Homework 8 — One protocol through three lenses

Due: Tuesday 24 November 2026, 23:59 · folder week-08/hw/ in your repo

Goal

Take one DeFi protocol and analyse it the way you would analyse a fintech before a partnership: as a business (what it sells, what it keeps), through its data (real, with a snapshot), and through its risks (the five-lens canvas). Finish with its MiCA classification and a justification you could defend to a compliance officer.

Pick a protocol (one)

An exchange (Uniswap, Curve, PancakeSwap, a Solana DEX) — quantitative element: price impact or LP-vs-hold on one of its real pools.
A lender (Aave, Compound, Morpho, Spark) — quantitative element: utilisation and interest-rate curve, or bad-debt exposure under a price shock.
A staking or restaking service (Lido, Rocket Pool, EigenLayer, Jito) — quantitative element: staked share, exit-queue depth, or the stETH/ETH discount history.
A stablecoin issuer (Sky/DAI, Ethena, Frax) — quantitative element: collateral composition and peg deviation history.
Your own, if DefiLlama has fees and TVL for it. Message me first if it is not on that list; your capstone protocol is the obvious choice.

Deliverables (in `week-08/hw/`)

protocol.ipynb — runs from a clean runtime. Must contain, in this order:
- data acquisition through the three-tier loader (live → dated snapshot in data/ → labelled synthetic); the snapshot must be committed so I can run it without network;
- business-model table: TVL, annualised fees, annualised revenue, market cap, and the ratios fees/TVL, take rate, revenue/mcap, with a sentence per ratio saying what it means for this business;
- at least one quantitative element specific to the protocol type (see the list above), computed from real data, with a chart whose title states the conclusion;
- a comparison against one peer on the same ratios (a second DefiLlama slug is enough);
- the risk canvas as a DataFrame with the five lenses, one fact and one link per row, and a severity 1–5 with a sentence justifying it;
- a governance or documentation text summarised by llm() (mock allowed) followed by the number-check cell, with your own example of a wrong summary the check would pass.
memo.md (500–700 words) with these headings:
- What it sells and what it keeps (the business-model ratios in one paragraph)
- The number that matters (your quantitative element: what it shows and what would change your view)
- Risk canvas (the five rows, one line each, and the one row you would act on first)
- MiCA classification — category of the protocol’s main token (ART / EMT / other / out of scope), the article or definition that supports it, and the service (if any) that would need a CASP authorisation in the EU
- What the assistant got wrong (at least one: denominator, percent vs decimal, a hallucinated number, a category with no article)
- AI-use disclosure

Constraints

Real data, with the snapshot committed. Synthetic fallback is allowed only as a labelled pipeline test and never as a number in the memo.
Every percentage in the memo has a denominator stated next to it. “Impermanent loss of 5%” without “versus holding” is a mistake.
The MiCA paragraph must cite the reading or the regulation text; a category with no justification scores as absent.
Fees and revenue are DefiLlama’s definitions; say which dataType you used.

Scope

4–6 hours. The canvas and the MiCA paragraph are the part I will read most carefully; the notebook is the evidence for them. If the protocol’s data is thin (no fees endpoint, no governance text), say so and use what exists — a canvas row that says “no public audit found, severity 5” is a valid finding.

Self-check

Restart and run all: no errors, with and without network.
Snapshot CSVs in week-08/hw/data/ and committed.
Ratio table with a one-sentence reading per ratio.
Quantitative element from real data; chart title states the conclusion.
Peer comparison present.
Risk canvas: five rows, fact + link + severity each.
Number-check on the LLM summary, plus your own uncaught-error example.
MiCA category with article/definition; CASP service named or ruled out.
Denominator stated next to every percentage; AI-use disclosure present.

How this feeds the capstone

The protocol analysis is a full chapter of the portfolio: the business-model table becomes the domain section of the decision memo, the canvas is the first draft of the risk and controls canvas (week 9 adds model risk and the AI Act rows), and the MiCA paragraph is the regulation section. If your capstone is on a digital-asset question, do this homework on that protocol and reuse everything.