xln

xln crisis review and release priorities

xln has a coherent financial thesis and substantial working machinery. The current constraint is closing a release across its growing scope. My architectural judgment is that implementation breadth has outrun evidence that one fixed product can be used, fail safely, recover and exit. Another rewrite would worsen that gap.

Baseline: f1464e84e916a65146c046b576a976c4359bcd25, 2026-10-11, with pre-existing uncommitted settlement/contract work. Evidence separates new checks, historical records and interpretation. This is an engineering and strategy review, not an exhaustive audit or independent assurance.

Continuation: current AGENTS.md narrows launch to TypeScript, React and Ethereum same-J. Rust and Svelte wallet development are frozen; companies and other jurisdictions are deferred. The owner’s later October 11 instruction includes lending: savers lend to a bank, borrowers borrow from it, and the bank retains its liability to savers after borrower default. Earlier broad-scope conclusions below describe the initial review, not the current release denominator.

The idea and the outcome

J/E/A models existing finance: jurisdictions enforce, entities authorize, and bilateral accounts record obligations. Runtime executes and durably delivers their transitions. Signed accounts let counterparties verify agreement and invoke J when cooperation fails; ordinary operations need not all be published globally.

MML remains accounts supporting 51% of world GDP provable and disputable by 2050. Broad bilateral relationship coverage is how to reach it. Counting accounts would reward empty relationships; counting payment volume would reward turnover. GDP attribution must identify underlying value added once. Until an attribution method is validated, report useful customer activity and enforcement coverage without converting either into a GDP percentage.

Owner clarification: credit limits cap accepted unsecured risk; only the collateral-backed entitlement is secured. Signed unsecured debt is still debt. A claim of 100 with 70 allocated collateral has 30 unsecured exposure. A 50 credit limit permits that exposure; it does not guarantee its repayment. Secured recovery still uses actual evidence, challenge windows, inclusion and finality.

The principle for development is: a financial promise is implemented only when its failure path is executable. This connects MML to today’s work more directly than feature count, a throughput slogan or a model’s approval.

What the last six months show

Reachable non-merge history in the inspected April 11 to October 11 window contains 3,752 commits, including 584 with a refactor prefix and 241 refactor titles mentioning splits, moves, extraction, location, size or facades. These are subject counts, not effort measurements; many refactors establish legitimate ownership.

Period Observed development Implication
April and May Runtime adapter and recovery work; release CI at 31753252e on May 20 A complete release path was already an explicit objective
June and July 432 and 1,335 non-merge commits respectively; July includes 385 labelled refactors; native TRON integration at 5ba0c7c5a Large structural investment needs stable user-flow evidence to retain its value
August 850 non-merge commits; Rust commitment primitives begin at 3d9aa1af2 on August 23, followed by live/parity work A second engine creates an ongoing obligation to prove the same financial semantics
September and October Native-chain, recovery, typed-reject, wallet and gas-bound work; current scope includes two engines, two wallets, three jurisdictions, lending and companies Completion must be judged against this explicit matrix, not one green subset

The tracked documentation inventory contains 147 Markdown files and 82,414 lines, including generated release snapshots and historical reports. Size is not a defect count. The actionable problem was stale authority: the owner profile excluded lending and prioritized Svelte while the current execution loop required lending and equal UI priority. The active night handoff had 3,566 lines of mixed history.

Architectural mistakes and their replacement

Judgment Concrete evidence Better action
Asset admission was confused with custody observation The runtime imported 127 unsolicited assets into the signed proof; the historical expanded-proof fixture needed 12,657,990 gas Custody events update agreed assets but cannot add assets. The original signed one-asset/32-payment proof still exits after those donations in 921,213 conservative gas in the new fixture
Too many implementations became simultaneous completion obligations The initial scope required TS/Rust, React/Svelte and three Js The updated scope now selects TS, React and Ethereum; finish their complete money and recovery paths before reopening deferred features
Structural neatness became easier to demonstrate than release closure Hundreds of refactor-labelled commits coexist with a blocked release preflight Require a measured reduction in failure, duplication or time to evidence before more structural changes; file counts and line limits are constraints, not success metrics
Instructions and evidence have multiple competing homes Conflicting owner-profile scope; dated mainnet.md results presented as current; stale process handles in the active handoff One live queue, a short execution handoff, canonical specifications and explicitly historical evidence
Public claims outran the risk model Landing copy said loss was bounded by collateral plus credit, and a caption implied one signed frame sufficed to reclaim funds Explain secured entitlement, chosen unsecured exposure and actual dispute execution consistently

These are judgments about sequencing and control, not a claim that six months were wasted. Preserve the useful work: deterministic transitions, canonical money math, signed authority, WAL/outbox boundaries, exact replay and typed rejection. They are essential infrastructure for enforceable accounts.

What is still missing from release evidence

  1. Complete bounded exit and funding compatibility. The owner’s asset-list clarification led to a smaller runtime fix: custody cannot enlarge the agreed proof. The collateral-positive contract guard and first-funding settlement detour were removed. First funding, top-ups, offline exit, signed reopening and external withdrawal pass for both sides on the existing contract. React funding/exit, recovery, payment and same-J swap pass seven browser cases. One shape under 5M does not prove every admitted shape or correlated exit capacity.
  2. Lending within the current scope. Named borrower facilities, bilateral covenants, bounded signed late interest and repayment authorization are now implemented; lend.md defines their limits. Local maturity/default and crash recovery pass. The owner’s common-bank-pool requirement remains unimplemented: allocation still ties each loan to a saver position. A named regression rejects a 100-unit draw against two committed 60-unit deposits. These existing green tests do not prove pooled-bank semantics. Current public-chain enforcement and immutable release evidence remain required; neither local tests nor the bank’s promise guarantee unsecured repayment.
  3. One current candidate with full evidence. October 9 replay/native-J results and October 10 tutorial results are useful historical evidence. They do not certify the changed October 11 runtime candidate. The current local contract suite passes 202 cases; final replay matches all 498 frames. The recorded Sepolia deployment fails current-bytecode identity for all five core contracts (block 11889009). Redeploy/verify before real-clock secured, debt and offline stale-dispute drills. Stable-candidate CI, live load and operations remain open.
  4. Proven repeat use and executable liquidity. The pilot has concrete recruitment and operating targets. It labels 15 interviews and five partners as targets, not achieved relationships. This review establishes no customer count. A real recurring desk workflow and reliable quotes are the next adoption evidence; no contact or spending is authorized by this report.

What to stop and what to keep

Keep financial correctness, executable dispute, crash/key recovery and usable funding/payments/same-J swaps on the current TS/React/Ethereum scope. Preserve pre-freeze Rust and Svelte evidence as history; do not spend the release cycle repeating checks that the current scope explicitly excludes.

Stop expanding scope while stabilizing the candidate. Defer optional new chains, new clients, more model-ranking exercises, unmeasured optimization and cosmetic file splitting. Do not defer a required assertion or hide a currently offered financial feature merely to improve a completion percentage.

Release timing and the improvement loop

There is no defensible release date while public-chain enforceability and the executable gate remain unresolved. The local exit boundary is now green. Align the production profile, verify the current Sepolia stack, then freeze a candidate and measure the remaining complete run. The owner excluded the soak gate; do not reintroduce it as a release requirement. A failed money/recovery gate restarts the affected candidate verification after its fix. Only an exact reviewed build and deployment/rollback package goes to the owner for promotion approval and signatures.

Use the existing improvement loop: one gap, one owner, shortest reproduction, smallest canonical fix, same-boundary rerun, then related evidence. Review the method every 30 minutes. After two attempts without new data, change observation or hypothesis. Compare net closed gaps per engineering hour and time to first useful failure over the next two work blocks.

The requested /1000 metric measures evidence coverage for a frozen candidate and denominator. FAIL and UNKNOWN mandatory gates still block release; 1000 never means defect-free. This review does not assign an overall score because that candidate and complete denominator have not been frozen.

The second pass must challenge the retained artifacts under the same criteria. It must not approve this report because its author agrees with it. The active queue remains todo.md, not a new framework or permanent background job.

Changes and verification from this review

Reconciled owner-profile scope, separated current instructions from archived history, clarified MML and the credit/collateral distinction, and corrected risk and inbound-capacity copy in all ten landing languages. The English/Russian browser views show the correction with no captured console errors or warnings. Broader landing claims about unlimited scale, competing protocols and CBDCs remain outside this financial-copy correction and need alignment with the qualified claims in the canonical introduction before public promotion.

Initial copy review: 17/17 admission/preflight tests and 21/21 caption tests; frontend build/Svelte check with 0 errors/warnings. Its contract-artifact drift failure and proposed 22,824-byte Account patch remain historical evidence, not the selected fix.

Autonomous continuation: 73 TS tests / 437 assertions, 50 native Account tests, 232 native Entity tests, and three real-contract funding/exit regressions pass. The new receipt rule reports excluded unadmitted rows and preserves their custody on-chain; it does not promise that an unwilling peer will admit donated assets. A newer local dispute draft also no longer invents native counter-proof authority: the named h474 regression failed before that fix and passes after it.

One fresh production WAL contains 498 frames, 9,586 Entity inputs and 9,470 outbox envelopes. Final TS/Rust W1/W4/W8 replay after the asset fix matches every root and ordered output; the native evidence predates the freeze. This recording used 130 users and is not a valid production TPS benchmark. bun run check passed before subsequent parallel scope/package edits. The active handoff has the current first failure and remaining gates. No push or deployment has occurred in this task.

The actual public-proof operator command subsequently exposed stale factory imports, a missing signed proposer direction and reused EOA nonces. Those paths now pass a real-clock local deployment/funding/dispute/withdrawal regression (14 assertions, 49.42 seconds in the latest run), including a code-identity rejection before signing. React’s hosted build passes; its deployment now selects an exact reviewed SHA rather than whichever commit origin/main reaches while testing. Its regression also rejects untracked source before checkout/build, preserving the existing build and revision. React source now participates in the signed release scope and source-cleanliness gate.

Payload preparation exposed a separate correctness defect: an internal marker bypassed strict consensus projection, changing bytes for long hex values and admitting undefined. The fix passes 26 codec/hash checks, exact TS W1/W4/W8 replay of all 498 frames and seven freshly rebuilt React browser journeys. It leaves the frozen core unchanged; it is not a demonstrated performance optimization.

The 1,000-user live gate remains red: all 20,000 payments deliver, but W8 drains in 30.387 seconds and W4 in 32.184 against the 25-second limit. Instrumented W8 repeats also fail; a W32 trial drains in 29.381 seconds. No valid TPS or isolated capacity improvement is claimed. Worker profiling reproduced the entire immutable WAL and located signature recovery and serialization costs, but provided no evidence for another worker-count trial. Production defaults remain unchanged.

The metrics entrypoint now points to its actual CLI, clearing the unwired-file failure. The latest full check passes 38 source gates plus the frontend and React checks. These are working-tree results, not an immutable release candidate; public-chain verification, production-profile closure, live load and the deployment and recovery package remain open.

The next real-clock rehearsal now covers unsecured shortfalls from both Account directions as well as secured exit: three operator cases, 64 assertions. An 850-unit collateral allocation plus a 150-unit unsecured claim pays 950 immediately from collateral and the counterparty’s 100 free reserves; the remaining 50 is verified in the on-chain debt row. Existing debt and changed bytecode reject before another transaction. These are atomic token amounts and controlled local identities.

The automatic watchtower drill now uses a real 30-second response window. The stale proof would give the counterparty 200 of the protected wallet’s 300 collateral units. The production daemon locks the newer proof 12 seconds before the deadline and finalizes it after expiry; the counterparty receives zero and the protected wallet’s transaction nonce does not change. Both tower cases pass. Public Sepolia receipts and real network conditions remain unverified; local clock realism does not remove that gate. The subsequent full source check passes again.

The latest owner instruction restores lending to launch scope, permits open mainnet participation and authorizes autonomous Sepolia key management/funding/ deployment. A dedicated test signer now lives in macOS Keychain; readback and the canonical deployment dry run pass. Faucet access remains unresolved and its balance is zero. No public transaction was broadcast. The former global funding cap blocker is removed; 16 policy tests retain financial and recovery requirements.

Two lending defects are now reproduced and repaired. A valid overdue repayment previously failed with LENDING_REPAY_LOAN_NOT_ACTIVE. A matured 10,000-unit deposit at 1% returned only 10,000 when there was no borrower. Late repayment now works, and the saver independently receives 10,100 at maturity when the bank has Account payout capacity. Borrower interest belongs to the bank. Repayment or default during/after saver payout cannot change the agreed payout or recreate paid capital. Existing committed earnings remain preserved. Early cancellation does not earn unelapsed deposit interest.

Verification: 46 focused tests / 238 assertions, two freshly rebuilt React lending journeys with zero skipped tests, and 38 source gates plus frontend and React checks. Maturity/default cases are Account/Entity evidence, not a public chain or browser maturity drill. Signed loan/deposit covenants, on-chain term enforcement, exposure approval and restart evidence remain open; this is not a release-candidate or mainnet-readiness claim.

The next two reproduced failures are repaired. Distinct savings rates previously produced identical executable proofs. Savings now live in the signed subcontract tree and the stock transformer settles principal plus earned simple interest, frozen at dispute start. Cooperative payout removes that exact claim. Sixteen Depository cases include both directions and bank insolvency: collateral pays; any shortfall becomes debt. Delayed finalization cannot inflate interest.

Savings interest and later spending could also exceed the signed credit limit. The Account now checks aggregate ordinary debt, savings obligations and reserved work, including projected collateral withdrawals. Limit reductions preserve prior claims, and top-ups remain admissible. The related suite passes 130 tests / 5,291 assertions; both rebuilt React lending journeys pass. Sixteen gas tests include 80 claims with maximal starter arguments (3,520,680 conservative finalization gas) and 18 full-width claims (3,460,520), below the 5M transaction budget.

Only DeltaTransformer runtime bytecode changes: 6,714 to 7,745 bytes; new hash 0x0252d347d8c11dbcda19cba4022c431b26994b39809a135dd1b739ff5f5bfda6. Artifacts and TypeChain are synchronized. The full check still fails its committed-artifact drift gate in this shared, uncommitted tree; the source, frontend and React checks pass. No release or public deployment is claimed.

The next executable failure is the borrower side: loans owing 2,525 and 2,550 produce identical dispute proofs. Savings storage round-trip evidence is not a full lending process/WAL restart. Loan authority, borrower enforcement, recovery, public Sepolia receipts and the live load deadline remain open. The dedicated Sepolia signer still has zero test ETH; the browser CAPTCHA confirmation remains pending. The 05:55 UTC method review retains one reproduced financial boundary at a time and stops blind faucet/worker retries. Migration validation stays at the storage boundary rather than rescanning historical lending receipts for every proof.

2026-10-11 07:11 UTC — signed borrower claims and real wallet boundaries

The borrower proof collision is fixed: both savings and loans use the signed TermLoan program. Thirty-two real Depository settlement cases cover both claim kinds, directions, timing and insolvency, with nine existing cases also passing. The 16 gas tests pass. DeltaTransformer runtime is 7,746 bytes with hash 0x63fba61efba77634fbceb13be1453d24c7129acea5397889b9c529cfc7d74139; all eight other contract bytecodes match HEAD literally. Old unpaid unsigned claims reject during restore rather than acquiring fabricated signatures.

A real browser WAL now replays exactly across TS W1/W4: 111 frames, 382 worker operations each, with roots and ordered outputs checked per frame. The generic HLT exporter could not consume this WAL because ordinary frames lack recorded canonical-state hashes. Canonical journal replay verifies the recorded output hashes directly and compares richer evidence between worker counts. No absent recorded root was fabricated; this is not TPS or full failure-injection evidence.

Wallet shortfall accounting had omitted 50 units of J debt after a 950-unit reserve payout. Complete debt summaries now preserve the 1,000-unit asset while keeping the 50 unsecured and unavailable for payment. A 441-debt regression covers both compact-view truncation dimensions. Automatic Account opening also silently granted 10,000 USDC of credit; it now grants zero. Savings interest asks for explicit extra risk consent. The real loan journey then exposed a full borrower receiving limit; the existing receiving-capacity control is reused. Three of four browser journeys passed before that final UI fix; the rebuilt candidate is running. The adapter regression suite passes 97 tests/647 assertions.

Method review: stop expanding scope until the same production journey passes. Reuse the canonical receive planner, retain full pre-truncation financial aggregates and distinguish source checks from an immutable release. Approved bank facilities, signed late terms, current public Sepolia receipts and live-load drain remain open. Dedicated Sepolia balance is still zero. No public deployment, commit, push or release-candidate claim is made.

2026-10-11 07:29 UTC — current wallet verification and unfunded requests

All four rebuilt wallet journeys pass (53.8s). The remaining payment failure was stale routing data: the recipient had 8 USDC incoming capacity, but the borrower still saw 0.05. User profile heartbeats can be one hour. Signed credit policy changes now publish immediately; normal balance updates retain the heartbeat policy. The authenticated-socket regression verifies both properties. The broad related suite passed 226 tests / 5,953 assertions, and the original 111-frame lending WAL remained exact across TS W1/W4 on this source.

The next production boundary was reproduced: a valid request with no matching pool threw after Account commitment. It now returns a closed liquidity-unavailable outcome, records a signed bank-frame decline, retains the consumed request ID and allocates no funds. Twelve local lending tests pass. The rebuilt browser submits that request, then receives another 100 USDC through the same bank. Both lending journeys pass together (27.0s); related tests pass 182 / 1,185 assertions. The resulting 113-frame WAL replays exactly across TS W1/W4, including roots and ordered events/effects/outbox evidence. Digest: ee4e4943fa52e48b243e9232f2f22f86a94c8817be6412e8f5e82c105c75da85.

The 38 source gates and three outer gates pass. Full bun run check still fails committed-artifact drift. Current-bytecode Sepolia dry-run passes; the dedicated address has zero test ETH and no broadcast occurred. Bank approval/pricing and borrower-facing request status remain incomplete. Exact replay does not prove maturity/default crash handling or public-chain operations. No immutable RC claim.

October 11 lending authority closure

Bank Entity governance now publishes saver/borrower prices and grants named, expiring facilities bound to the borrower Account epoch. Actual disbursement still needs payment capacity and accepted Account risk. Three browser journeys and 129-frame TS W1/W4 replay passed before the next consent correction.

A targeted regression then showed that a bank could create coupon debt without a prior borrower request. The fix binds funding to the exact signed request, maximum rate, deadline and epoch, consumes it once, and preserves that authority through recovery. Local evidence: 44 tests / 518 assertions, 219 related tests, three browser journeys and exact 132-frame TS W1/W4 replay. All source gates pass. This is release work, not an additional scoring or audit campaign.

Two earlier boundary fixes are also retained: close requests remain retryable after bank liquidity recovers, and authenticated Account finality retires opening loan reservations without inventing repayment. A further free Sepolia faucet returned HTTP 500, Faucet balance is too low; the dedicated signer remains unfunded.

The next product pass added bank-officer controls to the existing React wallet: publish prices, approve a borrower facility and revoke future draws through the normal authenticated command lane. Borrowers see committed request status. The current candidate passes 225 tests / 1,652 assertions, three real browser journeys, 130-frame TS W1/W4 equality and all source gates. The full release gate still stops at uncommitted contract artifacts. These are dirty-tree results, not mainnet approval.

October 11 signed late-interest boundary

The missing accrual regression expected 10,150 but returned 10,100 after maturity. Bank-published, borrower-authorized late windows now bind the Account covenant and Solidity proof. Interest stays simple and capped. Savings remain independent. A maximum repayment authorization of 2,575 pays only 2,537 when that is the committed Account quote; the later Entity clock cannot charge extra. Admission reserves the whole possible coupon. Old unsigned fields fail restore instead of gaining consent.

Current evidence: 263 tests / 1,959 assertions; three browser journeys / zero skips; 133 exact TS W1/W4 frames; 40 settlement cases, nine stock cases and 16 gas vectors. The 80-claim stress finalization uses 3,623,080 conservative gas against 5M. Only DeltaTransformer bytecode changes; its 7,867-byte runtime hash is 0x0771c06db45d23cadb711ffc6c78278399132674c8fd7c739aa7838c6b34dff9. All source gates pass. Full check still stops at artifact drift against HEAD.

The browser exposed an omitted strict-ingress field and a fixture startup race; both were reproduced and fixed before recording the passing WAL. The first gas failure was an inadmissible test claim exceeding uint256, not an exit-budget regression; the corrected full-width claim keeps the original 5M assertion. Public Sepolia remains unfunded. Runtime crash/outbox maturity drills, public real-clock exits, valid load and immutable release identity remain open.

October 11 Runtime lending crash recovery

The first real process crash left the bank recording 2,575 repaid while the borrower remained at 12,500 with an unacknowledged Account frame. Forty Runtime cycles did not repair it: local outputs were never saved in the WAL outbox. Retaining them only until RAM enqueue also failed after a capped frame and a second crash. The canonical ordered outbox now retains local outputs until the exact authenticated recipient payload commits. There is no second durable queue.

Four SIGKILLs across five processes now recover origination, controlled-clock default, independent saver close and exact late repayment. The borrower finishes at 9,925, the saver at 10,050; redelivering recovered output bytes causes no second debit. These balances are explicit bank IOUs, not fabricated collateral funding.

The current candidate passes 319 tests / 2,173 assertions, three rebuilt browser journeys and exact 136-frame TS W1/W4 replay. Evidence digest: 3ccb007dc22ca4073df35f8e07005e482a7e7ebe62cfc9e7e178cb389f0781c6. All 38 source and three outer gates pass. Full check still rejects uncommitted contract artifacts. Public real-clock exits and valid live load remain open; the next step is a current 1,000-user baseline before any performance change.

October 11 measured proof construction

The unique-operation ledger confirms 20,000 payments across 1,000 Accounts, with 40,000 unique locks and resolves across their two Account legs. A 1,448-sample W1 replay profile places 26.6% of stacks inside dispute-gas admission. Prefix batching repeatedly ABI-encoded every growing prefix. The canonical builder now tests the whole ordered batch first: when its storage atom and both independent uint256 allowances fit, all prefixes fit. Existing greedy overflow splitting is unchanged.

All 352 related tests / 2,253 assertions and three rebuilt wallet journeys pass. Both the 136-frame lending and 35-frame payment WALs remain exact before/after and across TS W1/W4. All 38 source and three outer gates pass. Full check still rejects uncommitted artifacts. The live drain changes from 30,849ms to 28,911ms, but both fail the 25,000ms requirement. These are N=1 shared-desktop runs, not a stable causal speedup estimate; no TPS claim follows. Continue from the remaining measured cost.

October 11 routing publication and lending freshness

A real Entity chat commit cloned the configuration shell, invalidating the profile cache’s reference comparisons despite unchanged policy. Compare exact policy values and routing edges from the existing canonical descriptor instead. A second regression proves that first collateral on an existing zero row must publish the newly visible token even when credit limits and row count stay equal. No signature verification, financial admission or heartbeat deadline is weakened.

Removing accidental gossip exposed a real React dependency: the bank could commit an active loan after the borrower’s last ACK, while the wallet kept displaying pending funding. Serial bank-view refresh with superseded-response rejection now passes the unchanged browser assertions. All three rebuilt journeys pass with zero skips; 375 related tests and four isolated profile tests pass. A prior grouped profile run aborted in Bun teardown after passing assertions, and remains failed evidence. The fresh 110-frame lending WAL is exact through TS W1/W4, including all R/E/A roots and ordered outputs. All 38 source and three outer checks pass.

The current live run settles 20,000 payments with no pending ACKs in 27,665ms, still above 25,000ms. Hub signature recoveries fall from 19,289 to 52; this does not establish a stable speedup from N=1 shared-desktop measurements. A native hex projection experiment showed no established live benefit and was reverted. The canonical load uses durable Hub WAL/fsync and RAM-only user lanes; it is not evidence for persistence of all 1,000 users. Full check still rejects uncommitted contract artifacts. No public deployment or immutable release candidate exists.

October 11 checkpoint and remaining live boundary

Gas admission now builds the previous proof only when resolution or idempotent asset admission needs a before/after comparison. Every successful transition still checks its entire resulting proof. A new adversarial vector proves an ordinary payment cannot slip through an already oversized proof. Parent signature priming now omits dispute signatures that only isolated Account workers verify; worker verification and canonical financial outputs remain unchanged.

The fixed 35-frame payment and 110-frame lending WALs remain exact before/after through TS W1/W4. Checkpoint coverage expands to 554 tests / 448,400 assertions, three real-clock local funding/exit cases, two offline-watchtower cases and three rebuilt React journeys with zero skips. Two combined test commands hit the 60s budget; isolated real-clock stands pass without shortening dispute windows.

Production-default load still fails: 20,000 settled, zero pending ACKs, 26,077ms against 25,000ms. Diagnostic counters were disabled; financial checks and Hub WAL/fsync stayed enabled. Source payments were acknowledged by 20,094ms, Hub ingress completed at 23,564ms, all delivery at 25,909ms and drain at 26,077ms. The remaining completion tail needs evidence, not a relaxed deadline. N=1 does not establish a stable causal speedup. Public Sepolia remains unfunded.

Repeatedly treating intentional synchronized artifacts as an unresolvable drift was a workflow mistake. A reviewed coherent source/tests/artifacts checkpoint is the correct next step. Frozen Rust and foreign CI edits remain unstaged; a local checkpoint does not grant release readiness or mainnet promotion.

October 11 common bank book and elapsed repayment

The saver matcher was the wrong abstraction. Independent saver liabilities and borrower facilities now replace per-deposit loan allocation; no loan names a saver. The bank’s existing authenticated facility is its exposure policy, including approval above deposited principal. Account risk and payment capacity remain separate. Default cannot replenish synthetic liquidity or change a saver claim.

Cooperative repayment now reuses the existing dispute-clock formula. Half-term repayment no longer charges the entire term coupon. This deletes a financial branch without changing Solidity bytecode. The React balance projection consumes the independent savings quote instead of removed allocation counters.

Evidence: 558 related tests, 87 focused financial cases, three rebuilt React journeys, 102 fresh WAL frames exact through TS W1/W4, and full check. The common book also survived four actual SIGKILLs across five processes. All six product choices are now resolved; demand savings, queued payouts and partial repayment remain implementation work. No mainnet-readiness or public-deployment claim follows.

October 11 partial repayment through the same signed claim

The next cutover implements partial repayment without a separate payment engine. The existing covenant carries remaining principal, rate, accrued interest and fractional carry. Both cooperative repayment and unilateral exit use that clock; interest is paid before principal and never compounds. The original end does not reset. An exact preceding-claim hash prevents a stale partial debit from executing twice, without adding another durable receipt. The bank uses the same settlement function, and principal repayment replenishes the approved facility.

Evidence: 562 related tests; 65 production-contract cases; three React journeys; 108 production WAL frames exact through TS W1/W4; five real SIGKILLs across six processes preserve fractional accrual and outbox redelivery. All 16 gas vectors pass: 80 claims finalize at 3,930,280 / 5M gas. Only DeltaTransformer bytecode changes among nine artifacts; runtime grows from 7,867 to 8,326 bytes. Synchronized ABI/types reject old claim schemas rather than reading them through a fallback. Full source check passes on source checkpoint 5a2025a81: 38 source gates and three outer gates. A rejected non-null assertion was removed by sharing the validated Account, lowering existing debt from 617 to 616; all 108 replay frames remain exact. Demand savings, ordered payouts and public Sepolia deployment remain open; this is not a mainnet-ready release.