Algorithmic stablecoins and why they broke

1. TL;DR

Purely algorithmic designs are now a small fraction of the roughly $308 billion total stablecoin market (as of 2026-08, CoinLaw, secondary aggregator), having lost most market share to fiat-backed and delta-neutral models after 2022. An algorithmic stablecoin tries to hold a $1 peg using supply-adjustment rules — minting and burning a companion volatile token — instead of holding off-chain cash or heavy on-chain overcollateralization. The idea traces to Robert Sams’ 2014 “Seigniorage Shares” proposal and was refined across years of Ethereum research-forum threads (PID controllers, MetaCoin, FRAX, USM) trying to make the feedback loop stable. In practice, the largest deployments failed catastrophically under stress: TerraUSD (UST) and its companion token LUNA collapsed from a combined value in the tens of billions to near zero in eight days in May 2022, and Iron Finance’s TITAN suffered an earlier, smaller version of the same death spiral in June 2021 — both because the mechanism that was supposed to restore the peg instead accelerated its collapse once confidence broke.

2. Explain it simply

Analogy

A thermostat keeps a room at 70°F by switching a furnace on when it’s cold and off when it’s warm — a simple feedback loop that works fine for ordinary weather. But if someone leaves a window wide open in a blizzard, the furnace can run at full power and still never catch up: the loop that normally stabilizes the room instead burns through fuel while the temperature keeps falling. Algorithmic stablecoins are thermostats for price instead of temperature, and a large enough shock can overwhelm the loop the same way.

ELI5

An algorithmic stablecoin tries to hold a $1 price using only code, not a pile of real dollars in a bank. When the price dips below $1, the code offers people a reward — often letting them trade the stablecoin for a second, more changeable token — to buy it back up. This works fine when only a few people are worried. But if a lot of people get worried at the same time, that reward mechanism can flood the market with the second token, crashing its price, and once that token is worthless the reward stops working at all — taking the $1 coin down with it.

Step-by-step walkthrough

Scenario, simplified from Terra’s actual UST/LUNA mechanism: UST trades at $1.00 and LUNA at $80. A smart contract always lets you burn $1 worth of LUNA to mint 1 UST, or burn 1 UST to mint $1 worth of LUNA, at the oracle price. Anchor Protocol pays roughly 19.5% APY to UST depositors, subsidized by the protocol.

Before (2022-04-05, LUNA near its peak of $119.18). UST is stable near $1; LUNA supply is roughly 1 billion tokens.
Trigger (2022-05-07). Two large addresses withdraw 375 million UST from Anchor; some of that UST is sold into a large Curve pool, pushing UST slightly below $1.
Intended correction. Arbitrageurs are supposed to burn cheap UST for $1 of newly minted LUNA and sell it, which should push UST back toward $1 — but this expands LUNA’s supply every time it happens.
Feedback loop takes over. As confidence erodes and LUNA’s price falls under selling pressure from redeemers, each further $1 of UST redeemed requires minting proportionally more LUNA (since LUNA is worth less per token). Researchers who studied the event describe the result plainly: “over just three days, the LUNA supply increased from 1 billion to 6 trillion and the LUNA price decreased from $80 to almost zero” (Liu, Makarov & Schoar, NBER Working Paper 31160, 2023).
After. Both UST and LUNA are effectively worthless within about a week; a run that started with a few large withdrawals had become self-reinforcing.

Common misconceptions

Myth: “Algorithmic” means backed by nothing at all. Reality: many designs are hybrids — FRAX, for instance, started fully collateralized and only fractionally reduced its collateral ratio as protocol revenue accrued, rather than relying purely on a companion-token mint/burn.
Myth: UST failed because of a single hack or exploit. Reality: post-mortems describe a bank run triggered by large withdrawals following an unsustainable subsidized yield (Anchor’s ~19.5% APY, whose daily cost reached $6 million by April 2022), not a smart-contract bug.
Myth: A PID controller (a standard control-theory tool for smoothly correcting a system toward a target) automatically makes a stablecoin’s peg stable. Reality: the original ethresear.ch thread proposing PID control for stablecoins was met with direct skepticism — one reply warned that “PID control is liable to be extremely unstable when dealing with something as uncertain as an economic system of this magnitude” (fubuloubu, ethresear.ch, 2020).
Myth: seigniorage-share, mint-and-burn designs are a fringe idea unique to Terra. Reality: the lineage runs from Robert Sams’ 2014 whitepaper through a 2018 ethresear.ch thread on PID gains, MetaCoin’s 2020 governance-minimized proposal, and FRAX’s 2019 fractional-algorithmic design — Terra was one large-scale instance of a much older pattern.
Myth: overcollateralized designs can’t fail the same way. Reality: they fail differently — through bad debt or a collateral-linked de-peg (see /stablecoins/cdp-stablecoins/ and /stablecoins/pegs-and-psm/) — rather than a self-reinforcing reflexive spiral between two tokens.

If you only remember one thing

An algorithmic peg is only as strong as the belief that the mechanism will keep working, and once that belief breaks, the same mint/burn loop that defends the peg in calm markets can instead print a companion token into worthlessness.

3. How it works

Seigniorage Shares: the base pattern

Robert Sams’ original proposal (cited in the ethresear.ch discussion below) pairs a stablecoin with a “shares” token: if the stablecoin trades above peg, new stablecoin supply is minted and paid to shares holders as a dividend, expanding supply to push the price back down; if it trades below peg, shares are sold to buy back and burn stablecoin, contracting supply to push the price back up. The simplest version is a proportional (“P-only”) controller: for a price $p$ and current supply $S$ ,

\Delta S = \alpha \cdot S \cdot (p - 1)

where $\alpha$ is a gain constant. The thread’s original poster found this “overly simplistic”: “if the price of the coin goes down by 1%, this formula would remove 1% of the supply. This seems overly simplistic to me. It’s essentially a P-controller with a P-gain of 1” (k26dr, ethresear.ch, 2018-04-05).

PID control and its limits

A full proportional–integral–derivative (PID) controller adjusts supply using not just the current price error $e(t) = p(t) - 1$ but its accumulated history and rate of change:

u(t) = K_p\, e(t) + K_i \int_0^t e(\tau)\, d\tau + K_d \frac{de(t)}{dt}

In practice, the thread’s implementation used only a heavily damped proportional term — “a 1% movement in price leads to a 0.1% change in supply per cycle (P-gain of 0.1),” cycling roughly five times a day (k26dr, 2018-04-06) — because tuning integral and derivative gains without a real market test environment was judged too risky. Vitalik Buterin proposed an alternative in the same thread: maintaining large standing buy/sell orders at $0.99/$1.01 rather than periodic discrete supply changes, but flagged the core vulnerability of any on-chain market-maker peg defense — arbitrage loss “if the attacker has a more up-to-date view of the price than the blockchain feed” (Buterin, ethresear.ch, 2018-04-06), the same oracle-staleness problem discussed for AMMs in /oracles/oracle-manipulation/.

MetaCoin: governance-minimized PID with a DEX oracle

MetaCoin (2020) proposed replacing MakerDAO’s governance-voted stability fee with an immutable, on-chain PID controller: “the controller is dumb and parameters is not be able to be updated once launched,” trading adaptability for censorship resistance while warning of “the risk of runaway feedback loops” (ameensol, ethresear.ch, 2020-02-09). It used a DEX-basket VWAP oracle instead of a curated committee, which critics pointed out simply relocates trust to “the custodians of those stablecoins” instead of removing it.

FRAX: a fractional-algorithmic hybrid

The original FRAX proposal starts fully collateralized (1 DAI in for 1 FRX out) and only fractionally reduces the required collateral ratio as lending revenue accrues on the reserve: “the difference in the reserve ratio… must be paid in FXS as a fee (which is burned out of circulation) so that value isn’t leaving the system but instead captured by the investment token” (EazyC, ethresear.ch, 2019-09-19). If price falls, the accrued cash flow is used to buy back FRAX and “walk back” the reserve ratio toward full collateralization — an explicit attempt to make the algorithmic component a slow-moving efficiency gain layered on top of collateral, rather than the sole source of stability the way Terra’s mechanism was.

Terra’s mint-burn mechanism

Terra’s UST used no off-chain collateral at all: “UST was not supported by off-chain collateral but by a smart contract that allowed an exchange of one unit of UST to $1 worth of Terra’s native currency, LUNA, and vice versa,” which the NBER researchers characterize as “like infinite maturity convertible debt with a face value of $1 backed by LUNA” (Liu, Makarov & Schoar, 2023). Demand for UST was propped up by Anchor Protocol’s subsidized yield rather than organic use, meaning the peg depended on a subsidy that was, by the researchers’ own account, unsustainable well before the collapse.

4. Worked numeric example

The following is an illustrative, simplified model of the mechanic the NBER researchers describe (not the exact historical order-flow, which involved many participants) — it shows why redeeming UST for LUNA during a LUNA price crash is inherently supply-inflationary:

Round	LUNA price	UST redeemed	LUNA minted (= UST redeemed / LUNA price)	Cumulative LUNA minted
1	$80	800,000,000	10,000,000	10,000,000
2	$8 (10x drop from selling pressure)	800,000,000	100,000,000	110,000,000
3	$0.80 (10x drop again)	800,000,000	1,000,000,000	1,110,000,000
4	$0.008 (10x drop again)	800,000,000	100,000,000,000	~101 billion

Each round redeems the same 800 million UST, but because LUNA’s price is collapsing under sell pressure from the previous round’s newly minted LUNA, the amount that must be minted to honor the same dollar redemption grows by roughly the same factor LUNA’s price falls. This illustrative arithmetic mirrors the actual data: “over just three days, the LUNA supply increased from 1 billion to 6 trillion and the LUNA price decreased from $80 to almost zero” (Liu, Makarov & Schoar, 2023) — a roughly 6,000x supply expansion, consistent with the table’s order of magnitude compounding across more, smaller rounds than shown here.

5. Where it’s used

Ethereum

Seigniorage Shares (Robert Sams, 2014) — the foundational proposal; never deployed at scale as specified but the direct conceptual ancestor of every design below.
MetaCoin (2020) — a governance-minimized proposal combining ETH collateral, a DEX-basket price oracle, and an immutable PID controller; remained a design proposal rather than a deployed protocol in these sources.
FRAX — launched as a fractional-algorithmic hybrid starting near full collateralization; its collateral ratio and algorithmic component have evolved substantially since the 2019 proposal referenced here, including moves toward higher collateralization after Terra’s collapse.
USM — launched October 2021, an immutable two-token (USM debt / FUM equity) design using a “median of three” price oracle (Chainlink plus two Uniswap v3 TWAP feeds) rather than a governance-set peg mechanism (jacob-eliosoff, ethresear.ch, 2021-10-16) — closer to overcollateralized than purely algorithmic.
Iron Finance (TITAN) — a partially collateralized algorithmic stablecoin (IRON) paired with a floating seigniorage token (TITAN); see §6.

Solana

n/a — no Solana-native seigniorage-share or PID-controller-style algorithmic stablecoin has launched at comparable scale. Solana’s best-known undercollateralized failure, Cashio (March 2022), was a collateral-validation exploit in an overcollateralized CDP-style design, not an algorithmic peg-mechanism failure — see /stablecoins/pegs-and-psm/.

6. Risks, attacks, and incidents

TerraUSD (UST) / LUNA collapse, May 2022. Triggered by 375 million UST withdrawn from Anchor Protocol on 2022-05-07 following a governance-approved reduction of Anchor’s ~19.5% subsidized yield; LUNA’s supply grew from roughly 1 billion to 6 trillion tokens while its price fell from about $80 to near zero over three days, and UST never recovered its peg (Liu, Makarov & Schoar, NBER Working Paper 31160, 2023-05). Total value destroyed across UST and LUNA has been estimated in the tens of billions of dollars; the researchers found that “wealthier and more sophisticated investors were the first to run and experienced much smaller losses,” while “poorer and less sophisticated investors not only ran later and had larger losses.” Root cause: an uncollateralized mint-burn peg whose defense mechanism (minting LUNA to redeem UST) itself accelerated LUNA’s price collapse once confidence broke.
Iron Finance (TITAN) collapse, 16–17 June 2021. TITAN, the floating token behind the partially collateralized IRON stablecoin, fell from an all-time high of $65 to near zero within a day, with the protocol’s total value locked collapsing from over $2 billion; Iron Finance’s own post-mortem described it as crypto’s “first large-scale bank run.” As large liquidity providers withdrew and sold TITAN into IRON rather than redeeming IRON directly, TITAN’s price fall reduced the value backing IRON, breaking IRON’s own peg (settling around $0.69) and triggering further TITAN selling — the same reflexive structure as Terra’s collapse, a year earlier and at smaller scale (CoinDesk, Federal Reserve “Runs on Algorithmic Stablecoins,” 2021-06 and 2022-06).
Naive proportional control can oscillate. The original PID thread’s own concern — supply changes overshooting and causing price oscillations — was a design worry from the start (k26dr, 2018-04-06), a related instability mode distinct from the confidence-driven runs above.
Oracle dependence and manipulation. Systems relying on a DEX-basket VWAP (MetaCoin) or a small set of Uniswap TWAP feeds (USM) inherit the manipulation risks discussed in /oracles/oracle-manipulation/; USM’s own developer acknowledged the oracle question “ended up consuming maybe half the total time we spent on USM.”

7. Open problems

Is a purely algorithmic peg ever stable under a large enough shock, or always a confidence game? Sources disagree: the Richmond Fed’s account emphasizes the mechanical breakdown of arbitrage once “the redemption capacity was so binding that the arbitrage mechanism of algorithmic stablecoin was no longer working,” while Harald Uhlig’s NBER paper “A Luna-tic Stablecoin Crash” models the same event primarily as a run driven by loss of confidence — both views appear without a settled resolution.
Tuning feedback-loop gains without a real test environment. k26dr’s original problem — setting PID gains for an economic system with no reliable simulation environment — was never definitively solved in the thread; fubuloubu pointed to formal control-theory concepts (observability, controllability, robust control) as the right tools, but these were not applied in a shipped protocol in these sources.
Governance minimization vs. crisis responsiveness. MetaCoin’s immutable PID parameters and Terra’s initially slow response to Anchor’s unsustainable subsidy illustrate the same tension: hard-coding parameters removes attack surface but also the ability to intervene before an imbalance becomes a run.
Where fractional-algorithmic hybrids sit on the risk spectrum. Whether FRAX-style gradual, revenue-funded fractionalization meaningfully avoids Terra’s failure mode, or merely delays a similar dynamic at smaller scale, is not conclusively answered in these sources.

8. Ethereum vs Solana

Aspect	Ethereum	Solana
Seigniorage/PID-style designs	MetaCoin (proposal), FRAX (fractional-algorithmic), USM (minimalist, market-median-oracle)	n/a
Largest algorithmic collapse in this space	Iron Finance/TITAN (June 2021, ~$2B TVL wiped)	n/a (Cashio was a CDP exploit, not algorithmic)
Cross-chain-relevant collapse	TerraUSD/LUNA (May 2022) — ran on its own Cosmos-SDK chain, not Ethereum, but shaped Ethereum-native design debates and had bridged exposure via Wormhole	n/a
Undercollateralized-stablecoin activity today	Largely wound down or shifted to overcollateralized/delta-neutral models after 2022	None at scale; Solana’s dollar liquidity is fiat-backed and delta-neutral instead

Algorithmic, seigniorage-style stablecoin design has been almost entirely an Ethereum-ecosystem research and deployment story — Terra itself ran on a separate Cosmos-SDK chain rather than Ethereum or Solana, but its collapse (and Iron Finance’s a year earlier, which did run on Ethereum-compatible chains) shaped how every DeFi ecosystem, including Solana’s, subsequently approached stablecoin design: toward fiat-backed and delta-neutral models (see /stablecoins/fiat-backed-and-cctp/ and /stablecoins/delta-neutral/) rather than pure algorithmic seigniorage.

9. Reference doc

The reference

PID control with stablecoins — k26dr, ethresear.ch, 2018-04-05. ethresear.ch/t/1623

Summary of the reference

The thread opens with the author implementing Robert Sams’ Seigniorage Shares whitepaper in Solidity and objecting to its proposed supply-adjustment formula as “overly simplistic” — a pure proportional controller with a gain of 1, meaning a 1% price move removes exactly 1% of supply. The author asks whether full PID control (adding integral and derivative terms) would do better, while worrying that tuning the gains is hard without a good test environment.

Replies quickly narrow the design space. MicahZoltu pushes back that “simple is better,” prompting the author to reduce the gain to 0.1 (a 1% price move produces a 0.1% supply change per ~5-times-daily cycle) rather than pursue full PID. Vitalik Buterin proposes an alternative mechanism entirely: maintaining large standing buy/sell orders at $1.01/$0.99 rather than periodic discrete supply changes, but immediately flags the central risk of that approach — an attacker with a more current price than the on-chain feed can arbitrage the standing orders for free, a concern fubuloubu and Buterin refine into a broader discussion of bounding arbitrage losses via mechanism design. The thread pivots to price-discovery mechanisms: the author settles on a reverse Dutch auction (modeled on the EOS crowdsale) for minting/burning, arguing it needs only infrequent, imprecise price feeds and isolates mis-pricing damage to slow supply-adjustment errors rather than direct arbitrage losses — though Buterin identifies a specific low-volume exploit against this auction design where an attacker profits by bidding low when minting and high when redeeming across two auction cycles. fubuloubu closes with a broader methodological point: that PID control “is liable to be extremely unstable when dealing with something as uncertain as an economic system of this magnitude,” recommending formal control-theory tools (observability, controllability, robust control) and simulation frameworks like cadCAD rather than ad hoc gain-tuning.

Key quotes

“If the price of the coin goes down by 1%, this formula would remove 1% of the supply. This seems overly simplistic to me. It’s essentially a P-controller with a P-gain of 1.” (Original post)

“A 1% movement in price leads to a 0.1% change in supply per cycle (P-gain of 0.1).” (Reply #7, k26dr)

“Why not just have large active sell orders at $1.01 equivalent and buy orders at $0.99 equivalent… There’s probably some mathematical formalism that can get you pretty good bounds on loss from arbitrage if the attacker has a more up-to-date view of the price than the blockchain feed. That’s the main thing I would be worried about.” (Reply #9/#11, vbuterin)

“PID control is liable to be extremely unstable when dealing with something as uncertain as an economic system of this magnitude, but thankfully there are solid areas of research in Control Theory that will be better suited to this type of system.” (Reply #7, fubuloubu)

How to read the original

Background needed: what a proportional/PID controller does in general (see §3’s formula), and the basic seigniorage-shares idea of a stablecoin paired with a floating “shares” or equity token. Skip the deep mechanics of the reverse Dutch auction on a first pass unless mint/burn price-discovery specifically interests you. The hardest part to read correctly is Buterin’s exploit in reply #24: it looks like a minor nitpick about low-volume auctions, but it is actually a general demonstration that any discrete-cycle mint/burn auction with an approximate price feed creates a round-trip arbitrage (mint low, redeem high) whenever price moves between cycles — the same structural vulnerability that, at far larger scale and under panic conditions, is what let Terra’s mint-burn arbitrage become supply-destructive rather than stabilizing (§3–§4 above).

What changed since

MetaCoin (2020) and FRAX (2019) both cite this thread’s PID-vs-simple-proportional debate directly in their own designs; USM (2021) sidestepped the debate by using a market-median price oracle instead of an active supply-control loop.
TerraUSD’s 2022 collapse (§6) is the largest real-world test of exactly the failure mode this thread worried about in the abstract — a discrete mint/burn mechanism between a stable and a floating token, breaking under stress the thread’s participants anticipated only in miniature (a single low-volume auction exploit) rather than at systemic scale.
Post-Terra, regulatory frameworks like the GENIUS Act (see /stablecoins/payments/) have moved toward requiring full reserve backing for payment stablecoins, effectively excluding purely algorithmic designs like the one this thread discusses from that regulatory category.

Secondary references

Announcing MetaCoin (ameensol, ethresear.ch, 2020) — read for a concrete governance-minimized PID design building directly on this thread’s ideas.
DeFi Algorithmic Stablecoin: FRAX (EazyC, ethresear.ch, 2019) — read for the fractional-algorithmic hybrid alternative to pure seigniorage shares.
13 dev takeaways from developing the USM stablecoin (jacob-eliosoff, ethresear.ch, 2021) — read for a shipped, oracle-median-based alternative that avoided an active supply-control loop entirely.

The reference

Anatomy of a Run: The Terra Luna Crash — Jiageng Liu, Igor Makarov, Antoinette Schoar, NBER Working Paper No. 31160, 2023-05 (summarized at corpgov.law.harvard.edu, 2023-05-22). nber.org/papers/w31160

Summary of the reference

Using granular on-chain data, the authors reconstruct the mechanics and timeline of Terra’s collapse. They describe UST’s design as fundamentally uncollateralized: “UST was not supported by off-chain collateral but by a smart contract that allowed an exchange of one unit of UST to $1 worth of Terra’s native currency, LUNA, and vice versa” — economically equivalent, in their framing, to “infinite maturity convertible debt with a face value of $1 backed by LUNA.” They trace demand for UST substantially to Anchor Protocol’s subsidized savings rate: “the Anchor protocol offered a very high yield of 19.5% to UST depositors, which generated significant inflows of deposits and led to a large increase in UST issuance,” a subsidy that had grown unsustainable by early 2022 — “by April 2022, a daily subsidy level reached $6 million, prompting the Terra community to pass a proposal to gradually decrease the 19.5% interest rate… starting on May 1, 2022.”

The paper places the run’s origin at “May 7, 2022, when two large addresses withdrew 375M UST from Anchor,” shortly after the yield-reduction proposal passed — the researchers use this timing to argue the yield cut, not an external shock, catalyzed the run. They document the death spiral’s speed: “over just three days, the LUNA supply increased from 1 billion to 6 trillion and the LUNA price decreased from $80 to almost zero,” noting LUNA had already been under pressure, having declined from an April peak: “following its peak value of $119.18 on April 5, 2022, the value of LUNA experienced a decline in conjunction with a general downturn in the value of cryptocurrencies.” A central empirical contribution is distributional: using wallet-level data, the authors find “wealthier and more sophisticated investors were the first to run and experienced much smaller losses. Poorer and less sophisticated investors not only ran later and had larger losses, but a significant fraction of them attempted to buy into the run” — evidence of an information asymmetry in how the run propagated, not just a uniform collapse.

Key quotes

“UST was not supported by off-chain collateral but by a smart contract that allowed an exchange of one unit of UST to $1 worth of Terra’s native currency, LUNA, and vice versa.” (Mechanism)

“The Anchor protocol offered a very high yield of 19.5% to UST depositors, which generated significant inflows of deposits and led to a large increase in UST issuance.” (Anchor subsidy)

“The first signs of the run appeared on May 7, 2022, when two large addresses withdrew 375M UST from Anchor.” (Timeline)

“Over just three days, the LUNA supply increased from 1 billion to 6 trillion and the LUNA price decreased from $80 to almost zero.” (Death spiral)

“Wealthier and more sophisticated investors were the first to run and experienced much smaller losses. Poorer and less sophisticated investors not only ran later and had larger losses, but a significant fraction of them attempted to buy into the run.” (Distributional findings)

How to read the original

Background needed: the mint-burn mechanism described in §3 above, and a basic sense of what a “bank run” model of instability means (withdrawal decisions depend on beliefs about other withdrawal decisions). Skip the full econometric wallet-clustering methodology on a first pass; the timeline and quotes above capture the mechanism-level story. The hardest part to read correctly is the causal claim that the Anchor yield-reduction proposal, not an external market shock, triggered the run: the paper’s evidence for this is the tight timing between the proposal’s passage and the May 7 withdrawals, which is suggestive but (as with any single historical event) not a controlled experiment — Harald Uhlig’s companion NBER paper “A Luna-tic Stablecoin Crash” offers a formal run model consistent with this story but frames the underlying cause more generally as a coordination failure among holders.

What changed since

The GENIUS Act (signed 2025-07-18, see /stablecoins/payments/) was drafted with Terra’s collapse as a direct backdrop, and effectively excludes UST-style uncollateralized designs from the “payment stablecoin” category it regulates.
Delta-neutral designs like Ethena’s USDe (see /stablecoins/delta-neutral/) emerged partly as an attempt to offer high, sustainable-seeming yield without repeating Anchor’s subsidy-funded model — itself now subject to its own scrutiny over funding-rate sustainability.
Later analyses (e.g., the Richmond Fed’s “Why Stablecoins Fail”) have offered complementary mechanical accounts emphasizing the point at which “the redemption capacity was so binding that the arbitrage mechanism of algorithmic stablecoin was no longer working,” consistent with but distinct in emphasis from this paper’s run-based framing.

Secondary references

Harald Uhlig, “A Luna-tic Stablecoin Crash,” NBER Working Paper — read for a formal run-based model of the same collapse.
Richmond Fed, “Why Stablecoins Fail: An Economist’s Post-Mortem on Terra” (2022-07) — read for a shorter, policy-oriented account emphasizing the arbitrage mechanism’s breakdown.
Federal Reserve, “Runs on Algorithmic Stablecoins: Evidence from Iron, Titan, and Steel” (2022-06) — read for the closest documented precedent, a year before Terra, with the same reflexive-collapse structure.

10. Sources

PID control with stablecoins — k26dr, ethresear.ch — 2018-04-05 — https://ethresear.ch/t/pid-control-with-stablecoins/1623
Announcing MetaCoin—The Governance-Minimized Decentralized Stablecoin — ameensol, ethresear.ch — 2020-02-09 — https://ethresear.ch/t/announcing-metacoin-the-governance-minimized-decentralized-stablecoin/6897
DeFi Algorithmic Stablecoin: FRAX (feedback wanted) — EazyC, ethresear.ch — 2019-09-19 — https://ethresear.ch/t/defi-algorithmic-stablecoin-frax-feedback-wanted/6169
13 dev takeaways from developing the USM stablecoin — jacob-eliosoff, ethresear.ch — 2021-10-16 — https://ethresear.ch/t/13-dev-takeaways-from-developing-the-usm-stablecoin/11020
Anatomy of a Run: The Terra Luna Crash — Jiageng Liu, Igor Makarov, Antoinette Schoar, NBER Working Paper No. 31160 — 2023-05 — https://www.nber.org/papers/w31160
Anatomy of a Run: The Terra Luna Crash (summary) — Harvard Law School Forum on Corporate Governance — 2023-05-22 — https://corpgov.law.harvard.edu/2023/05/22/anatomy-of-a-run-the-terra-luna-crash/
In Token Crash Postmortem, Iron Finance Says It Suffered Crypto’s “First Large-Scale Bank Run” — CoinDesk — 2021-06-17 — https://www.coindesk.com/markets/2021/06/17/in-token-crash-postmortem-iron-finance-says-it-suffered-cryptos-first-large-scale-bank-run
Runs on Algorithmic Stablecoins: Evidence from Iron, Titan, and Steel — Federal Reserve — 2022-06-02 — https://www.federalreserve.gov/econres/notes/feds-notes/runs-on-algorithmic-stablecoins-evidence-from-iron-titan-and-steel-20220602.html