xln

XLN-1000

A 0..1000 steering score for the 2026-10-11 release cut (TypeScript engine, React wallet, Ethereum same-J). Run it with bun tools/xln1000/cli.ts (package script xln1000). It steers priorities; it does not release.

Ideology

  1. Value counts only where it can be reclaimed, and only at this SHA.
  2. A financial promise is implemented only when its failure path is executable.
  3. Activity scores zero: transactions, TPS, commits, checks run, documents written.
  4. Positive evidence must come from a clean committed SHA and expires when its covered paths change.
  5. Negative evidence (a reproduced exploit, a red test) counts from anywhere, committed or not.
  6. Unknown is not proof: missing evidence scores 0, and a reviewer may only lower a verdict.
  7. The collector is a script, not an LLM; agreement between agents is not evidence.

Commands

Inputs are separated so a run is reproducible for its inputs:

Input Source Default Deterministic
tree git objects at the SHA (core snapshot cached in the OS temp dir by SHA) always yes
worktree negative evidence files present only in the working tree of --repo always while files unchanged
ledger <repo>/.logs (or --ledger-dir): QA sqlite (APFS-cloned before reading) and .log files off; --ledger yes for a past SHA, see below
ci GitHub Actions build-and-test at the SHA via gh off; --ci networked; a gh failure is an error

Ledger windows end at the commit time, so runs and logs made later never feed an older SHA: F reads logs modified and QA runs created in the 21 days ending at the commit, H reads QA runs from the 30 days ending at it, and J reads clean passing runs whose SHA is the measured commit or an ancestor. The JSON fingerprint holds only these window-bounded counts. The one wall-clock comparison is the integrity check (a row dated more than a day after “now” is corrupt).

Axes

Axis W Formula Anti-gaming
R Reclaimability 200 2·Σ w·d·e over 12 off-chain state types (Σw = 100); d ∈ {0, ½, 1} design verdict cited to contract file:line; e = 1 iff a listed adversarial test passed in CI at the SHA and none failed Frozen type list and weights; a test must assert exact payout against a non-cooperating counterparty; frozen or deferred types count until their code is deleted with owner approval
J Journey reality 150 1.5·Σ w·L/3 over 12 journeys; L1 headless test, L2 browser E2E through the visible UI, L3 public J + fresh profile + real clock React ui/tests specs only; enqueueEntityTxs caps at L1, clock fast-forward caps at L2; strict L2 needs a clean passing run on the SHA or an ancestor with core, ui/src, contracts unchanged since
X Exit economics 120 executable·120·(0.7·cost + 0.3·capacity); cost = log scale on worst exit $ / 1% from $100 (1) to $1M (0); capacity = (log10 worst exits/day at 50% of the reference J − 3)/3 Worst admitted shape plus one counter-dispute, never the typical one; gas and the 5M budget come from committed evidence; prices fixed (10 gwei, $3000/ETH, 60M gas per Ethereum block, block time from jurisdictions.json); an open negative evidence or a phase above the budget sets executable = 0
H Evidence honesty 120 120·(0.3 clean + 0.2 SHA-bound + 0.2 durable + 0.2 CI + 0.1 ledger) Counts artifacts, not claims; clean = QA runs with dirty = 0 and a SHA in the 30 days ending at the commit; durable = evidence/ refs vs /tmp and .logs refs in markdown; a future-dated ledger row sets ledger = 0
K Kernel size and clarity 100 100·(0.35 size + 0.25 long + 0.15 types + 0.25 isolation) on the transitive value-import closure of the kernel roots, code lines only Moving files does not help while they are still imported; long = Σ max(0, len − 40) of outermost functions; types = (casts + non-null) per kLOC
S Single canonical path 90 90·(0.7·mean(1/n per role) + 0.3·(1 − extra helper definitions/400)) Implementations come from a path manifest (≥1k lines counts); frozen or deferred code still counts until deleted; renaming does not help
F Feedback latency 90 90·log scale on T from 5 min (1) to 480 min (0); T = Σ stage minutes / Π p, p Laplace-smoothed Fixed stages (check, full unit, lock-ahb, E2E payment) in the 21 days ending at the commit; skips and todos count as failures; a stage with no timed sample makes F unknown (0), so dropping slow-stage logs never helps
P Process weight 50 50·(0.7 prose + 0.3 scripts); prose = process lines per kernel kLOC, log 20..200; scripts linear 40..300 Archives and JSON count; root-level markdown and any *plan* / *handoff* markdown under docs/ are included automatically
D Public-J reality 40 40·(0.4 HEAD bytecode on Sepolia + 0.4 real-time dispute on a public J + 0.2 mainnet canary) Needs a bytecode-hash match plus transaction hashes
M MML edges 40 40·log(1 + U)/log(1 + 3.65e9); U = USD-days outstanding between unaffiliated entities (secured ×1, unsecured ×0.5) Only unaffiliated pairs count; a hub-routed payment counts once, on the commercial edge; hubs, operators, MMs and test entities are excluded; outstanding |Δ|, never volume or limits

Manifests, weights, anchors and reference prices are owner-only. Any formula change bumps the tool version and re-scores the last 5 history points so deltas stay comparable (current: xln1000/2, which made the ledger opt-in and commit-anchored and scores a stage with no samples as unknown). A second pass on the same SHA and inputs must reproduce the JSON byte for byte.

Relation to the release gate

The binary evidence-coverage gate in improvement-loop.md (passed / frozen required checks; any failed or unknown mandatory gate blocks release) stays authoritative for release, together with mainnet-acceptance-gate.md. XLN-1000 never releases anything: it only orders the work. A high score with a failed required check is still a no-go; a vetoed axis is worked first.

Self-improvement loop

Baseline at f1464e84e

Inputs Total R J X H K S F P D M
tree only (clean clone) 122.6 0 0 45.9 13.7 22.2 29.7 0 11.1 0 0
tree + worktree (main tree, default) 76.7 0 0 0 13.7 22.2 29.7 0 11.1 0 0
+ --ledger 77.8 0 0 0 13.7 22.2 29.7 1.1 11.1 0 0
+ --ledger --ci 183.8 106 0 0 13.7 22.2 29.7 1.1 11.1 0 0

Ledger windows ending 2026-10-10T22:50:40Z: 112 QA runs (30 days) and 214 logs (21 days); every row reproduced byte for byte on a second run.

Not yet measurable (data source TBD)