00 Protocol specification · draft v0.9 · 26 October 2026

Every chain that issues
liabilities needs
a load line.

Ztakora is the collateral sufficiency layer for on-chain liabilities.

It computes the largest liability an issuer may put on-chain right now — from collateral that has actually been attested, stressed by the time it would take to sell. The number is live, it decays when evidence goes stale, and the contract refuses issuance above it.

A specification, not an offering. No token exists, no contract is under audit, and the figures in the worked example are illustrative inputs rather than a live deployment. Everything that is a claim about the outside world is sourced at the end.

A circle bisected by a horizontal band: the international load-line markL · ceilingfreeboard
The load line mark of the international convention: a circle bisected by the deepest safe waterline. Above it, the vessel is still afloat. Below it, it is already sunk.

What is true today

Status
Specification v0.9, reference implementation running on a public testnet.
Audits
None commissioned. Scope published in §12.
Token
None issued. None for sale. See non-goals.
Deployed
Eight components on Base Sepolia (chain 84532), root contract 0x95d445…cf15, live since block 50 982 076. Epochs advance, books publish, verifiers abstain by name when a book is not there. No epoch has yet settled on a quorate attestation on that chain, and the reason is the cost of a transaction there, not a rule of this mechanism. Read it from the chain.

01 At a glance

A ceiling, not a certificate

Tokenised assets have grown far faster than the machinery that proves they are covered. The industry answer is the reserve attestation: a periodic statement that at one instant, assets matched liabilities. Ztakora replaces the statement with a limit. It asks a different question — how much may this issuer put on-chain today, given what has been proven, how good it is, and how long it would take to sell? — answers it every epoch, and puts the answer where it binds: in the contract that mints.

  • 01

    Continuous, not periodic

    The ceiling is recomputed each epoch from rolling evidence. Nothing expires into a grace period where the old number keeps working.

  • 02

    Decay, not revocation

    Stale evidence multiplies the ceiling down by itself. No committee has to notice a problem and vote to act before protection ends.

  • 03

    Stressed by time-to-sell

    Haircuts come from a liquidation horizon and market depth, not a fixed percentage applied to every asset class alike.

  • 04

    Binding at the mint

    The issuer's contract calls the network before it creates. Above the line the call reverts, so an over-issue has to be a deliberate bypass.

The gap the specification is aimed at

0
Value of distributed real-world assets on-chain today, excluding fiat-backed stablecoins, from $389bn including them. [R1]
0
Tokenised assets projected by 2033 — BCG and Ripple, revising their own 2022 estimate upward. Roughly 485× today's non-stablecoin figure. [R2]
none
Open standards in wide deployment that compute a maximum issuable liability from attested collateral: none. Attestation is reporting. Ztakora proposes the limit.

The category split is where the risk concentrates: tokenised Treasuries ≈$15.7bn, credit ≈$8.0bn, commodities ≈$4.9bn as of September 2026 [R3]. The first is nearly a load line already — short-dated, deep, one day to unwind. The other two are where h and q do all the work, and they are the fastest-growing part of the book. Trackers disagree on totals by more than a factor of three depending on whether stablecoins and bridged duplicates are counted; Appendix C cites the range rather than the flattering number.

02 The problem

A snapshot cannot stop a mint

Proof-of-reserves was the right instinct and the wrong instrument. It answers were you covered when we looked, at a date chosen in advance, using a valuation supplied by the party being measured. Three properties make it reporting rather than protection, and all three survive a good auditor.

  1. It is retrospective. The attestation describes an instant. The issuance that breaks it happens in every other instant, and the gap between them is where the liability actually accumulates.
  2. It is binary. Covered or not covered. An issuer at 100.4% coverage and an issuer at 100.01% both pass, and the second one has no reason to slow down before the next snapshot.
  3. It ignores how much a sale costs. A reserve of long-dated private credit and a reserve of T-bills can be identical at face value and incomparable at the moment of a run. Coverage measured without a liquidation horizon measures the wrong thing.

The failure state is not fraud. It is a solvent issuer on the day of the snapshot and illiquid on the day of the redemption queue.

What has to be true instead

For the limit to bind, five properties have to hold simultaneously. Each one has a corresponding mechanism later in this document, and each mechanism is named where it appears.

Required properties and the mechanism that supplies each
Forward-lookingThe limit applies to the next issuance, not the last report.§04 Ceiling
ContinuousRecomputed on a schedule nobody can choose.§05 Epoch
Independently sourcedNo single party supplies both sides of the comparison.§06 Three classes
Time-stressedValue discounted by how hard the asset is to exit.§04.2 Haircut
Self-removingEvidence that ages stops counting without anyone deciding.§05.3 Decay

The honest caveat. A limit computed from attested inputs is only as good as the attestation. Ztakora does not make false statements impossible; it makes them expensive, detectable, and short-lived, and it removes the benefit of the doubt when evidence stops arriving. §11 states this as a risk rather than a solved problem.

03 Foundations

Seven constraints, fixed before the design

Written to make the design falsifiable. A change to any of these is not a parameter update — it is a different protocol, and it should be argued as one.

  1. 1

    Coverage is a rate, never a badge

    The network's only output is a number per issuer per epoch. There is no "verified" state an issuer can put on a website. Anything that can be displayed on a marketing page will be optimised for the marketing page.

  2. 2

    Refusal beats warning

    A warning attached to a mint is a document nobody reads at the moment it matters. The default enforcement point is a revert. Warnings exist, but for operators, not as the protection.

  3. 3

    The ceiling only ever needs one input to fall

    Any single attestation class going silent, stale or adverse lowers the ceiling. It never takes agreement to reduce protection and it always takes agreement to raise it. Asymmetric by construction.

  4. 4

    No emissions

    Security is paid for from the fee an issuer saves by not having to convince a counterparty one at a time. Inflation to buy participation would price the network to grow rather than to be right.

  5. 5

    Parameters move on evidence, never on votes

    A token-weighted vote on a haircut constant is a request for the largest holder to set it. Every parameter change carries a required evidence pack — §10.

  6. 6

    The guarantee is capped or it does not exist

    An uncapped backstop is a promise the network cannot keep. Every guaranteed figure has a hard ceiling above which it says "shortfall", and says it in the contract output.

  7. 7

    Everything is reproducible from public data

    Inputs, weights, and the arithmetic are published each epoch with the result. An analyst who disagrees can recompute the number rather than argue about the report.

03.1 Explicit non-goals

What this protocol refuses to be

Not a credit rating.

The ceiling uses observable market quantities — price, depth, time-to-sell — and legal facts about custody. It does not opine on whether an issuer is well run. An investment-grade issuer with unauditable custody gets a low ceiling and a fair one.

Not a custody provider.

Ztakora never holds, keys, moves or freezes an asset. Custody stays with the issuer's regulated custodian; the network only records whether it proved to exist.

Not an exchange or a venue.

No order flow, no price discovery, no listing decisions. The network has one output and it is a limit.

Not a lender, insurance carrier, or yield product.

The backstop is a capped, loss-bearing reserve for a defined settlement failure, not a product that returns to depositors. Backers can lose their stake; holders can be paid short.

Not a token sale.

Nothing here is an offer of a security, an invitation to invest, or a promise of future network access. A token specification appears in §08 because the mechanism needs one to bond — not because one is being distributed.

Not a substitute for law.

Enforceability of the underlying claim is an input qa, weighted by counsel opinion per jurisdiction. Where the legal chain is broken the number goes toward zero, which is the honest answer, not a workaround for it.

04 The load line

One formula, four inputs, no discretion

This is the whole protocol in a line of arithmetic. Everything after §05 exists to feed it honestly and to make its output bind.

Li=(Σa ∈ AiVa·(1 − ha)·qa)×(1 − ci)
Li is the ceiling: the maximum liability issuer i may have outstanding at the end of this epoch. Summed over every attested asset it holds.
Va

Attested market value

Source
Two independent price feeds at the epoch close, median taken, per asset class rule.
Range
Absolute, base currency of the liability.
Fails as
Understates if the feed is stale — price ageing is treated as a market-value haircut, not a data error.
ha

Liquidation haircut

Source
Computed, §04.2. Not a table a human maintains.
Range
0 ≤ h < 1. Above the cap it is not a haircut, it is ineligibility.
Fails as
Too small in quiet markets, which is exactly when a real one is cheap. §11.2 covers the feedback.
qa

Evidence factor

Source
Product of attestation completeness, proof freshness and legal enforceability — §05.3.
Range
0 ≤ q ≤ 1. Zero means "we cannot count this", not "this is worthless".
Fails as
Decay. Silence lowers the ceiling on its own schedule.
ci

Concentration penalty

Source
Herfindahl of the collateral book, §04.3.
Range
0 ≤ c ≤ cmax, governance-capped at 0.40.
Fails as
A nudge toward diversification that never forbids a book — a single-asset issuer can still operate, just smaller.

04.2 · How the haircut is computed

A haircut should answer one question: if this issuer has to unwind position a starting today, how much of its value does the unwinding itself destroy? That is a function of volatility, of how long the position takes to exit, and of how thick the market is while it is being exited. So the network computes it as:

ha = min ( zp · σa · √Ta  +  λ · ( Pa ⁄ Da0.5 ) ,  hcap )

Haircut symbol glossary
zpStress quantile of the return distribution at confidence p (baseline 99%, 1-day).
σaRealised volatility of the asset over a 126-day window, expressed as a one-day figure — the contract computes zp·σ·√T against a horizon counted in days, so an annualised σ here would silently divide every haircut by √365.
TaLiquidation horizon: days in which ≤ 25% of average daily volume can be absorbed, scaled to the position size.
PaPosition value the issuer would need to realise at once.
DaDepth: two-sided liquidity within 100 bps, sampled at the epoch.
λImpact coefficient, per asset class, fitted from realised execution slippage. Governance-gated, §10.
hcapHard ceiling at 0.60. Beyond that the asset is ineligible rather than heavily discounted.

The square root in the volatility term is the standard √t-time scaling and is deliberately conservative for assets whose returns autocorrelate — private credit, insurance-linked notes, anything without a continuous tape. Where the tape does not exist, σ is inherited from the nearest observable proxy and the inheritance is recorded as its own evidence item, so a reader can see that a haircut borrowed from a proxy rather than measured.

04.3 · The concentration penalty

Diversification is worth something at the moment of a forced sale, so it is priced into the ceiling rather than praised in a policy document. Using the share of each asset class in the covered book:

ci = min ( hmax , γ · (Hi − Htarget)+ )   where   Hi = Σa wa2

H is the Herfindahl index of the collateral book, so 1⁄n for an even book of n classes and 1 for a single-class book. The penalty is zero at or below the target Htarget = 0.30 and rises from there; with γ = 0.35 a book that is entirely one asset class takes c = 0.245, so cmax = 0.40 never binds under the default sensitivity — it is a guardrail against a future gate raising γ above 0.57, not a term that does work today. Stating which parameters are inert is part of publishing them.

A worked book, same face value, different shape
BookCompositionΣ weighted coveredciCeiling
A$100m T-bills only$97.90m0.245$73.9m
B$50m bills + $50m corporate$95.55m0.070$88.9m
C$40m bills + $35m corporate + $25m gold$95.68m0.016$94.2m
Identical $100m of face value, and B and C carry slightly less covered value than A because corporate credit and gold haircut harder than bills. They still support 20% and 27% more issuable liability than A — entirely because the book is spread, which is the whole argument for pricing concentration separately.

The numbers here are inputs chosen to make the shape of the mechanism visible, not a description of any live issuer. The calculator below recomputes them from your own assumptions.

04.4 Interactive

Move the line

Change the collateral book and watch the ceiling, the freeboard and the mint decision recompute from the same arithmetic the contract runs. Nothing here is a saved state; reload and it returns to Book C above.

h = 2.1% · q = 1.00 · deepest, shortest horizon

h = 6.8% · q = 1.00 · 4-day horizon

h = 4.4% · q = 1.00 · storage attested

h = 24.0% · q = 0.55 · no tape, σ inherited

Reserve-proof curve: 2-epoch grace, then ×0.70 per epoch. q falls whether or not anyone notices.

Draft gauge: outstanding liability against the computed ceilingceilingliability
Covered value Σ V(1−h)q
—
Concentration penalty ci
—
Load line Li
—
Freeboard
—
Mint decision
—

Recomputed on every input change from L = Σ V(1−h)q × (1−c).

05 Evidence

Four stages from a claim to a number that binds

An epoch is 24 hours, aligned to host-chain block height rather than wall-clock, so the schedule cannot be chosen by whoever benefits from a different Tuesday.

01 · Declare

The manifest is the only step the issuer performs alone, and it is the step with no authority: a declaration that contradicts the other three classes does not raise the ceiling, it collapses it. The asymmetry is the design.

  • Signed by the issuer's registered key
  • Hash committed before proofs are requested
  • Immutable for the epoch once committed
Stage artefacts
Artefactliability-manifest/<epoch>.json
Required fromIssuer
Failure modeAbsent → ceiling is the prior epoch's, decayed by one
Latency budget≤ 40 blocks

05.3 · Evidence decay

The factor q is not a switch. Each attestation class carries its own grace period, after which its contribution decays geometrically per epoch until it reaches zero. Nothing is revoked, no committee decides, and the issuer cannot stall the decay by disputing it.

Grace and decay by evidence class
ClassGraceDecay/epochEffectively zero at
Price and depth1 epoch×0.50epoch +6
On-chain reserve proof2 epochs×0.70epoch +11
Custodian confirmation6 epochs×0.85epoch +28
Counsel enforceability opinion90 epochs×0.93epoch +140

The ordering is the point: fast-moving quantities must be re-proved quickly and slow-moving legal facts are allowed to age, which means an issuer can keep operating through a weekend and cannot keep operating through a month of silence.

05.4 · Quorum, and what breaks it

A ceiling commits when verifiers holding at least two-thirds of the bonded voting stake agree within tolerance — a maximum relative spread of 1.5% between their independent computations. Below that the epoch does not commit and the prior ceiling decays.

  • Two of three classes required. An issuer cannot satisfy the network with only its own on-chain proofs, however complete.
  • Adverse beats absent. A custodian that affirms the liability but disputes the asset is a hard zero for that asset, not a missing input.
  • Verifier inaction is loss. A bonded verifier that skips an epoch earns nothing and is slashed for the skip; agreeing with a demonstrably false result costs the bond.
  • No emergency authority. The network cannot freeze a token, claw a redemption back, or halt an issuer's contract. It can only stop raising the line.

The last item is a deliberate reduction in capability. A system that can rescue an issuer can also decline to, and the discretion involved would need governance — which §03 rules out as an input to protection.

06 Participants

Five roles, and who cannot hold two at once

I

Issuer

Does
Declares liabilities, integrates the mint hook, pays coverage fees.
Holds
Registration, bonded collateral, no vote over its own ceiling.
Cannot
Supply price, attest custody, or be a verifier.
II

Attester

Does
Confirms the off-chain side: custodian, fund administrator, counsel, or a regulated transfer agent.
Holds
A signed, revocable statement attributed to a named legal entity.
Cannot
Be paid by the issuer in a way that is not disclosed in the epoch pack — the relationship is data, not a secret.
III

Verifier

Does
Runs §04 independently, submits result plus proof, signs the commit.
Holds
Bonded stake; earns the verifier share of fees.
Cannot
See another verifier's submission before signing — commits are blind for the first two-thirds of the epoch.
IV

Backer

Does
Bonds into the capped settlement reserve that pays holders on a proven shortfall.
Holds
Fee share proportional to risk actually absorbed, per tranche.
Cannot
Withdraw while a shortfall is unresolved, and is not promised a return — tranches are first-loss by design.
V

Integrator

Does
Builds the mint hook into a chain, vault, or issuance platform.
Holds
A published interface conformance test and nothing else.
Cannot
Re-label the ceiling. "Ztakora-verified" is not an available state; only the number is.

The role separation exists to make one specific failure expensive: an issuer producing its own evidence. It is not a policy preference, it is why q is a product of independent terms.

07 Architecture

Three moving parts, and where each one lives

Deliberately minimal: the network computes a number and a contract refuses a transaction. Anything more ambitious is a different product.

Edge

Issuer SDK & mint hook

A Solidity and Vyper interface, one Rust binding, one Move binding. requireCeiling(amount) resolves against the committed ceiling for the epoch, atomically with the mint. Fails closed: no reachable ceiling, no issuance.

  • Solidity ≥0.8.24
  • Vyper ≥0.3.10
  • Move 2024 ABI
  • Account abstraction safe
  • ~40k gas per check
Settlement

Ceiling registry

Per-issuer, per-epoch committed Li with the input hash. Read-only for issuers, writeable only through quorum commit. Holds the decay clock, the breach ledger, and the capped reserve's remaining capacity.

  • Deterministic commit
  • Epoch-indexed
  • Single-block finality on host
  • Upgradeable behind a timelock, §10
Consensus

Verifier network

Bonded set, blind submission, aggregate signature over the computed result. Runs §04 locally from the published pack; the network does not trust a verifier's arithmetic, it checks that a sample of them reproduce it.

  • Blind-commit, 2/3 window
  • BLS aggregation
  • Proportional slash on skip
  • Target ≥ 64 verifiers at Phase 3
Evidence

Attestation layer

Three separate transports, deliberately not unified: on-chain Merkle proofs, an out-of-band signed channel for custodians and counsel, and market data with its own provenance. Unifying them would create the single point §06 exists to avoid.

  • Reserve Merkle roots
  • Ed25519 signed manifests
  • Two-feed median pricing
  • Per-epoch public pack
Public

Epoch archive

Every input, weight, computation and result, published and reproducible forever. This is where a disagreement becomes an argument with evidence in it rather than a claim about a black box.

  • Append-only
  • IPFS + origin mirror
  • CSV and JSON per epoch
  • No per-issuer confidential field

07.6 · Failure posture

What happens when each part is unavailable
Verifier network downNo commit. Ceiling holds at the prior epoch and decays. Issuance keeps working until it doesn't.
Market feed staleq for price decays by ×0.5 per epoch. A one-day outage is a meaningful cut, not a halt.
Custodian silentGrace 6 epochs, then decay. Long enough for a holiday, too short for a problem.
Origin archive lostHistoric packs reproduce from host-chain commitments; the archive is a convenience, not the authority.
Registry compromisedIssuance is refused, not authorised, by a wrong low value. A wrong high value is the attack, and it needs a quorum of bonded stake — which is why the bond is sized against the ceiling it can misstate.

07.7 · What the network is not

Ztakora holds no assets, executes no transfers, and settles nothing itself. It is a measurement and refusal service: the smallest amount of trust that still has teeth. Two consequences are worth stating plainly rather than hiding.

First, a holder who never checks the ceiling can still be repaid short — the mechanism protects the issuer's contract, not a counterparty who ignores it. Integration is therefore aimed at issuance platforms, not at end users, and §09.3 sets an adoption gate out of that fact.

Second, the fee is paid by the party whose limit is being enforced. That is normal for compliance infrastructure and it is still a conflict, so the price the issuer pays buys a computation it cannot influence rather than a judgement it can shop for. Where that distinction fails, §10 makes the evidence public.

07.8 · The whole integration surface

Two view functions and one require. The protocol's power is proportionate to how little it can ask of an issuer's contract — anything larger gets bypassed, and §11.7 is the honest consequence.

// SPDX-License-Identifier: MIT
pragma solidity ^0.8.24;

interface IZtakoraRegistry {
    /// @notice The committed load line for `issuer`, valid until the next commit.
    /// @return line   Maximum liability that may be outstanding, in base units.
    /// @return epoch  Host block height at which the value was committed.
    function loadLine(address issuer) external view returns (uint256 line, uint32 epoch);

    /// @notice Freeboard remaining, so a frontend can show headroom without arithmetic.
    function freeboard(address issuer) external view returns (int256 headroom, bool over);
}

abstract contract CeilingGuarded {
    error ABOVE_LOAD_LINE(uint256 requested, uint256 line);
    error STALE_EPOCH(uint32 committed, uint32 required);

    IZtakoraRegistry private immutable _registry;
    uint32 internal _maxEpochAge = 2;          // fail closed, but cache the last line

    /// The ceiling is read atomically with the mint. No oracle round-trip, no
    /// off-chain check an operator can choose to skip under load.
    function _requireCeiling(uint256 outstandingAfter) internal view {
        (uint256 line, uint32 epoch) = _registry.loadLine(address(this));
        if (block.number - uint256(epoch) > _maxEpochAge) revert STALE_EPOCH(epoch, block.number);
        if (outstandingAfter > line) revert ABOVE_LOAD_LINE(outstandingAfter, line);
    }
}
Why each line is shaped the way it is
Read is a view, not a pullThe issuer never asks permission to mint and never receives a signature to verify. A permission call is a dependency; a number in storage is not.
_maxEpochAge = 2An issuer keeps operating through two missed commits, then halts. §11.8's "degrades to no new issuance" is this line, not a policy.
Reverts carry valuesrequested and line are in the error, so an operator debugging a refusal at 3am does not have to reconstruct the epoch from an archive.
No admin functionThere is nothing here to upgrade, pause or override. The registry is the upgradeable component and it is behind the §10 timelock.

07.9 · Why the shape matters more when the minter is not a person

Every argument in §02 assumes a human counterparty who reads something. That assumption is already eroding: issuance platforms are being wired to act on behalf of a treasury, a hedging routine, or an agent that decides to mint at a hour when the risk committee is asleep. Payment and authorisation rails for machine-initiated settlement are the fastest moving part of this market, and they are being built before the limits that should contain them.

A limit is the only control that survives that transition intact. Disclosure depends on someone reading it. An attestation depends on someone remembering it expires. A warning depends on someone choosing to heed it. A revert depends on nothing but the arithmetic, and it is the same revert whether the caller is a treasury desk, a solver bot, or an agent with a payment channel and no supervisor.

Which produces the strongest argument for building this now rather than later, and also the sharpest risk in the specification:

  • Machine-rate issuance breaks the epoch. A 24-hour recompute is coarse against an agent that can mint continuously. Either epochs shorten to block-rate for high-frequency issuers, or the ceiling has to be issued with intra-epoch sub-limit reservations. §15 lists this as unsettled question one, and it is the one most likely to force a redesign.
  • Machine-rate redemption breaks the waterfall. §09 assumes losses arrive slowly enough for the sale-of-collateral step to run before the reserve is reached. A queue that clears in four minutes inverts steps 2 and 3.
  • No badge, for a reason that is now practical. An agent integrates an interface, not a brand. §03.1's refusal to sell a "verified" state is what makes the output machine-readable at all — a badge cannot be required, and a number can.

08 Economics

The fee is the whole subsidy

Security funded by the thing it protects, with no inflation leg. If the network is only worth protecting what already pays for it, that is a finding, not a flaw.

08.1 · Coverage fee

Issuers pay per epoch on outstanding attested liability, quoted annualised so it can be compared with what the same issuer already pays an auditor and a custodian.

Fee by risk tier. Tiers set from horizon and evidence class, not negotiation.
TierTestAnnualisedPer epoch
Tier 1h ≤ 5%, horizon ≤ 1 day, all three classes current8 bps0.022 bps
Tier 2h ≤ 12%, horizon ≤ 5 days16 bps0.044 bps
Tier 3h ≤ 30%, or any inherited σ27 bps0.074 bps
Tier 4h > 30%, capped tranche35 bps0.096 bps
A $500m Tier-2 book pays roughly $800k a year. The same issuer's current annual cost for a periodic attestation, counsel opinion and custodian reporting is not materially lower, which is the honest comparison and the reason the fee is per-epoch rather than per-mint.

No fee per transaction, no fee per holder, no fee taken from a redemption. Charging the mint would tax activity; charging the epoch prices the service actually delivered, which is a continuously valid limit.

08.2 · Distribution and the bond

Where each unit of fee goes
Verifiers55%per correctly-reproduced epoch
Backstop accretion20%builds the capped reserve
Attestation channel15%paid to the attester class, disclosed
Protocol treasury10%upgrades, archive, incident cost

Verifiers and backers must bond. The sizing rule is the one worth arguing about: bonded stake must exceed the aggregate ceiling the verifier set can misstate in a single epoch, by a factor of at least 1.2. If a quorum could wrongly raise a ceiling by $40m, the stake exposed to getting it wrong must be at least $48m — so the profitable attack requires buying more than the damage it does.

08.3 · The token, stated plainly

One token, PLM, exists for two purposes and no others: it is the unit in which verifier and backer bonds are posted, and the instrument in which fee claims are settled. It is not a gas token — gas belongs to the host chain — and it is not a claim on revenue in the way equity is.

  • No emissions. Supply enters only as bond, and leaves escrow only when bond is posted. There is no issuance schedule to bootstrap a network that has no security yet, which is the honest version of the objection in §11.6.
  • No sale. No public sale, private sale, or pre-sale is planned in Phase 0–2.
  • Bond ≠ stake. A bond is slashable against a specific misstatement, and is refundable when the exposure closes. It is a performance deposit, not an investment.

This is the weakest section of the specification and it knows it. A network with no emissions and no sale has no obvious way to acquire its first bonded stake without a counterparty that already believes in the fee. §09.4 states the minimum viable bootstrap; anything more confident than that would be a pitch.

09 The capped guarantee

What happens after a breach that should not have happened

The ceiling is prevention. This section is what happens when prevention is wrong, and it is deliberately smaller than readers expect.

09.1 · Loss waterfall

  1. 1Issuer bondPosted at registration at 3% of the last committed ceiling, floored at $250k. Always the first loss; the issuer is never rescued from its own misstatement.
  2. 2Collateral actually soldBefore any reserve is touched, attested assets are liquidated through the issuer's own redemption channel. The reserve exists for the gap between realising and paying, not for bad assets.
  3. 3Backstop, first-loss trancheCapped at 40% of reserve capacity. Bearers are paid the highest fee share and can be reduced to zero.
  4. 4Backstop, senior trancheCapped at the remaining 60%. Pays a lower fee, absorbs only after first-loss is exhausted.
  5. 5Shortfall ledgerWritten down pro rata, named, and public forever. The protocol says "this much could not be paid" in its output rather than inventing capacity.

Reserve target is 5% of aggregate committed ceiling with a hard maximum of 12%. That number is not a safety estimate — it is the largest multiple of historical stressed loss the fee stream can plausibly rebuild within four quarters. A reserve that cannot be rebuilt after being drained is not a reserve.

09.2 · What the guarantee explicitly excludes

  • Any loss from the market value of the collateral falling, once it is sold.
  • Loss of principal in the tokenised asset itself.
  • Redemption queues, suspension of redemptions, or timing losses.
  • Anything caused by a holder or integrator choosing not to read the ceiling.
  • Shortfall above reserve capacity — the ledger, not the reserve, absorbs it.

The list is short because the guarantee is small. A backstop that covers "whatever happened" would be priced like an insurer, licensed like one, and — at 12% of ceiling — would not be one.

09.3 · Adoption gate, and 09.4 the bootstrap

The mechanism protects an issuer's contract, so its value depends entirely on integrators. Phase 3 does not open until three independent conditions hold, measured rather than asserted:

  • The mint hook is live in production on at least two issuance platforms, not merely available as a package.
  • ≥ 80% of aggregate ceiling across registered issuers has a working fallback: if Ztakora stops, their mint degrades to manual review rather than halting.
  • Median breach-detection-to-refusal latency under one epoch for 60 consecutive epochs.

The second condition matters more than it looks. A ceiling that issuers cannot live with is a ceiling that gets bypassed, so the bootstrap target is a network that is inconvenient to break and easy to degrade — the opposite of the usual launch posture.

10 Governance

Eight conditions, all required, none of them a vote

Every constant in §04 and §05 is a parameter. A parameter change is admissible only when the pack below is complete and public. Quorum ratifies that the evidence is present; it does not ratify the evidence.

  1. Two independent reviews. Separately authored, by parties that do not share a fee stream with the proposer.
  2. Twelve-month backtest. Against published epoch data, showing the effect on ceiling, false-attestation rate and refusal count.
  3. Thirty epochs in shadow. The candidate value runs and publishes alongside the live one without binding.
  4. No adverse simulation. No scenario in the pack where the change raises a ceiling for an issuer whose evidence deteriorated.
  5. Thirty days' notice to affected issuers, individually, not by announcement.
  6. Two of three regional quorums. Verifiers are grouped by jurisdiction and legal system; a single region cannot move a constant alone.
  7. Named reversal path. The change ships behind a timelock with a rehearsed rollback, and the rollback itself is not gated.
  8. Post-mortem obligation in advance. If the change is implicated in a breach within 180 days, the pack and the dissent are republished with a written analysis.

10.1 · What is not governable at all

Hard-coded, outside the gate, and therefore outside anyone's reach:

  • The ceiling may never be raised for the current epoch by any mechanism.
  • hcap ≤ 0.60 and cmax ≤ 0.40.
  • Reserve capacity ≤ 12% of aggregate ceiling.
  • No path from any key to an issuer's assets, ever, including emergency.
  • The epoch pack remains public; no confidentiality mode exists.

10.2 · Why not a vote

Token-weighted voting over a haircut constant hands the pricing of risk to whoever holds the most of the instrument that bears it least. Verifier-weighted voting over the same constant hands it to whoever is paid to compute it, which is closer but still a preference disguised as arithmetic.

So the design removes the decision rather than moving it: the question a parameter change has to answer is empirical, and where the evidence does not answer it, the incumbent value stays. That is a real cost — this protocol will be slower to change than one that votes — and it is taken on purpose, because the thing being changed is a safety limit.

11 Risk register

Eight ways this is wrong, and which one is fatal

Each row is the mitigation class it uses — remove, cap, detect, recover — because "monitoring" is not a mitigation and is not listed anywhere in this document.

Risks to the protocol as specified, not to a hypothetical deployment
#RiskClassTreatmentResidual
11.1Attester collusion. A custodian confirms assets that do not exist.capNo single attester's word above q=0.55 for any asset; two classes needed; false statement is a slashable event for the verifier set that signed it. Detection is disclosure of the fee relationship, not loyalty.High. Unavoidable. Stated in §02 as the honest caveat.
11.2Procyclical haircuts. Quiet markets make h small, which raises ceilings right before liquidity disappears.removeCeiling uses the 126-day maximum of the haircut rather than its current value, and depth is sampled at stress percentiles. A calm tape cannot lower the line's floor.Medium. Costs headroom in good times, which is the point.
11.3Legal non-enforceability. The claim on the asset is not enforceable in the issuer's jurisdiction.detectCounsel opinion is its own evidence item with 90-epoch grace; a withdrawn opinion decays q without waiting for anything else.Medium. Depends on counsel quality the network cannot audit.
11.4Guarantee illusion. Holders read a ceiling as a promise and stop diligencing.removeNo badge, no "verified" state, no marketing artefact (§03.1). The only output is a number with a timestamp, and §09.2 is contractual rather than advisory.High. A framing risk; framing is the weakest tool available.
11.5Regulatory reclassification. The network is judged a ratings agency, custodian, or insurer.capRefusal-only interface, no opinions, no custody, capped guarantee — each is chosen because it narrows what the service can be said to be doing.Unquantified. Jurisdiction-dependent and outside the specification's control.
11.6Cold start. No emissions and no sale means no obvious first bond.recoverPhase 1–2 run attestation-only with a verifier set of known operators under a published, explicit trust assumption — which is written into the epoch pack so nobody mistakes the early network for the final one.High. The most likely reason this project does not exist in three years.
11.7Bypass. An issuer integrates, then routes around the hook.detectRegistration is revocable for future epochs, and the registry publishes the fact that an issuer stopped calling it — which is the only sanction available and is honest about being only that.Medium. Strong with platforms, weak with unilateral issuers.
11.8Single host-chain failure. The registry's host halts or reorganises.capCeiling reads are cached with TTL; a halt freezes issuance at the last committed value rather than authorising issuance at zero.Low. Degrades to no new issuance, which is the safe direction.

Two rows are marked High residual and they are not accidents: attester collusion (11.1) and the guarantee illusion (11.4). If either is unacceptable to you, this protocol does not solve your problem, and §15 says the same thing as a recommendation.

12 Security posture

Nothing is audited. Here is the scope that would have to be.

A protocol that protects issuance cannot claim a security history it does not have. This section is a plan with named gaps, in the order the gaps have to close.

Audit and verification plan, by phase
WorkPhaseStatus
Ceiling registry + mint hookbefore Phase 1Not commissioned
Commit/blind-subset protocolbefore Phase 1Not commissioned
Backstop tranche accountingbefore Phase 4Not commissioned
Formal model of §04 arithmeticPhase 2Spec draft only
Penetration test of attestation channelPhase 2Not commissioned
Bug bountyPhase 1 + 30 daysTerms not drafted
Key ceremony for the registryPhase 1Procedure drafted

No firm is named here on purpose. Naming an auditor before engaging one implies a relationship that does not exist, and a specification that borrows a reputation has already told you how it will treat evidence.

12.1 · The threat model, in one page

  • Adversary with issuer keys. Can mint below the line freely; cannot raise the line without a quorum of bonded stake.
  • Adversary with n/3 of verifier bond. Can stall commits — the ceiling then decays, which harms the issuer it is trying to help. Griefing is available; forgery is not.
  • Adversary who is the custodian. The strongest position, because the network cannot observe a physical asset. Capped by §11.1 rather than solved.
  • Adversary in the price feed. Two independent feeds at the median bound the damage to one epoch, and the 126-day maximum haircut bounds the benefit.
  • Failure of the origin archive. No effect on validity; commitments are on the host chain.

12.2 · What would make this unsafe to deploy

Three findings would stop Phase 1 rather than delay it: a way for an issuer to influence its own q without a counterparty noticing; a commit path that lets a verifier see another's result before signing; or a reserve accounting error that can make capacity look larger than it is. Any of those is a redesign, not a patch, and the gate is written down before the audit begins so the gate cannot be moved afterwards.

12.3 · Regulatory posture, stated as design constraints

Not a credit rating agency.

Output is a computed limit with published inputs, not an opinion with a letter attached. The distinction is load-bearing, which is why §15 forbids a rating scale even as a convenience for integrators.

Not a custodian.

No key, no withdrawal path, no freeze authority. If the network cannot move an asset it cannot be the reason an asset is missing, which is the single most valuable thing this design gives up in exchange for.

Not an insurer.

The backstop is a capped, loss-bearing reserve for one defined event, with the exclusions in §09.2 written into the terms rather than the marketing. Licensing analysis has to precede Phase 4 in every jurisdiction that reads it differently, and no such analysis has been commissioned yet — the entity record in the footer of every page says the same.

Not an offer.

No token exists to sell and none is planned for sale in Phase 0–2, so this document can be read as engineering rather than as a prospectus. Published by Ztakora Pte. Ltd., Singapore; the registration number and the licensing analysis that Phase 4 depends on are both outstanding.

Each constraint above is enforced by an omission in the architecture, not by a policy page. That is deliberate: a compliance posture that depends on staff behaviour is a promise, and one that depends on the absence of a function is a fact a regulator can verify from the source.

13 Implementation roadmap

Six phases, and what each one has to measure before the next

No dates. A phase advances when its gate holds, and the gate is a measurement, not a milestone.

0
current

Specification

This document, reviewed publicly. Registry types and the epoch-pack schema fixed as RFCs.

Gate
≥ 3 independent external reviews published, with every dissent left in the record rather than resolved.
1

Attestation-only testnet

No bonds, no mint hook, no guarantee. The network publishes ceilings for synthetic books so the arithmetic can be attacked while it costs nothing.

Gate
180 consecutive epochs committed; 100% reproducible from the public pack; ≥ 16 verifiers.
2

Shadow mode on live books

Two or three consenting issuers, ceilings computed and published but not binding. The only question: would the line have stopped something real?

Gate
Every refusal the line would have made is reviewed with the issuer; false-refusal rate < 2%; zero unexplained ceiling jumps > 15%.
3

Binding pilot, capped

Mint hook enforced. Aggregate ceiling hard-capped at $250m and the backstop is empty — breaches are recorded, not paid.

Gate
The three §09.3 conditions. Detection-to-refusal under one epoch for 60 epochs.
4

Backstop live, pool capped

Bonds posted, reserve funded to target, tranche accounting exercised by a rehearsed simulated breach before a real one can reach it.

Gate
Two full simulated breaches settled end-to-end, with the shortfall ledger written and reversed. Reserve rebuild time under four quarters at then-current fees.
5

Open registration

Any issuer meeting the evidence classes. The cap lifts only if the reserve target has been met for 12 months.

Gate
All eight §11 residual risks re-reviewed with 12 months of real data, and the review is published whether or not it is favourable.

13.7 · Measurement set — all must hold before expansion

≤ 0.5%

False-attestation rate, measured as signed results contradicted by later evidence.

< 1 epoch

Median breach-detection-to-refusal latency.

> 99.5%

Epoch commit success over a rolling 90 days.

100%

Published packs reproducible by an outside analyst with the reference implementation.

< 2%

False-refusal rate, reviewed jointly with each affected issuer.

≥ 64

Independent bonded verifiers, no single operator above 8%.

14 Worked example

An issuer drifts past its own line, epoch by epoch

Fictional, fully specified, reproducible. Every figure below follows from §04 with the parameters in Appendix A, so a reader can check the story rather than accept it.

Ridgeway Digital Notes issues a note series backed by a collateral book. It begins conservatively, then a corporate mandate pushes it into private credit — the asset class where the network has the least visibility and the most conservative decay curve. Nothing about the issuer's intentions changes. That is why the mechanism does not need them to.

The interesting part is what the protocol never does: it never calls the book insolvent, never freezes a token, and never asks anyone's permission to reduce protection. It just lowers a number until the mint stops.

Assume epoch 0 is a normal Tuesday, the mandate is reasonable, and no one is lying at the start.

14.1 · Ceiling drift across fifteen epochs

Amounts in $m, computed with the Appendix A defaults
EpochBookΣ coveredciLiLiabilityDecision
0$60m bills + $40m corporate96.00.07788.670.0mint ok
435 / 35 / 30 — credit enters the book79.40.01278.572.0mint ok
920 / 30 / 50 — credit is now half the book68.40.02866.574.0over line
10unchanged68.40.02866.574.0refused
13Custodian silent — grace ends, q falls43.60.02842.474.0refused
15Redemption queue; assets sold stressed——0.074.0shortfall
Books are stated as bills / corporate / private credit in $m. Every row is the §04 arithmetic on the Appendix A defaults, which means the table can be recomputed line by line. At epoch 9 the liability already exceeded the ceiling, and the protocol's only response is that no further note can be created. The queue at epoch 15 is the same event either way; what changed is that it arrived at $74m of outstanding liability rather than at $140m.

14.2 · How the loss is allocated

Breach arithmetic on the worked example, $m
Realisable after stressed saleOutstandingGap
58.074.016.0
Loss allocation in waterfall order, $m
Source, in orderAbsorbsRunning uncovered
Issuer bond — 3% of the last committed ceiling, $66.5m at epoch 9, floored at $250k2.014.0
Backstop first-loss tranche8.06.0
Backstop senior tranche, capacity for this issuer6.00.0
Shortfall ledger0.00.0
Holders are made whole on these inputs — and the useful reading is the opposite one: the reserve absorbed $14m of $16m, which is 87.5% of the capacity available to every issuer on the network. One book of this size drains it. §09's cap and §11.6's cold-start problem are the same problem seen from two sides.

15 Decision brief

Approve the baseline, not the launch

Recommended

Fund Phase 1 and Phase 2 — attestation-only plus shadow mode against consenting issuers. Do not fund the backstop, the bond sizing, or any token work until the §13 Phase 2 gate has answered the only question that matters: would a line at this height have stopped something real? If the answer after 60 days of shadow data is no, the protocol should be withdrawn rather than repositioned.

Do not

  • Do not market coverage numbers to holders before Phase 3. §11.4 is the risk this product creates most easily for itself.
  • Do not name an audit firm in collateral material before engagement.
  • Do not raise a bond pool to "get the network started" — the sizing rule in §08.2 is the security model, not a growth constraint.
  • Do not add an opinion, a badge, or a rating letter. Each converts a computation into a regulated product with a different legal shape.

Open questions this draft does not settle

  1. Is a 24-hour epoch the right granularity for assets that trade continuously and settle at T+2? Shadow data will say; the intuition here is no.
  2. Should q for the counsel class decay at all? An enforceability opinion does not become less true with age, but a jurisdiction can change.
  3. Who is the counterparty that values a ceiling enough to be the first verifier set, given no emissions? §11.6 marks this High residual and it is still unanswered.
  4. Does a hard cap on reserve capacity make Phase 4 economically pointless for small books? Possibly, and if so the answer is fewer issuers, not a bigger promise.

Appendix Parameters, terms, sources

Everything a reader needs to recompute §04

A · Parameter table

All constants named in this document, with the gate that moves them
SymbolMeaningDefaultRangeMove via
pStress confidence behind zp0.990.95–0.995§10 gate
λMarket-impact coefficientper classfitted§10 gate
hcapHaircut beyond which an asset is ineligible0.60fixednot governable
γConcentration sensitivity0.350.10–0.60§10 gate
HtargetHerfindahl at or below which c = 00.300.20–0.45§10 gate
cmaxCeiling on the concentration penalty0.40fixednot governable
εVerifier agreement tolerance1.5%0.5–3%§10 gate
—Commit quorum of bonded stake2/3fixednot governable
—Epoch length, host-chain equivalent24 h6–48 h§10 gate
—Reserve target and hard cap, of aggregate ceiling5% / 12%cap fixednot governable
—Bond multiple over the misstateable ceiling1.2×1.0–2.0§10 gate
—Issuer bond, of last committed ceiling3%2–8%§10 gate

B · Glossary

Load line, Li
The maximum liability an issuer may have outstanding at the close of an epoch. The protocol's only binding output.
Ceiling, freeboard
The line, and the distance between it and the liability already outstanding. Negative freeboard is not an error state; it is the state in which no further issuance is possible.
Epoch
The recompute interval, aligned to host-chain block height. Nothing in this protocol is scheduled on wall-clock time.
Epoch pack
The published inputs, weights, intermediate terms and result that let anyone reproduce an epoch independently.
Evidence factor, q
The term by which attested value is discounted for completeness, freshness and enforceability.
Liquidation horizon, Ta
Days over which a position can be unwound while absorbing no more than a quarter of average daily volume.
Adverse, absent
Two different inputs. An attester disputing a fact is adverse and zeroes the term; an attester who has not spoken is absent and decays it.
Shortfall ledger
The public record of what could not be paid. Its existence is why §09 is a bounded statement rather than an open one.

C · Sources, and how much each one is worth

  1. [R1] — primary, high confidence. Live tracker totals read 26 October 2026: $389.32bn of tokenised assets including stablecoins, $289.24bn of stablecoins, and $39.01bn of distributed non-stablecoin real-world assets. app.rwa.xyz
  2. [R2] — primary release, high confidence, and a forecast. BCG with Ripple project $18.9tn of tokenised assets by 2033, revising upward from BCG's own 2022 estimate of $16tn by 2030. ripple.com · BCG 2022 baseline (PDF) — a projection is not a measurement, so §01 uses it only for direction of travel and never as an input.
  3. [R3] — secondary, medium confidence. Category split of tokenised Treasuries ≈$15.65bn, credit ≈$7.96bn, commodities ≈$4.85bn at September 2026, reported by Eco from rwa.xyz exports. eco.com — rounded to two significant digits in §01 because the underlying export has not been re-derived from the primary tracker here.
  4. [R4] — primary, high confidence, and it disagrees with [R2]. The Citi Institute and Citi GPS forecast $5.5tn base / $8tn bull of tokenised assets by 2030 — materially below the BCG–Ripple path for a period that overlaps it, from a bank with real exposure to the market it is forecasting. Citi GPS, Tokenization 2030 (PDF)
  5. [R5] — primary, and the closest precedent to §02's objection. Circle's monthly USDC reserve attestation, produced by a Big Four firm with per-asset detail. circle.com The comparison Ztakora has to win is against the strongest version of the reporting model, not against a weaker one.

Trackers disagree on RWA totals by more than a factor of three, for reasons that are legible rather than suspicious: whether fiat-backed stablecoins count, whether one asset bridged to four chains counts four times, and whether private credit is marked at par. [R1] carries the headline because it is the only figure read from a tracker rather than from a report about one, and [R2] against [R4] is left visible because two forecasts of the same market differing by 3× is itself the argument for measurement over narrative. No figure in §04, §08, §09 or §14 comes from outside this page: those are parameters and worked examples defined here, and saying so is what makes them checkable.

— Live deployment

The specification has bytecode, and it is on a chain

Everything above this line is a design. Everything in this section is a fact about a running system, read from the address book the deployer wrote and from the chain itself. It is a testnet deployment of a protocol with no token, no partner and no audit, and it is presented as exactly that: the eight contracts below are live on Base Sepolia, the roles publish and audit against them every epoch, and the mechanism's central claim — that issuance above the line is refused — has been demonstrated on a rehearsal chain and is waiting on a funded quorum on this one.

The eight components of the live deployment, in dependency order. Runtime code length is what the chain itself reports at eth_getCode.
ComponentAddressWhat it is responsible forCode, bytes
Ztakora0x95d4…cf15The root: the epoch clock, the load line, and the only place a mint is refused.20215
ParameterSet0xe65e…7868Epoch to parameter set, hashed, and changeable only through the controller.5953
BondRegistry0x8fd7…32c8Verifier stake, approvals and slashing — the bond a quorum is measured in.6139
AttestationRegistry0x8508…f037Blind commit, reveal, quorum, settlement: the evidence an epoch is built from.9109
LiabilityRegistry0x332c…0d87Books, series, issuer cover, and the mint and burn that stay inside the line.11552
Backstop0xc3f6…3b2dThe reserve: its cap, its fee, and the rules under which it releases.3119
ShortfallLedger0x1b4e…b225The part of a loss that could not be paid, named and remembered.3345
GovernanceGate0x9da5…79b9Endorsements, observation period, signal and band for every parameter change.6642

What is deployed, in one column

Chain
Base Sepolia · 84532
Creation transaction
0xede5…f512
Deployed at block
50982076
Parameter hash
0x2159b9646ac5ec57e0fbd8d05ad25096e8c38988bdb77f7023d3daa0a0461b16
Epoch
720 blocks · commit 240 · reveal 240 · settle 240
Quorum of bonded stake
666667 ppm
Required verifier bond
37393598585152 wad
Book collateral
45000000000000 wad
Read RPC
https://sepolia.base.org
Address book in the repository
state/base-sepolia-v4.json
These figures read from the chain
2026-10-28, block 51121485

Check it yourself, in five commands

No explorer login, no API key, no trust in this page. An address with code is a contract; the code length is the compiler's; the calls are the roles'. The explorer will tell you the source is unverified — that is a publishing decision, not a missing deployment, and the contracts page says what verification would cost.

# 1. Does a contract live at the root address? ("0x" would answer: nothing here)
curl -s https://sepolia.base.org -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"eth_getCode","params":["0x95d445fad4d54e8e5f0375a75779bf79cdf4cf15","latest"]}'

# 2. Who created it, in which transaction, and is its source published?
curl -s https://base-sepolia.blockscout.com/api/v2/addresses/0x95d445fad4d54e8e5f0375a75779bf79cdf4cf15 \
  | jq '{is_contract, creation_status, creator_address_hash, creation_transaction_hash, is_verified}'

# 3. The calls it has actually received
curl -s https://base-sepolia.blockscout.com/api/v2/addresses/0x95d445fad4d54e8e5f0375a75779bf79cdf4cf15/transactions \
  | jq '.items[:5][] | {hash, block_number, timestamp, status, method}'

# 4. Every component in the book against the code this repository compiles
node state/check-code.mjs state/base-sepolia-v4.json https://sepolia.base.org

# 5. Protocol state, the same way the operator CLI reads it
node node/src/cli.ts status --rpc https://sepolia.base.org --book state/base-sepolia-v4.json --chain-id 84532

One honesty note, because the section would be worth little without it. The root above is the 4th deployment on Base Sepolia, and the 3 before it still have code on the chain: the most recent, 0x8e8825389ab3e5d7a3e151efd2e915fe87d3f84e, created at block 50 958 156, then 0x3687…0f60 and 0xde00…cc5a. None of them are readable by a node built from this tree, because the strings a commitment is hashed over were renamed and those constants are load-bearing. A rename mints new addresses and the old ones stay on the chain, holding nothing. The generations, their domain strings and what is at each address are tabulated on the contracts page.

Changelog

VersionDateChange
v0.92026-10-26First public draft. §04 haircuts made time-scaled rather than class-tabulated; §09 reserve cap reduced from 18% to 12% of aggregate ceiling after the §14 example showed a single book draining 87.5% of it; §09.3 adoption gate added after review that a limit nobody integrates is a report.

Corrections are more useful than endorsements, and there is no published route to send one through: no inbox is staffed behind this site, and an address that bounces is worse than no address. When a reader establishes a mistake, the fix lands in the page it belongs to and in the changelog above. Every substantive change to a constant in Appendix A requires the §10 gate, including changes made because a reader pointed out an error.