally) all effectively controllable by one principal. Foundation holds ~34% of supply, founder allocations were 45% of initial supply. The USDT-on-Tron dependency adds Tether as a meta-override that can blacklist addresses unilaterally. TRX is in the same OP=4 cohort as Bitcoin and Ethereum, but the speed and scope of override execution is on the opposite end of the spectrum.
What this is NOT
The Centralization Fingerprint is two integers and a paragraph. It is deliberately not a single composite score that pretends to summarize all risk. It is not stable in five-minute increments; we don't intend to refresh it daily. It is not a substitute for reading the underlying research page, where the full claims, risks, breakers, and analytical context live.
If you compare two protocols using only the TA and OP counts, you will misread several of the cases in this dataset. Sky and Aerodrome both score 5 TA / 7 OP, but Sky's overrides require 18-hour governance approval and Aerodrome's factory upgrades execute instantly. Read the note. Click through to the research page. The numbers anchor; they do not conclude.
What's next
Three follow-ups are flagged in the project's TASKS.md, each conditional on signal that the current surface is being used:
- Timelock Depth as a third atomic count (Sky 18h, Aerodrome 0h, Bitcoin years). Surfacing this turns the row into 3 atomic numbers without ballooning complexity. The Aerodrome no-timelock finding is the strongest argument for adding it.
- Per-asset Centralization Fingerprint sparklines showing how the counts have drifted over time. Today the data is a single snapshot. Drift over multiple quarters would surface protocols whose structural shape is changing without governance announcements.
- Position-level risk pages for common DeFi positions (leveraged stETH loop, GHO mint vs USDC borrow, basis trade). Captures the "Capital at rest vs active usage" gap that protocol-level data misses. Defers Risklayer-style PSL Monte Carlo simulation in favor of TI's manual analysis approach.
Browse the full map
Sort by Trust Assumptions ascending to see the most concentrated profiles. Sort by Override Points descending to see the largest control surfaces. Filter by category to compare within DeFi, Layer-1, or infrastructure.
Open the Centralization Map →Sources and methodology disclosures
Every Trust Assumption and Override Point cited in this report is enumerated from one or more of: the asset's TI research page (which itself cites public protocol docs), defiscan.info reviews, AllCoreDevs / governance forum records, public Foundation disclosures, and incident reports for events referenced (rsETH on Aave, Ventuals vHYPE on Hyperliquid, Pump.fun May 2024 insider exploit). The full backing lists per asset are queryable at /data/asset-risk-scores.json. No counts in this dataset are LLM-inferred; the explicit Rule Zero from the project's CONSTITUTION.md applied to every entry.