Eight DeFi yield-tranching chassis, side by side. Seven untested theories. One that has actually absorbed real losses on real defaults.
A tranche is French for "slice" – in structured finance, a slice of the same loan book carved out by where it sits in the loss waterfall. Junior takes the first dollar of loss and gets paid a premium for it; Senior is protected up to a defined cushion and accepts a thinner, steadier coupon. Tranche Research is a 13-slide research brief on the eight protocols trying to rebuild that machinery on-chain, and on the single question that separates them: when the losses actually arrive, does the waterfall hold?
Seven of the eight have never found out. The eighth, Goldfinch V1, has – three documented defaults and ~$17.9M of writedowns, ~$7.0M of it landing on Junior exactly as designed. That asymmetry is the spine of the deck. Goldfinch has since wound down – GIP-87 passed in June 2026 – which makes it a closed book rather than a running experiment, and the loss data final rather than a snapshot.
Written to answer an insurance-vault design question – which chassis could actually sit under a vault that has to pay claims out of a subordinated capital cushion – and published because the side-by-side didn't exist anywhere else.
→ Read it at trancheprotocol.com
- Eight chassis compared on one matrix – core abstraction, senior and junior instruments, loss waterfall, both yield sources, yield-split mechanism, lockups, underlying assets, live products, and the closest TradFi analogue for each.
- The case against looping – why one risk profile for a whole vault socializes loss, why phantom TVL hides insolvency, and what ranked loss changes. With the two 2025 worked examples: the wstETH:ETH unwind on Aave and Stream Finance's ~$285M cross-protocol blast radius.
- The counterpoint, stated fairly – tranching only delivers ranked loss when the underlying is predictable. Over an actively managed vault, the senior tranche is mostly a leveraged bet on the curator. Procyclical buffers, oracle and contract risk hitting both tranches pari passu, and binary DeFi tails all get their own bullet.
- A worked example, end to end – $100 into Royco, split ~$70/$30, run through five loss states from +8% yield down to −40%, showing where Junior is wiped and Senior starts to bleed.
- The only real proof point – Goldfinch V1's three defaults (Tugende Kenya, Stratos, Lend East) broken out by pool size, total writedown, Backer capital wiped, Senior portion, and what was backstopped off-protocol. Plus the June 2026 GIP-87 wind-down and why it leaves the V1 tranching analysis intact: the waterfall ranked losses as specified, the underwriting behind it is what ended the protocol.
- Five cross-cutting lenses – where the first dollar of loss lands, who sets the yield split, how general the chassis is, liquid vs. locked, and what happens to Senior once Junior is exhausted.
- An August 2026 postscript on Morpho Midnight – why the fixed-rate launch isn't a ninth chassis but is the benchmark every senior coupon now has to clear against.
| # | Chassis | Shape | Status |
|---|---|---|---|
| 01 | Royco Dawn | Two-tranche wrapper over any yield source; dynamic utilization curve; observation-period loss recognition | Live |
| 02 | Infinifi | Stablecoin issuer with a three-tier deposit waterfall (iUSD → siUSD → liUSD) | Live |
| 03 | Strata | The purist generalized two-tranche chassis; guaranteed senior minimum + uncapped upside | Live |
| 04 | Covenant | Yield Coin / Leverage Coin per market; pro-rata senior haircut, no time priority | Live |
| 05 | Knox Finance | Fixed-maturity three-tier (Senior / Spectrum / Junior); the only contractual fixed senior coupon | Live |
| 06 | Goldfinch V1 | Per-loan Senior/Junior over real-world private credit | Tested – and wound down |
| 07 | Lotus Protocol | LLTV-ordered tranches inside one lending market; RWA yield floor under Senior | Pre-launch |
| 08 | Mezzanine | Managed tranched stablecoin yield; mechanics partly undisclosed | Pre-launch |
Lotus and Mezzanine are covered narratively but held out of the matrix – neither is live with public on-chain data.
index.html ← the entire deck: 13 <section class="slide"> blocks,
│ inline <style>, no framework, no JS build step
├─ fonts → Fraunces + Inter, Google Fonts (only external request)
├─ favicon.* → .ico / .svg / .png / apple-touch-icon
└─ og.png → 2000×1050 social card, absolute-URL'd in the meta tags
↓
wrangler.jsonc ← static-asset Worker, name "tranche-research", assets dir "."
↓
git push origin main
↓
Cloudflare auto-build → https://trancheprotocol.com/
https://tranche-research.blankm.workers.dev/
One file, no dependencies, no toolchain. Open index.html in a browser and you have the whole thing. That is deliberate: the deck should still render in ten years without a lockfile resolving.
| Path | Purpose |
|---|---|
index.html |
The deck. All 13 slides, all styling, ~75 KB |
og.png |
Social preview card (2000×1050) |
favicon.ico · favicon.svg · favicon.png · apple-touch-icon.png |
Icon set |
wrangler.jsonc |
Cloudflare static-asset Worker config |
LICENSE |
Boost Software License 1.0 |
git clone https://github.com/mishablank/tranche-research.git
cd tranche-research
python3 -m http.server 8000 # then open http://localhost:8000Or, to preview exactly as Cloudflare serves it:
npx wrangler devThe repo is git-connected to Cloudflare, so any push to main redeploys automatically – no manual step, no CLI. Pushes typically go live within a minute.
If you fork it and want your own deployment: change name in wrangler.jsonc, then either connect the repo in the Cloudflare dashboard (Workers & Pages → Create → Connect to Git, framework preset None, build command blank, output dir /) or run npx wrangler deploy. Keep the config name and the Worker name in sync – a mismatch silently creates a second Worker instead of updating the first.
Editing the deck means editing index.html directly. Slides are numbered in two places – the data-slide="NN / 13" attribute on each .top-bar and the footer – so adding a slide means renumbering both.
Every figure is drawn from protocol documentation, public defaults write-ups, or DefiLlama, and is dated in place rather than presented as evergreen. Where a number moves – Infinifi's tier APYs, Midnight's TVL – the deck states the capture date next to it. Knox deliberately has no platform-wide APY because its parameters are set per pool at deployment, and the deck says so instead of inventing an average.
The Goldfinch default table is reconstructed from public documentation and the figures are approximate; it is the one dataset in the deck describing losses that actually happened, so it is flagged as approximate rather than quietly rounded.
The Goldfinch wind-down figures – GIP-87's 1,052,820 GFI cast against a 250,000 quorum, ~$56M outstanding across eight borrowers, ~$1.63M remaining TVL, the ~30% depositor recovery and Warbler Labs' $7M – are from The Defiant's coverage and the DefiLlama numbers it cites, as of June 2026.
- Mezzanine remains partly undisclosed – tranche token names, exact waterfall ordering, lockup terms and live TVL were all unpublished as of the last revision.
- Lotus is still pre-launch, so its loss-containment claims are documentation, not observed behaviour.
Boost Software License 1.0 © 2026 Mikhail Blank
The analysis is research, not investment advice. Figures are approximate and dated; verify anything load-bearing against the protocol's own documentation before acting on it.