xln

XLN release execution loop

Owner direction, 2026-10-11: autonomously prepare a usable mainnet candidate, reduce unnecessary code and follow the release scope in AGENTS.md: TypeScript, React and Ethereum same-J, first Sepolia and then a mainnet canary. The current scope and freeze decisions live in owner-profile.md. Production lending remains deferred; the owner lifted the Rust freeze on October 14, with TypeScript still the launch engine. Svelte changes are limited to landing/static copy and build health. October 15 reopens company shares and control for an experimental local demo, not mainnet company admission.

Owner request, 2026-10-11: diagnose the development trajectory, improve the repository and make this loop learn from results. The crisis review records the evidence and priorities. Its governing principle is a financial promise is implemented only when its failure path is executable. Keep the current release scope; sequence its work through complete user journeys.

Outcome and authority

The deliverable is a reproducible candidate with executable financial journeys, recovery evidence and a concrete deployment/rollback package for xln.finance. The first jurisdiction is Ethereum. Use the current Depository and EntityProvider artifacts and require deployed-code identity before public-chain evidence. Older multi-chain plans are historical, not additional gates for this candidate. See the mainnet bar.

Prepare code, local testnet resets, evidence and releases without repeated owner questions. The owner retains seeds and signs mainnet deployment transactions; production starts only after approval of the exact build. This authorization creates no new spending grant or external-model review. The owner’s October 12 request separately authorizes subagent review and a 30-minute method-review heartbeat in the current chat; it does not renew historical monetary budgets. External audit is optional; internal review is not independent audit evidence. Ask only for an unresolved material protocol/authority choice, never routine implementation. Frozen-core approval remains owner-only.

One candidate, one next failure

  1. Freeze a source checkpoint before broad tests. Bind evidence to its SHA or complete source digest, engine, command and artifacts. Keep one live queue in todo.md; use night-work-plan.md for the current recovery handoff. Earlier chat claims are hypotheses until matched to code and artifacts; do not reuse old green results as current acceptance.
  2. Execute the earliest missing production boundary. For each wallet journey, require the visible action, exact committed economic result, and applicable cancellation/rejection/reload behavior. A tutorial chapter opening is not a payment, delivered swap, recovered wallet or activated company board.
  3. Fix the first observed cause. Before adding code, inspect its owner and reuse the existing canonical path. Delete a duplicate only with preserved behavior proven by the same test. Never shrink assertions or financial proof limits to make a gate green. Net lines removed are context, not a quality target.
  4. Rerun that boundary, then its related tests. Broaden only when the candidate works. One machine holds one heavy stand; read stand status and retain the lock for its entire run. Poll the existing process, preserve failing evidence, and copy temporary artifacts before starting the next isolated stand.
  5. When all current gates pass, prepare the exact build, contract bytecode/hash manifest, genesis/network configuration, deployment transactions, operator permissions, backup/restore drill, monitoring and rollback instructions. Verify the package locally before requesting approval of that concrete build.

Reducing the human audit surface

Use the existing XLN-1000 kernel dependency closure and canonical-path analysis; do not introduce another quality score. Report reachable production code, semantic implementations per responsibility, and the actual source delta separately. Moving a file, deleting an oracle or hiding a frozen directory does not establish a smaller financial implementation.

For each deletion, name the surviving owner, all callers, differing authority or input assumptions, and the smallest counterexample that must still pass. One counterexample outranks a clone detector’s match. Keep independent test oracles and separate Runtime/Entity/Account trust boundaries. Wallets consume exact runtime aggregates; bounded detail pages are never financial totals.

Inventory every requested file, then separate mechanical indexing, focused semantic reading and executed evidence. Parallel reviewers own disjoint source sets at one SHA; only one implementer changes each area. Review a stable diff before broad verification. The October 12 review records the first full inventory and concrete proposals; todo.md remains the only live queue. Do not make completing that inventory a new release gate.

Evidence ladder

Treat expensive legal exit as a release blocker even when ordinary payments work. A local Anvil proof establishes local execution; it cannot substitute for current Sepolia bytecode, real public inclusion and the real dispute deadline.

Reflection that changes the next action

Owner direction, 2026-10-11: apply Japanese continuous-improvement practices to the existing loop. These are engineering choices inspired by the Toyota Production System, not a new framework or a claim about all Japanese design.

Use the existing evidence counts, time to a working journey and reopened failures to judge these changes. Do not add ceremonies, dashboards or quality scores merely to make the process look more Japanese. Simplicity is compatible with continuous revision; aesthetic imperfection never excuses incorrect financial state.

Every 30 minutes of active work, spend at most two minutes reviewing: which acceptance gap closed, what actually failed, and how much time was spent without new evidence. After two failures with unchanged evidence, do not repeat the command: preserve its inputs and first divergence, narrow observation or change the hypothesis. Examples: inspect signed state instead of adding sleep; isolate a single restore before rerunning all tests; measure the expensive contract call before refactoring.

Change at most one working-method rule per review and evaluate it over the next two work blocks. Stop parallel edits when they invalidate evidence; one owner per area. Subagents are used only when explicitly requested and review stable diffs. Agreement between agents is not proof. External model calls require explicit current authorization and the existing cumulative budget accounting.

Measure closed/reopened release gaps, time to reproducible artifacts and retained failure evidence. Report done/total for a named stage, not a guessed readiness percentage. If blocked on a real owner decision, state the exact fork and continue independent authorized work. Do not claim background execution after the turn.

For each 30-minute work block, record only the result in the existing handoff: the gap attempted, command/artifact, elapsed time and whether it closed or reopened. Use net closures per engineering hour and time to the first reproducible failure to compare methods over the next two blocks. A smaller denominator, weaker assertion or changed scope is not an improvement. Keep an unsuccessful method change only if it produced new evidence that explains the next action. This improves the development process and retained instructions; it does not claim to modify model weights or become more intelligent automatically.

Reviewing the review

Owner request, 2026-10-15: improve execution and periodically test whether the improvement mechanism itself is useful. Keep two feedback levels only; a third recursive review would consume the work it is supposed to improve.

The reflection budget is at most thirteen minutes per two active hours. The rest belongs to implementation and verification. No historical external-model budget, usage reset, new spending or production authority is renewed by this loop.

Evidence coverage out of 1000

The requested metric is 1000 * passed required checks / frozen required checks, using the existing release criteria for one identified candidate. List PASS, FAIL and UNKNOWN counts alongside it. UNKNOWN means unverified, not defective. Freeze the denominator before collection for the declared engine, interface and jurisdiction; do not increase the score with redundant tests, split trivial checks, removed features or skipped failures. Repeated runs count once. Stale evidence becomes UNKNOWN when its relevant source/configuration changes.

Any failed or unknown mandatory financial, recovery or operational gate blocks release regardless of the number. 1000 means complete evidence for that declared scope, never universal correctness or a measure of MML adoption. A second pass uses the same criteria and retained artifacts; it cannot grade by appearance. Until the candidate and required-check denominator are frozen, report counts for the actual checks run and leave the overall score unassigned.

MML remains the owner ambition: useful unique economic value, not self-transfers, routed hops or test-token throughput. Reliability of a user’s first payment and recoverable ownership comes before claims about adoption or GDP.