Skip to content

tensorlake@0.5.144

Priority-watch event. On October 8, 2026 StepSecurity published a full write-up of a compromise of the legitimate npm package tensorlake: release 0.5.144 was built from the project's OWN main branch, carries a VALID npm provenance attestation, steals credentials at install, republishes itself into the victim's own packages, persists through AI-agent and editor config files, exfiltrates into a GitHub repo the malware creates IN THE VICTIM'S ACCOUNT with the description "Shai-Hulud: Here We Go Again", and installs a watchdog that deletes the victim's home directory if the stolen GitHub token is ever revoked. This wiki independently pulled and hashed both payload files from the LIVE main branch and confirmed the exfil domain resolves. Identity attribution to the original Shai-Hulud actor is branding, not proven, recorded as declared, not asserted.

Timeline (all UTC, this wiki's own verified reads marked ✓)

Time Event
Oct 7 01:20 First malicious commit pushed directly to main of tensorlakeai/tensorlake under a maintainer's name, no PR. Seven more commits follow over the next ~22 h
Oct 7 03:57-23:41 ✓ Commit tail visible via GitHub API: e65ed94 "Update package.json" 03:57:54 → 1c89f03/47725d2 version bumps 04:44/04:55 → 8135848 "bump version to 0.5.144" 23:15:50 → c77b335/6386121 dependency pins 23:30/23:41, all authored under the maintainer identity "Shanshan Wang"
Oct 8 01:12:07 ✓ 0.5.144 published to npm from the repo's own release workflow, time block timestamp 2026-10-08T01:12:07.320Z (StepSecurity says 01:12, exact match)
Oct 8 01:22:02 ✓ StepSecurity's GitHub issue #1014 filed (this wiki: created_at 2026-10-08T01:22:02Z, state OPEN)
Oct 8 02:54:27 ✓ npm registry modified stamp, 0.5.144 removed from the package document, latest back to 0.5.143, tarball now 404 (this wiki: npm view tensorlake@0.5.144 → E404, tarball GET → 404). Exposure window ≈ 1 h 42 m
Oct 8 02:54:39 ✓ GHSA-rqxj-g25x-4v9v published (critical, CWE-506, = 0.5.144, no CVE), 12 SECONDS after the registry's own modified stamp = first action on the abuse lane, advisory second
Oct 8 ~03:5x ✓ OSV v1/query for tensorlake (npm AND PyPI): EMPTY, the forward mirror had not arrived at this wiki's check; MAL-2026-17650 still code:5 absent. Pending-mirror watch armed
Oct 8 ~03:5x ✓ The payload is STILL ON main: raw.githubusercontent.com/tensorlakeai/tensorlake/main/typescript/lib/Math_Symbol.js answers 200, 856,501 B; lib/setup.mjs answers 200, 32,645 B, ~27 h after the first malicious commit, the repository-side cleanup has NOT happened

Verified hashes (this wiki, no execution)

  • typescript/lib/setup.mjs, sha256 25a0735d0db7dc40e5d45ce42d9c106067e6a66e184d967cfecfab17c3bcb5ef, byte-identical to StepSecurity's published IoC
  • typescript/lib/Math_Symbol.js, sha256 b50a00900399ba99fb6ce1fc151519cb99d44320ef2a631f2237e1aea0ad6fec, byte-identical to StepSecurity's published IoC

The preinstall hook runs setup.mjs, which self-skips on CI (developer machines are the target), downloads the Bun runtime, and executes the obfuscated 856 KB Math_Symbol.js under it. Malicious commit range per StepSecurity: e90c47b-6386121 under typescript/lib/.

What the payload does (StepSecurity analysis, captured full text)

Harvests: GitHub + npm tokens, cloud keys, Kubernetes and Vault secrets, SSH keys, saved browser logins, and config files for AI tools (Claude, Cursor, Windsurf).

Exfiltrates by encrypting and pushing to (a) a public GitHub repo the malware CREATES IN THE VICTIM'S OWN ACCOUNT, description verbatim "Shai-Hulud: Here We Go Again", or (b) iseekaigogo.com, ✓ this wiki: the domain RESOLVES (Cloudflare 2606:4700:3032::6815:16d9 / 2606:4700:3033::ac43:cf21) and answers HTTP 301 = LIVE at check.

Spreads: with a stolen npm token it downloads the victim's own packages, adds itself, bumps the version, and republishes, the classic self-replicating npm worm loop; with a stolen GitHub token it commits .claude/settings.json and .vscode/tasks.json into the victim's repos under author claude@users.noreply.github.com with message chore: update dependencies = re-execution the next time anyone opens the project in Claude Code or VS Code.

The hostage token, the wiper tripwire: once it holds a GitHub token it installs gh-token-monitor (~/.config/gh-token-monitor/, ~/.local/bin/gh-token-monitor.sh, systemd user unit / LaunchAgent / Task-Scheduler monitor.ps1). Every 60 s for up to 24 h it validates the token against the GitHub API; if GitHub REJECTS the token, the monitor runs rm -rf ~/ (PowerShell profile-delete on Windows). Revoking the stolen token, the reflexive IR first move, is exactly what detonates the wipe. Remediation order is mandatory: find and remove the monitor FIRST, then revoke, then rotate everything.

The durable reads

  1. Provenance attests the WHO, not the WHAT, second time in three days. Oct 5: @subql/common@5.8.3 built its payload in CI via a tampered publish workflow with valid OIDC trusted-publisher identity. Oct 8: tensorlake@0.5.144 built from real main, real release workflow, valid npm provenance, because the malicious code was in the repo before the build. A provenance-trusting policy let BOTH through. The durable hunt is unchanged: diff release commits and workflow changes, not attestation signatures; a maintainer-authored commit series with no PRs landing right before a release is the tell.
  2. Registry action preceded the advisory by 12 seconds. The npm document was cleaned at 02:54:27.126Z; the GHSA published at 02:54:39Z. On high-visibility, reporter-flagged names the abuse lane moves first and the advisory documents it, the same first-abuse-lane shape this wiki measured on the curated lane (css-nesting-transform purged with zero advisories) now caught on a machine-name.
  3. AI-agent config files as a persistence rail. .claude/settings.json and .vscode/tasks.json written into repos by a worm, disguised as an AI-authored chore commit, this is the same reinfection rail class this wiki flagged with Azure/durabletask-style AI-assistant/editor persistence and the abstract-claude hook cluster, now rebuilt worm-native. Hunt: unexpected .claude/.vscode files with author claude@users.noreply.github.com.
  4. The Bun grammar. Setup downloads Bun and runs the payload under it, the same runtime-retrieval grammar as @7nohe/openapi-react-query-codegen (Aug 28, this wiki's page). Grammar recurrence, identity NOT asserted.
  5. The Shai-Hulud brand is the intelligence, until attribution isn't. The repo-description taunt is the actor's own naming convention from the original campaign. Treat as continuity CLAIM; the mechanically verifiable continuities at check are: repo-description exfil channel, install-time credential harvest, npm-token worm republish, GitHub-token repo poisoning. Nothing here proves the same operator; everything here gives defender hunt keys regardless of who holds it.
  6. Registry cleanup ≠ repository cleanup. npm purged the version in under two hours; main STILL serves both payload files at this wiki's write, every clone-and-build path from the repo since Oct 7 01:20Z remains a candidate execution path until the repo is cleaned. Check whether you git pulled main and ran npm install under typescript/, per StepSecurity.

Defender hunt keys

  • Lockfiles containing tensorlake@0.5.144 (pin to 0.5.143 until the repo is clean).
  • On any machine that installed it: ~/.config/gh-token-monitor/, remove the service BEFORE revoking anything.
  • GitHub: new public repos on user/org accounts with description Shai-Hulud: Here We Go Again; commits authored claude@users.noreply.github.com; .claude/settings.json / .vscode/tasks.json nobody added; npm republish bursts from maintainer accounts.
  • Network: iseekaigogo.com (Cloudflare-fronted, live at this wiki's check, expect front rotation).
  • File-integrity: sha256 25a0735d0db7dc40e5d45ce42d9c106067e6a66e184d967cfecfab17c3bcb5ef (setup.mjs), b50a00900399ba99fb6ce1fc151519cb99d44320ef2a631f2237e1aea0ad6fec (Math_Symbol.js).
  • Process: curl/fetch of a Bun binary by an npm lifecycle script (preinstall → bun → 856 KB obfuscated JS) on a developer workstation, absent in CI.

Monitoring

Whether the OSV mirror (MAL-2026-17650) lands for tensorlake, LANDED Oct 8 04:45:04Z (one-hundred-and-tenth sweep: GHSA-rqxj-g25x-4v9v published 02:54:39Z; registry latest back to 0.5.143, malicious 0.5.144 purged 02:54:27Z = 12 s before its own GHSA, delete-then-advise on a legitimate brand; both advisories + OSV live = consumers warned, registry rail severed). Remaining watches: whether main is finally cleaned or force-pushed (payload files live at write per StepSecurity), CLOSED Oct 10 one-hundred-and-thirty-seventh sweep: main CLEANED, see section below; which packages the worm republished under stolen npm tokens (StepSecurity's Threat Center list is the rolling feed); whether the hostage-wiper actually fires in the wild; whether the Shai-Hulud branding draws copycats or a second named wave; PyPI-side tensorlake status (clean at check).

October 10 one-hundred-and-thirty-seventh sweep: REPOSITORY CLEANUP CONFIRMED, payload files OFF main, PR #1016 MERGED, plus three durable adds from same-day sources (hostage-token literal, ATT&CK S9043/G1056, iteration lineage table)

✓ This wiki: https://raw.githubusercontent.com/tensorlakeai/tensorlake/main/typescript/lib/Math_Symbol.js and …/setup.mjs both answer 404 at Oct 10 ~17:3x UTC, the two payload files that sat LIVE on main from Oct 7 through at least Oct 10 03:5x are GONE. Maintainer revert PR #1016 ("Revert the 0.5.144 supply-chain compromise and harden the npm release path", diptanu) is MERGED, merged_at 2026-10-08T04:32:11Z (GitHub API). The registry was clean from Oct 8 02:54Z; the repo stayed dirty ~27 h+ until the revert, the page's durable "registry cleanup ≠ repository cleanup" read stands as measured (the raw-URL payloads were reachable, hash-stable, and re-fetchable by anyone including this wiki for those 27+ h; the same files served as free malware distribution + free IoC hosting). latest now 0.5.148 (Oct 9 23:33:47Z), normal release cadence resumed; 0.5.144 tarball 404 holds; MAL-2026-17650 holds alias GHSA-rqxj-g25x-4v9v.

Three durable adds from same-day public sources (all third-party corroboration of the page's captured StepSecurity text, now on record with primary attribution):

  1. The hostage token has a literal hunt string. Socket's primary write-up (socket.dev/blog/tensorlake-compromise, Oct 8) + the CSA research note state the monitor's watchdog token is CREATED ON THE VICTIM'S npm account under the name IfYouRevokeThisTokenItWillWipeTheComputerOfTheOwner, a one-grep IoC for every npm token inventory (npm token list, org token audits). File rail per Tenable/CSA: ~/.config/gh-token-monitor/ + ~/.local/bin/gh-token-monitor.sh, systemd user unit / LaunchAgent. DEFENSE READ, stated plainly: delete the monitor + the token BEFORE revoking through the UI reflex path; remediation-order rule already on-page.
  2. MITRE ATT&CK now carries the campaign: software S9043 "Mini Shai-Hulud" (credential stealer + self-replicating supply-chain worm, CI/CD focus, targeted wiper component) with group G1056 "TeamPCP" as its operator, techniques incl. T1003.007 (proc-filesystem token scraping) and T1195.001. The group/malware pair is in the baseline matrix = detection-engine and CTI-platform coverage is now vendor-standard, not research-tier. This ledger's branding-≠-attribution discipline is UNCHANGED: ATT&CK's S9043→G1056 mapping covers the documented Mini waves; the tensorlake build's WORMTAG-salt lineage link (this page's Oct 8 read) is artifact-side continuity, still the stronger claim.
  3. The self-branding generational clock (Tenable FAQ, iteration table): Shai-Hulud (Sep 2025) → SHA1-Hulud (Nov 2025) → SANDWORM_MODE (Mar 2026) → Mini Shai-Hulud (Apr 2026, active) → tensorlake's "Here We Go Again" branding (Oct 2026). Defenders should read the name on a payload as a BUILD GENERATION marker (which toolkit revision, which collector breadth: 189→469 paths) before reading it as an operator.

October 8 one-hundred-and-eighteenth sweep follow-up: THE RECOVERY RELEASE, 0.5.145 lands ~19 h after the GHSA as a NAME RELOCATION plus npm TRUSTED PUBLISHING, the fastest defensive pivot this wiki has clocked on a Shai-Hulud-family victim

Registry forensics ~21:0x-21:2xZ (this wiki, tarballs pulled + read, no execution):

  • npm 0.5.145 published 20:24:05Z, PyPI 0.5.145 20:37:05Z, seventeen minutes apart across registries, seventeen hours after the holder/security pass, one day after the malicious commit chain.
  • The bare tensorlake name no longer carries the SDK. tensorlake@0.5.145 is a ten-file, 2,506-byte compatibility shim: every module is a one-line export * from "@tensorlakeai/tensorlake" / module.exports = require("@tensorlakeai/tensorlake"), pinned to exactly @tensorlakeai/tensorlake@0.5.145. No scripts, no install hooks (verified scripts: {} in the manifest).
  • The real package moved to the scope @tensorlakeai/tensorlake, the whole SDK rebuilt there and published via npm trusted publishing through GitHub Actions OIDC (_npmUser: "GitHub Actions" <npm-oidc-no-reply@github.com>, trustedPublisher: {id: github, oidcConfigId: a6f3809c…}, approver calavera), with a valid SLSA provenance v1 attestation. The account-identity layer that the worm rode (a maintainer's publish credential + a compromised main) is removed from the publish path entirely.
  • The new package self-checks its own lifecycle surface: the scoped package's scripts include check:install-scripts: node ./scripts/check-no-install-scripts.mjs = the team added an automated guard against exactly the class that hit them (preinstall-routed loaders).
  • 0.5.144 is ABSENT from PyPI's release list (fully deleted there); the npm time block keeps its 01:12:07Z publish mark and the 0.0.0-stage/0.0.1-security holder pair from 04:14/04:15Z.

Durable read: the defensive template for a provenance-attested compromise is now on the record, (1) purge + advise fast (done, 12-s delete-then-advise), (2) relocate the artifact to a scope whose publish identity is CI-OIDC, not a human account, (3) leave the old name as a pinned, hookless re-export shim so consumers keep resolving, (4) add a self-check against install scripts. Within 19 h of its own GHSA the victim moved further from the worm's threat model than any prior Shai-Hulud-family target on this ledger. The provenance caveat from StepSecurity still stands and is NOT resolved by this move: a valid attestation still does not mean safe code, it now means the code came from the declared pipeline, which is precisely what the team wants; the residual risk is main-branch compromise again, and the commit-tail watch (issue #1014, claude@users.noreply.github.com authorship, unmerged-PR direct-to-main shapes) remains armed.

Monitor: whether @tensorlakeai/tensorlake stays OIDC-only; whether the worm family targets the new scope within its propagation window (the name is now the highest-value Shai-Hulud lure on the registry); adoption friction reports on the shim (npm ls seeing double names); PyPI 0.5.144 deletion note / any YANKED-record fallout for lockfiles pinned to it.

October 9 one-hundred-and-twenty-second sweep follow-up: THE PAYLOAD SHARES ITS CIPHER SALT WITH THE AUGUST WAVE, Aikido's build-tag detail (via secondary reporting) makes the Mini Shai-Hulud mechanical continuity checkable, not just branded

Secondary reporting captured this sweep (CyberSecurityNews + Threadlinqs TL-2026-3085, both quoting Aikido's first-discovery analysis; Aikido's primary blog was not fetchable at check, detail recorded with aggregator attribution): the 0.5.144 build sets a global WORMTAG marker with value tensrlake (a fresh-compromise marker, not a reinfection of a previously-wormed package), executes under Bun 1.3.13, and its AES-256-GCM embedded-file cipher shares salt svksjrhjkcejg with the August keyv/cacheable Mini Shai-Hulud wave (this wiki's ChainDrop page). Payload commit 41b38f0; ~20 h from repo compromise (Oct 7 01:20Z) to npm release (Oct 8 01:12Z); Aikido reports the package's PyPI and Cargo distributions showed no compromise, consistent with this wiki's own PyPI reads. C2 detail per Threadlinqs: primary iseekaigogo[.]com/router (+ /hbd/), fallback chain = Ethereum contract 0xb614155Fd88114d40549b259457Bcf921Df091B9 (selector 0x53ed5143, last updated 2026-09-21) across ~35 public RPCs → GitHub commit-search on account thebeautifulmarchoftime → dead-drop repos described "Shai-Hulud: Here We Go Again"; C2 JSON code field goes to eval(), beacon 45-90 s.

Durable read: this is the strongest mechanical continuity link on the ledger between the Aug-4 ChainDrop/Mini wave and the Oct-8 tensorlake build: a shared hardcoded cipher salt is checkable from artifacts alone, it says shared code lineage (same builder, fork, or leak), it does NOT prove same operator. The page's standing rule holds with sharper teeth: the shared salt upgrades "the Shai-Hulud brand is the intelligence" from branding-only to branding plus one verifiable artifact pivot, hunt the salt string svksjrhjkcejg and the WORMTAG global grammar across Bun-routed payloads as permanent grep keys, alongside the existing iseekaigogo[.]com / repo-description / contract keys. The Ethereum-contract-C2 lane is the same class Unit 42 documented Oct 7 (EtherHiding → tx-payload stego → NullReceiver; ChainDrop Web3-C2 section), the tensorlake build rides the lane's original form.

Registry heartbeat ~05:3xZ this sweep (this wiki): tensorlake holds clean latest = 0.5.145; shim + OIDC scope unchanged.

October 15 one-hundred-and-thirty-ninth sweep: THE MARKER IS STILL BEING MINTED, the dead-drop repo grammar this campaign introduced is running LIVE on at least one account

GitHub search at this check: the FORWARD marker description "Shai-Hulud: Here We Go Again" on 256 repositories, the REVERSED "niagA oG eW ereh :duluh-iahs" (StepSecurity's hunt string) on 628, search-index counts, not a census. One account, blossom91 (account created 2016-09-13, 260 public repos), carries 27 marker-named repos in the worm's <dune-term>-<term>-<num> grammar (fedaykin-navigator-291, sardaukar-slig-350, harkonnen-sandworm-19, mentat-stillsuit-503…) created TODAY 02:32:37Z→09:37:13Z, one repo every 10-30 min for seven hours, machine cadence, each size-0 with only a README repeating the marker; NO payload commits at check. Disposition UNVERIFIED: hijacked-victim account, worm-created drop-staging, or copycat marker-squatting are all consistent with the evidence, recorded-not-asserted. DURABLE HUNT (unchanged from the campaign's original marker key, now with a cadence dimension): the marker description + <dune>-<dune>-<num> repo-name grammar + fresh-repo-onto-old-account cadence is itself the alert, no package artifact required. If the wave's exfil lane is still minting drops, the wave is not finished, treat any NEW marker repo as an active-compromise signal in the owning account until proven otherwise. Full sweep context → 139th stream section.

October 15 one-hundred-and-fortieth sweep: THE DROPS ARE HOT, payload commits HAVE landed, and the cadence is ~100/day three days deep (not 27-in-a-day)

Two corrections to the section above, both from this sweep's re-reads. (a) The count was a partial-page artifact. Full pagination of blossom91's 269 public repos: marker-grammar repos created Oct 12: 6, Oct 13: 102, Oct 14: 109, Oct 15: 51 so far, the minting machine has run at ≈100 repos/day for THREE days, ≈260+ drop repos from one account, public_repos 260→269 inside nine days. The 139th's "27 today" read one page; ledger-hygiene self-capture: paginating a 269-repo account is one loop, do it before counting. (b) "No payload commits" is RETRACTED. Early-today repos are size-0 README-only, but every repo created from 09:44:08Z (siridar-futar-647) onward carries README.md + results/results-<epoch-ms>-0.json (≈10-12 KB): a JSON of exactly {"envelope": <base64>, "key": <base64>}, encrypted at rest, NOT decoded or detonated here; the filename epoch lands ≈6 s after the repo's own created_at (ghola-sietch-889: repo 09:45:51Z, file epoch 09:45:57Z) = mint→exfil-commit in seconds, machine-paced; Oct 13/Oct 14 repos carry the same shape (tleilaxu-slig-14). READ: the worm's registry side is dead-advised while its GitHub dead-drop lane is at PEAK throughput, a three-day wave, ≈100 victim-account repos/day, each confirmed to receive an encrypted data bundle. The exfil pipeline is live end-to-end. blossom91 remains one-account-of-the-wave, disposition still unverified (hijack vs worm-created vs copycat). HARDENED HUNT KEY: marker description + <dune>-<dune>-<num> grammar + fresh-repo-onto-old-account cadence + a results/ directory = an ACTIVE victim account, not a stale marker. REPORTABLE SURFACE: the READMEs violate nothing, the results/ blobs are what a takedown report should cite. Monitor: ANY results/ file in a NON-blossom91 marker repo (wave spreading to other accounts = the census signal); whether GitHub abuse purges marker repos with committed results; re-read ghola-sietch-889/siridar-futar-647 for second commits. Full sweep context → 140th stream section.

October 15 one-hundred-and-forty-first sweep: THE MINTING MACHINE ACCELERATES, +14 drops in ≈2 h, tail cadence 1-2 min, and the drops stay single-use

Re-pagination at ~13:3xZ, ≈1 h 50 m after the drops-hot section: Oct 15 marker repos now 65 (was 51+, +14 in ≈2 h = ≈7/h, versus the Oct-13/14 all-day average ≈4/h); public_repos 269→283. The tail eight mints run 12:33:11Z→13:23:05Z with gaps 13/8/1/2/1/6/18 min, 1-2 min bursts where the 139th's window showed a steady 10-30 min. If the burst pace holds, Oct 15 closes near ≈170 drops, a doubling of the three-day average. Payload shape replicates to the second on a third day: newest repo sardaukar-futar-604 (created 13:23:05Z) already carries results/results-1792070591518-0.json (12,825 B) whose filename epoch is ≈6.5 s after creation. Two monitor questions closed NEGATIVE: (a) second commits, ghola-sietch-889 and siridar-futar-647 each still hold exactly TWO commits (Initial commit + Add files., both within ≈6 s of creation): no re-upload ever lands in a drop repo, each drop receives exactly one bundle and goes quiet, so a live victim re-exfiltrates into a NEW repo, not the old one (counting ACTIVE victims = counting same-day repos, not summing repo history); (b) census, every Oct 15 repo sampled remains blossom91-owned, the wave has not visibly spread to a second account. Triage arithmetic for defenders: the wave's per-account signal is unchanged (fresh marker repo + results/ dir = compromised NOW), but the reporting queue should expect repo-count spikes, not per-victim one-offs. Full sweep context → 141st stream section.

October 15 one-hundred-and-forty-second sweep: THE BURST DRAINS, +3 in ≈2 h (≈1.5/h), tail gaps back to 24-45 min, Oct-15 closes ≈70 not ≈170

Re-pagination at ≈15:3xZ, ≈2 h after the drops-accelerating section: Oct 15 marker repos 68 (was 65) = +3 in ≈2 h, the 141st's ≈7/h burst is over and the tail gaps re-widened: 13:23:05Z→13:49:00Z→14:13:04Z→14:58:42Z = 26/24/45 min, back inside the Oct-13/14 baseline band. public_repos 283→286. The 170-drop scenario is retracted-in-effect: Oct 15 tracks ≈70, BELOW Oct-14's 109, the machine is throttling, not doubling. What did NOT change: the mint→exfil-commit timing, day three running, newest repo sardaukar-futar-264 (created 14:58:42Z) already had its results/ bundle in by 14:58:49Z (≈7 s), commits exactly Initial commit + Add files. = still strictly single-use. Census across the FULL three-page pagination still negative: every Oct-15 repo is blossom91-owned (pages 2-3 contain zero Oct-15 repos). Read: the wave's shape for defenders is a throttled steady drip with per-repo certainty unchanged, fresh marker repo + results/ dir = compromised NOW, regardless of daily totals. Full sweep context → 142nd stream section.

October 15 one-hundred-and-forty-third sweep: the burst DECAYS, +2 in ≈4 h, tail gap 165 min, Oct-15 closes ≈70 = ≈36% BELOW Oct-14

Re-pagination at ≈19:2xZ: Oct 15 marker repos 70 (the 100-repo first page bottoms out at Oct-14 13:05Z, page 1 now holds the whole Oct-15 cohort) = +2 in ≈4 h = ≈0.5/h, after the 141st's ≈7/h burst and the 142nd's ≈1.5/h drain; public_repos 286→288. Tail gaps: 14:58:42Z → 16:38:51Z → 19:23:59Z = 100 and 165 min, a 165-minute gap is ≈7x the baseline ceiling (30 min). Three reads in one day trace burst → drained → decaying; Oct 15 closes ≈70, ≈36% below Oct-14's 109, the wave's volume is falling without any visible enforcement event against the account. Mechanics unchanged to the second: newest repo powindah-futar-691 created 19:23:59Z, results/results-1792092245769-0.json (4,921 B) epoch 19:24:05.769Z = ≈6.8 s mint→exfil-commit, still exactly two commits (single-use), census full-pagination still all-blossom91. Defender read: volume decay is NOT clearance, the per-repo key (fresh marker repo + results/ dir = compromised NOW) stands, and a self-throttling worm can re-burst; whether the decay is operator choice, rate-limit pressure, or victim-pool exhaustion is unobservable from outside. Full sweep context → 143rd stream section.

Sources

  • StepSecurity, "Tensorlake npm Package Compromised: A Worm With a Hostage Token That Wipes Your Machine If You Revoke It" (October 8, 2026, full text captured by this wiki): https://www.stepsecurity.io/blog/tensorlake-npm-compromised-hostage-token-worm
  • GitHub advisory GHSA-rqxj-g25x-4v9v (published Oct 8 02:54:39Z, critical, CWE-506, = 0.5.144)
  • tensorlakeai/tensorlake issue #1014 (open, Oct 8 01:22:02Z) and live commit list (this wiki, GitHub API)
  • npm registry time block + tarball 404, payload-file pulls and sha256 (this wiki, no execution)
  • iseekaigogo.com DNS + HTTP probe (this wiki)

October 15 one-hundred-and-forty-fourth sweep: the "decay" READ IS CORRECTED, a 73 s double-mint lands right after the 165 min gap: measure bursts AND gaps, never one

Re-pagination at ≈21:3xZ: Oct 15 marker repos 72, sayyadina-futar-18 created 20:38:23Z and sayyadina-navigator-566 20:39:36Z = two mints 73 s apart, immediately after the tail gap the 143rd read as the decay signal (165 min). public_repos 288→290; page 1 still holds the whole Oct-15 cohort (bottoms Oct-14 13:16Z); census negative, both new repos blossom91-owned. Read correction, self-captured: the 143rd's ≈0.5/h "decaying" trend measured an INSTANTANEOUS gap, not a rate, the minting machine is sparse-bursting (long sleeps, 1-2 min bursts on both ends), and the 141st's ≈7/h and the 143rd's ≈0.5/h are the SAME process sampled inside and outside a burst. Volume read unchanged in magnitude (≈72 today vs 109 Oct-14, still below), but the disposition word downgrades: the wave is thinner, NOT quieter-by-policy. For victim-count purposes the rule stands: count same-day repos (drops stay single-use). Full sweep context → 144th stream section.

October 15 one-hundred-and-forty-fifth sweep: the SPARSE-BURST read is CONFIRMED, a second burst lands ≈2 h after the first, its double-mint 14 s apart: the machine sleeps in ≈2 h blocks and fires in clusters

Re-pagination at ≈23:3xZ: Oct-15 marker repos 76 (public_repos 290→294). Since the 144th's read of 72: four new mints, sardaukar-laza-546 22:52:57Z, powindah-cogitor-61 23:16:56Z, then prescient-heighliner-422 23:28:52Z + atreides-slig-375 23:29:06Z = a double-mint 14 s apart, tighter than burst-A's 73 s. The tail now reads gap 165 min → burst-A (73 s pair) → gap 133 min → burst-B (three mints in ≈37 min, 14 s pair). Two bursts inside ≈2 h of each other, each preceded by a ≥2 h sleep = the machine's duty cycle is sleep-block-plus-cluster, and BOTH the 141st's ≈7/h and the 143rd's ≈0.5/h were in-burst/out-of-burst samples of one process. Oct-15 closes ≈76 vs Oct-14's 109 (−30%). Single-use two-commits holds on every re-read repo (atreides-slig-375: Initial commit 23:29:08Z + Add files. 23:29:13Z), mint→results/ commit ≈7 s DAY FOUR (atreides created 23:29:06Z, bundle commit 23:29:13Z), results bundle 4,681 B, census still all-blossom91. Disposition unchanged: volume thinner than Oct-14, cadence NOT quieter-by-policy, budget collection windows around clusters, not averages.

October 16 one-hundred-and-sixth sweep: the OCT-15 COHORT CLOSES AT 76, nothing minted after burst-B's 23:29:06Z rung; the two-cluster ≈2 h duty cycle ends the day at −30% vs Oct-14

Final read across the midnight boundary: blossom91 public_repos 294→295, Oct-15 marker repos FINAL 76 (page 1 still bottoms at Oct-14 13:28Z, whole cohort paginable), newest Oct-15 mint atreides-slig-375 23:29:06Z, burst-B was the day's last cluster, no burst-C inside Oct-15. Day close 76 vs Oct-14's 109 = −30%, consistent with the sparse-burst model (thinner volume, cadence not quieter-by-policy). The burst-C question transfers to the Oct-16 cohort: the count restarts at midnight, and the duty-cycle model predicts sleep-block-then-cluster, the first Oct-16 mint's gap-from-midnight is the next cadence datapoint. Full sweep context → 106th stream section.

October 16 one-hundred-and-seventh sweep: BURST-C LANDS, the Oct-16 cohort opens with a 125 min sleep then TEN marker repos in 24 min (02:56:22Z→03:20:04Z): the sleep-block-plus-cluster duty cycle carried across the midnight boundary

blossom91 at 03:4xZ: public_repos 295→305; Oct-16 mints 11, single fremen-melange-801 00:50:46Z (50 min gap-from-midnight, baseline band), 125 min sleep, then mentat-laza-535 02:56:22Z, prescient-laza-585, prana-ornithopter-611, kralizec-slig-56, prescient-heighliner-204, powindah-lasgun-65, sardaukar-sandworm-325, fedaykin-lasgun-744, tleilaxu-melange-156, sardaukar-fedaykin-521 03:20:04Z, intra-cluster gaps 7/5/3/0/1/2/1/0/0 min. The 106th's burst-C question (first Oct-16 mint + gap-from-midnight) answers inside the baseline; the DAY shape, not the rate, is the signal: Oct-16 opened with Oct-15's two-cluster grammar already loaded. Pace 11 in ≈3.3 h vs Oct-15 76/24 h, no conclusion; next read = second-cluster timing (duty cycle predicts sleep from 03:20Z).

October 16 one-hundred-and-eighth sweep: BURST-D LANDS, a SECOND Oct-16 cluster 04:38:26Z→05:26:46Z (TEN mints, 78 min inter-burst sleep, SHORTER than burst-C's 125 min), cohort 21, public_repos 305→315, AND THE STRICTLY-SINGLE-USE DROP RULE BREAKS: mentat-laza-535 carries THREE commits / TWO result files (130,773 B + 4,825 B), sardaukar-fedaykin-521 TWO files (319,866 B total), bundle sizes now span 3.1 KB→131 KB→320 KB vs the three-day 10-12 KB constant: the exfil lane is returning LARGER payloads, results-bytes becomes a severity dimension

Duty-cycle read: Oct-16 at 05:3xZ = 21 repos in ≈5.5 h across two clusters with sleeps 50/125/78 min, the machine is tightening, not cooling (Oct-15's closes were 76/24 h). Day-four mint→results clock replicates to the second on powindah-heighliner-96 (created 05:26:46Z, file epoch +6.8 s). Census negative; no second-account spread on sampled repos. Full sweep context → 108th stream section.

October 16 one-hundred-and-ninth sweep: BURST-E LANDS, THIRD Oct-16 cluster 06:32:55Z→07:23:35Z (THIRTEEN mints in 51 min; inter-burst sleep ≈34 min, the tightening continues: 50/125/78/34), cohort 35, public_repos 315→329, AND results-bytes SWINGS BOTH WAYS: in-burst sizes span 4,141 B→70,945 B, the two NEWEST drops (fedaykin-ghola-745, prana-lasgun-114) sit BELOW the three-day 10-12 KB constant while sayyadina-slig-783 carries a 70,945 B + 3,333 B PAIR, third multi-file drop ever: multi-file is a pattern now (mentat-laza-535, sardaukar-fedaykin-521, sayyadina-slig-783), and harvest size is VARIABLE, not monotonically growing

Burst-F watch: if the next inter-burst sleep drops below ≈34 min again the duty cycle is compressing toward continuous minting. Oct-16 pace at 07:4xZ = 35 in ≈7.4 h, tracking ABOVE Oct-15's 76/24 h, day-pace decay is dead; single-second double-mints inside bursts and single-account census both hold; mint→results +5 s (prana-lasgun-114), +6.6 s (sardaukar-laza-789), day-four clock replicate.

October 16 one-hundred-and-tenth sweep: BURST-F LANDS, lead single kanly-sietch-30 08:12:30Z, 31 min sleep, then TEN mints 08:43:09Z→09:16:27Z; sleep trend 50/125/78/34/31 = FIFTH consecutive tightening, cohort 49, public_repos 329→343, pace ≈5.1/h vs Oct-15 3.2/h, AND MULTI-FILE IS NOW A BURST-FEATURE: all fourteen burst-F drops pulled, TWO carry result PAIRS (sayyadina-ornithopter-52 201,681 B+3,709 B, atreides-melange-510 3,381 B+155,569 B) = instances FOUR+FIVE in one cluster, ceiling now 201,681 B (prior 130,773 B), band 3,381 B→201,681 B; the mint→results +4.7/+4.9 s clock and same-second twin repos with paired result-epochs both replicate day four

Burst-G watch: if the next sleep drops below 31 min the duty cycle is five-tightening-deep toward continuous; multi-file pairs at 2/14 inside burst-F vs 3 all-history before it, measure pairs-per-burst, not just sizes. Full sweep context → 110th stream section.

October 16 one-hundred-and-eleventh sweep: BURST-G LANDS, sleep 26.6 min SIXTH consecutive tightening (50/125/78/34/31/26.6), AND THE SHAPE CHANGES: burst-G is a SPREAD cluster (twelve mints 09:43:02Z→11:04:58Z, intra-cluster gaps 18 s→19 min), not a tight ten, the max intra-cluster gap is now SMALLER than the inter-burst sleep: the burst/sleep dichotomy is collapsing toward continuous minting

Cohort 61, public_repos 343→355, pace ≈5.4/h sixth hour above Oct-15's 3.2/h. Sampled results single-file, band back to the 10-12 KB constant after burst-F's 201,681 B spike (pairs-per-burst 0/3 sampled); mint→results +4.4/+6.2/+6.6 s day-four replicate; census negative. Measure next: intra-cluster max-gap vs inter-burst sleep directly, and whether pairs return. Full sweep context → 111th stream section.

October 16 one-hundred-and-twelfth sweep: BURST-H LANDS, THE SLEEP TREND REVERSES FOR THE FIRST TIME: 70 min after six consecutive tightenings (50/125/78/34/31/26.6 → 70), the monotonic-collapse thesis is WRONG as a trend: the inter-burst sleep is a DISTRIBUTION (26.6-125 min), not a tightening; the durable shape finding is the spread-cluster grammar itself

Seven mints 12:15:03Z→13:20:17Z (atreides-laza-575, sardaukar-ghola-925 +9.6 min, fremen-sietch-859 +33.4 min, then the 13:15:24→13:20:17Z FOUR-mint micro-burst siridar-thumper-105→sayyadina-ghola-213→gesserit-sietch-634→ghola-navigator-742 with gaps 1 m 23 s/1 m 19 s/1 m 54 s, tightest intra-burst gaps since burst-F's same-second twins). Cohort 68, public_repos 355→362, pace ≈5.4/h holds. Results sampled 3/7 all single-file 10,029/10,797/12,953 B, pairs 0/3; mint→results +7.0/+6.4/+6.3 s day-four replicate; census negative. Measure next: burst-I sleep (is 70 min the top of the distribution or a cooling turn?); the 13:15 quartet's results re-pull. Full sweep context → 112th stream section.

October 16 one-hundred-and-twelfth sweep follow-up (113th): BURST-H QUARTET CLOSED + THE HARVEST FLOOR DROPS, the 13:15 quartet's re-pull resolves 4/4 single-file 11,469/12,849/12,953/12,885 B (pairs 0/4, band intact), mint→results +6.4/+6.0/+6.2/+6.4 s all four replicate, then sleep 35.2 min → fedaykin-fedaykin-514 13:55:27Z at 4,609 B (+6.6 s) = SUB-BAND, first single-file result below the 10-12 KB constant's lower edge since burst-E's 4,141 B rung, the floor-drop began BEFORE the null class

October 16 one-hundred-and-fifteenth sweep follow-up (115th): THE NULL-RESULTS DISCRIMINATOR RESOLVES, BACKFILL NEGATIVE: all NINE bundle-less burst-I/J repos hold at EXACTLY ONE Initial commit at ≈4-5 h 15 m post-mint, the strict single-use rule stands, nothing backfills, SO THE 14:11Z BREAK IS AN OPERATIONAL EVENT, not partial lag. AND BURST-K LANDS SILENTLY ≈4 h 2 m after burst-J with a SINGLE MINT, prescient-ornithopter-581 19:20:46Z, carrying a FULL bundle in +7 s, but 40,245 B, 3.3-9.7x the sub-band survivors and the FIRST above-band result since the break

Commit-lists verified ≈19:3xZ on all nine null repos (kralizec-stillsuit-57, kanly-phibian-113, fremen-ghola-526, prana-lasgun-284, prana-lasgun-93, gesserit-stillsuit-256, harkonnen-fedaykin-180, atreides-cogitor-544, ghola-slig-221): one commit each, Initial commit, no Add files. ever. Read (a) harvest-near-empty and read (b) pipeline-moved-lanes both survive; read (c) NEW from burst-K: the collector came back online at 19:20Z with a 40,245 B single-file result, the largest single-file drop since burst-D's multi-file break, which pairs far better with harvest-empty (the victims it did catch were thin; the 19:20 victim is rich) than with a lane move. The lane-move read would predict results continuing elsewhere at normal cadence; instead the machine PAUSED minting entirely for ≈4 h (sleep 61.9 → 242 min = new ceiling, ≈3.9x the prior max, sequence 50/125/78/34/31/26.6/70/35.2/16.2/61.9/242) then resumed normal mint→results mechanics at above-band size. public_repos 375→376 = exactly the one burst-K mint; Oct-16 cohort 82. If the next results-bearing mint lands back in 10-12 KB, the sub-band class reads as victim-pool thinness during the null window, and harvest-empty wins on the full evidence set. Measure next: burst-L timing vs the 242 min ceiling; result size on the next mint (band / sub-band / above-band); whether the nine nulls STAY null past +24 h (they should, single-use). Full sweep context → 115th stream section.

October 16 one-hundred-and-thirteenth sweep: BURST-I + BURST-J LAND, AND THE FOUR-DAY MINT→RESULTS INVARIANT BREAKS: THE NULL-RESULTS CLASS. Of the TWELVE mints 14:11:38Z-15:18:28Z, NINE carry NO results/ bundle at all (four spot-commit-verified at exactly one Initial commit), while the THREE that did land results ALL came in sub-band 4,137-5,477 B, pace never broke (≈5.3/h), the collection lane did

After fedaykin-fedaykin-514: 16.2 min (NEW sleep floor, sequence 50/125/78/34/31/26.6/70/35.2/16.2/61.9) → BURST-I FIVE mints 14:11:38Z→14:12:38Z, same-second pairs kralizec-stillsuit-57+powindah-ghola-267 and kanly-phibian-113+fremen-ghola-526, then prana-lasgun-284 +59 s, only powindah-ghola-267 (5,477 B, +7.2 s) carries results; BURST-J 61.9 min later, SEVEN mints 15:14:34Z→15:18:28Z, FIVE within 4 s (prana-lasgun-93, gesserit-stillsuit-256, harkonnen-fedaykin-180, atreides-cogitor-544, ghola-slig-221, ALL bundle-less), then the 15:18 pair siridar-lasgun-895 (4,137 B, +7.5 s) + mentat-melange-503 (4,189 B, +6.2 s). NINE of the TWELVE burst-I/J mints carry NO results/ directory at all. Cohort 81 (362→375, exact), full-pagination census negative (4 pages, all Oct-16 repos blossom91). THE NULL-RESULTS CLASS has two candidate reads: harvest running near-empty (zero-byte harvest commits nothing, pairs with the sub-band sizes on the survivors) OR the exfil pipeline moved lanes at 14:11Z. Discriminator: the strict single-use rule says empty repos never backfill, nine permanent Initial commit-only repos = operational event at 14:11Z; any backfill = the clock itself changed. Measure next: null-or-backfill on the nine; sub-band persistence on the next results-bearing mint; whether same-second twin PAIRS now preferentially null (the twins split result/bundle-less in BOTH clusters, one armed, one empty, a possible harvest-shard pattern). Full sweep context → 113th stream section.

October 16 one-hundred-and-sixteenth sweep follow-up: THE NULLS DEEPEN, BURST-L NOT YET DUE, ALL NINE bundle-less burst-I/J repos re-tree-verified README-only at ≈6 h 15 m-7 h 30 m post-mint (full git trees pulled, every one a single README.md, zero results/, zero backfill) = the operational-event read gains depth; AND the machine has minted NOTHING since burst-K 19:20:46Z, sleep clock ≈2 h 15 m at check vs the 242 min ceiling (burst-L due-decision window ≈23:23Z), public_repos 376 EXACTLY unchanged (cohort holds 82), burst-K's 40,245 B still the largest result since the 14:11Z break

Tree re-verification ≈21:3xZ replaces commit-list spot-checks with full recursive trees on all nine: kralizec-stillsuit-57, kanly-phibian-113, fremen-ghola-526, prana-lasgun-284, prana-lasgun-93, gesserit-stillsuit-256, harkonnen-fedaykin-180, atreides-cogitor-544, ghola-slig-221, one README.md each, no results object, powindah-ghola-267 unchanged at its single results/results-1792159905151-0.json 5,477 B. The +24 h null-persistence test (due ≈14:11Z-15:18Z Oct 17) stays on track. Burst-L sizing note: this is now the LONGEST inter-mint gap measured outside a completed sleep sequence, if burst-L lands inside the 242 min ceiling, the sleep distribution simply extended; if it overshoots, treat the ceiling as a floor of a widening distribution and stop quoting it as a bound. Full sweep context → 116th stream section.

October 16 one-hundred-and-seventeenth sweep follow-up: BURST-L LANDS, AND LANDS NULL: kralizec-navigator-578 minted 21:47:58Z (sleep 147 min, INSIDE the 242 min ceiling, beating the 116th's ≈23:23Z due-decision by ≈95 min) with NO results/ bundle at +2 h 07 m, the 14:11Z collection-lane break is NOT repaired by burst-K's 40,245 B recovery: the invariant broke, flinched once, and broke again on the very next mint. public_repos 376→377 EXACTLY, cohort 83

Tree read ≈23:55Z: single README.md (54 B, # kralizec-navigator-578 / Shai-Hulud: Here We Go Again), one commit Initial commit 21:48:00Z, zero results/ objects, the tenth bundle-less repo, minted ≈6.5 h after the nine. Sleep sequence now 50/125/78/34/31/26.6/70/35.2/16.2/61.9/242/147: burst-K's single above-band mint reads retrospectively as a one-shot, not a restart, the harvest-empty-vs-lane-move question now has a THIRD shape (the lane moves intermittently: rich victim 19:20Z, null victim 21:47Z). The 242 min ceiling stands as a bound but is no longer a due-date heuristic (147 landed well inside it). Null-persistence clocks now ten deep: nine burst-I/J repos hit +24 h ≈14:11Z-15:18Z Oct 17, burst-L's own +24 h ≈21:47Z Oct 17. Monitor: burst-M sleep + whether burst-L ever backfills (a backfill now would rewrite the strict-single-use rule and re-open the partial-lag read wholesale). Full sweep context → 117th stream section.

October 17 one-hundred-and-eighteenth sweep follow-up: BURST-M NOT YET DUE AT READ, sleep clock ≈232 min vs the 242 min ceiling, due-decision window opens ≈01:50Z; kralizec-navigator-578 backfill-NEGATIVE at +3 h 52 m (full recursive tree still ONE README.md, zero results/) = the tenth bundle-less repo HOLDS null, null-class now ten deep across ≈11.5 h

blossom91 public_repos 377 EXACTLY (unchanged since burst-L), Oct-16 cohort closes at 83 (calendar day flipped, Oct-17 mints so far: zero). Sleep sequence 50/125/78/34/31/26.6/70/35.2/16.2/61.9/242/147, burst-M is the second-longest live sleep on the board. Read discipline carried from the 116th: if burst-M overshoots 242, the ceiling becomes the floor of a widening distribution and stops being a due-date heuristic. Null-persistence clocks: the nine burst-I/J repos hit +24 h Oct 17 ≈14:11Z-15:18Z, burst-L's own +24 h ≈21:47Z. Monitor: burst-L backfill-or-not (strict-single-use stands until it doesn't); burst-L lands FIRST results-bearing mint re-opens harvest-empty vs lane-move on a fresh datapoint; whether an Oct-17 cohort opens at baseline gap-from-midnight or continues the long sleep. Full sweep context → 118th stream section.

October 17 one-hundred-and-nineteenth sweep follow-up: BURST-M LANDS, AND LANDS RICH(SUB-BAND): fremen-futar-460 minted 01:38:49Z (sleep ≈231 min, INSIDE the 242 min ceiling for the third consecutive burst) carrying results/results-1792201132693-0.json 3,869 B in +4 s, the FIRST Oct-17 mint, cohort 1, public_repos 377→378 EXACTLY, the collection lane is INTERMITTENT per-mint, not state-shifted: rich 19:20Z (40,245 B) → null 21:47Z → sub-band 01:38Z

blossom91 page read ≈03:27Z: top repo fremen-futar-460 created 01:38:49Z, commits Initial commit 01:38:50Z + Add files. 01:38:53Z, tree README.md (48 B, standard Shai-Hulud: Here We Go Again stub) + results/results-1792201132693-0.json 3,869 B, the bundle filename epoch 01:38:52.693Z puts mint→results at ≈3.7 s, the TIGHTEST mint→results on the ledger (prior floor +6.2 s). Sleep sequence now 50/125/78/34/31/26.6/70/35.2/16.2/61.9/242/147/231, the 242 min ceiling holds as a bound for the third consecutive long sleep (242/147/231), and the due-decision heuristic the 118th suspended stays suspended: the distribution, not the ceiling, is the object. 3,869 B is BELOW the sub-band floor the 113th measured (4,137 B) = new single-file floor, continuing the thin-harvest drift; the size sequence since the 14:11Z break reads 5,477/4,137/4,189 → 40,245 → 0 → 3,869, one rich victim, two nulls, three sub-band, which fits HARVEST-VOLATILITY (per-victim yield) better than either pure read: the lane neither died nor moved, the victims vary. The Oct-17 cohort opened 4 h 51 m after midnight, one mint, repos exact 378 (exactly the one mint, no hidden cluster). kralizec-navigator-578 (burst-L) re-read at +5 h 50 m: still ONE README.md, zero results/ = tenth null HOLDS, strict-single-use unbroken, null-class ten deep. Clocks forward: nine burst-I/J nulls +24 h ≈14:11Z-15:18Z today, burst-L +24 h ≈21:47Z today, burst-N next sleep vs the ceiling. Full sweep context → 119th stream section.

October 17 one-hundred-and-twentieth sweep follow-up: BURST-N AND BURST-O LAND BACK-TO-BACK AND BOTH LAND RICH: sayyadina-laza-277 minted 04:43:40Z (sleep ≈185 min from burst-M, INSIDE the 242 ceiling) carrying results/results-1792212226948-0.json 11,541 B; sayyadina-stillsuit-220 minted 05:14:32Z just ≈31 min later carrying results-1792214082665-0.json 11,253 B, the 31 min second gap says the Oct-17 cohort reads as a BURST-PAIR, not two singletons: the intra-burst cadence is back

blossom91 read ≈05:30Z: public_repos 378→380 EXACTLY the two mints, Oct-17 cohort 3. Trees pulled recursive: laza-277 = README.md 50 B + results/results-1792212226948-0.json 11,541 B (Initial commit 04:43:42Z, Add files. 04:43:47Z, bundle epoch 04:43:46.948Z = mint→results ≈+6.9 s); stillsuit-220 = README.md 55 B + results/results-1792214082665-0.json 11,253 B (Add files. 05:14:44Z, epoch 05:14:42.665Z = ≈+10.7 s). Both sizes sit in the day-four 10-12 KB single-file band, ABOVE the sub-band the 119th measured, so the full post-break sequence reads 5,477/4,137/4,189 → 40,245 → 0 → 3,869 → 11,541 → 11,253: rich, null, sub-band, rich, rich. That is FIVE draws spanning 0 to 40 KB from one lane in ≈12 h, which is what HARVEST-VOLATILITY looks like when you keep sampling, the harvest-empty-vs-lane-move question stays closed on the volatility read, and neither of the two single-lane theories has re-opened. Sleep sequence now 50/125/78/34/31/26.6/70/35.2/16.2/61.9/242/147/231/185; the NEW signal is not the 185 but the 31: burst-N→burst-O is the first ≤35 min inter-mint gap since the Oct-16 spread clusters, i.e. the sleep→cluster duty cycle appears to be RE-ENGAGED, if burst-P lands inside ≈35 min of burst-O the cohort reads as a cluster in progress, not a third singleton. kralizec-navigator-578 (burst-L) re-read at +7 h 42 m: still ONE README.md, zero results/ = tenth null HOLDS, strict-single-use unbroken across FOUR subsequent result-bearing mints. Clocks forward: nine burst-I/J nulls +24 h ≈14:11Z-15:18Z today, burst-L +24 h ≈21:47Z today, burst-P due-decision from ≈05:45Z. Full sweep context → 120th stream section.

October 17 one-hundred-and-twenty-first sweep follow-up: BURST-P AND BURST-Q LAND AT AN IDENTICAL ≈38 MIN CADENCE, sayyadina-sandworm-707 05:52:28Z (4,193 B in +6.4 s) + tleilaxu-ornithopter-873 06:30:03Z (11,305 B in +10.2 s), the 120th's cluster condition is MET TWICE: the Oct-17 cohort is a RUNNING CLUSTER on a fixed ≈38 min intra-mint clock, repos 380→382 exact, cohort 5

blossom91 read ≈07:26Z: public_repos 380→382 EXACTLY the two mints. Trees recursive: sandworm-707 = README.md 54 B + results/results-1792216354396-0.json 4,193 B (Initial commit 05:52:30Z, Add files. 05:52:34Z, epoch 05:52:34.396Z = ≈+6.4 s); ornithopter-873 = README.md 56 B + results-1792218613225-0.json 11,305 B (Initial commit 06:30:04Z, Add files. 06:30:14Z, epoch 06:30:13.225Z = ≈+10.2 s). The decisive measurement is the PAIR OF GAPS: 05:14:32Z → 05:52:28Z = 37.9 min, 05:52:28Z → 06:30:03Z = 37.6 min, two consecutive intra-mint intervals inside 20 s of each other. After the 120th read the 31 min N→O gap as a cadence returning; P and Q make it a CLOCK. The Oct-16 spread-cluster grammar (intra-cluster gaps tighter and steadier than inter-burst sleep) has its first sustained repeat since the 14:11Z break, four result-bearing mints 04:43Z→06:30Z, every one RICH, no nulls inside the cluster. Size draws 11,541 / 11,253 / 4,193 / 11,305: the 4,193 B sits near the post-break sub-band floor while the rest hold the 10-12 KB day-four band, volatility inside a cluster, not a cluster-level mode. The cluster's own due-question: burst-R on the 38 min clock lands ≈07:08Z; at the ≈07:26Z read the repo list still ENDS at ornithopter-873 = +18 min past cadence, NOT landed, first loosening of the clock (or first late mint); a third consecutive ≈38 min interval would date the pipeline to the minute, silence past ≈1 h reads as the cluster sleeping out to the 185-242 inter-burst distribution. kralizec-navigator-578 re-read +9 h 42 m: ONE README.md, zero results/ = tenth null HOLDS, strict-single-use unbroken across SIX result-bearing mints. Clocks forward: burst-R loosening vs sleep-out next sweep; nine burst-I/J nulls +24 h ≈14:11Z-15:18Z today, burst-L +24 h ≈21:47Z today. Full sweep context → 121st stream section.

October 17 one-hundred-and-twenty-second sweep follow-up: THE 38-MINUTE CLOCK DOESN'T SLEEP OUT, IT RESHAPEs: burst-R lands +24 min LATE at 07:32:05Z and five mints run 07:32:05Z→07:50:32Z at 1.3/2.1/4.7/10.4 min gaps = TIGHTER than the clock it replaced, every one RICH; 154 min sleep, then the tightest pair since burst-F's same-second twins: fremen-heighliner-582 10:24:46Z + atreides-ghola-894 10:25:19Z = 33 s apart, both RICH, Oct-17 is 12-for-12 RICH, repos 382→389 EXACT

blossom91 read ≈11:3xZ: public_repos 382→389 EXACTLY the seven new mints, Oct-17 cohort 12. Trees recursive, all seven carry results/: harkonnen-sietch-546 12,869 B (+6.3 s), kralizec-phibian-866 12,309 B (+6.4 s), kanly-cogitor-276 12,917 B (+6.5 s), fremen-sandworm-264 4,677 B (+6.7 s), fremen-stillsuit-320 4,749 B (+5.6 s), fremen-heighliner-582 4,721 B (+6.3 s), atreides-ghola-894 11,541 B (+6.4 s). Three reads resolve the 121st's open question: (a) the loosening was a reshape, not a sleep-out, the ≈38 min fixed cadence broke after ornithopter-873, ran 62 min to the next mint, then ran FIVE mints at gaps 1.3-10.4 min, i.e. intra-cluster gaps an order of magnitude tighter than the clock it replaced, the cluster grammar holds, the specific number does not; quote the Oct-17 shape as burst → variable-gap cluster, not a fixed clock; (b) 12-for-12 RICH, every Oct-17 mint from fremen-futar-460 01:38:49Z through atreides-ghola-894 carries a bundle, the null class has NOT re-opened in ≈10 h of minting; (c) the 154 min pre-pair sleep sits inside the widening inter-burst distribution (242/147/231/185/62/154, no longer quoting due-dates). Size draws inside one cluster: three above-band (12,309-12,917 B, top of the 10-12 KB day-four band ever measured), then THREE sub-band (4,677/4,749/4,721), then 11,541, note atreides-ghola-894's 11,541 B is the EXACT byte-count of sayyadina-laza-277 six hours earlier (recorded as repetition, meaning unasserted). mint→results clock: seven consecutive measurements inside a 5.6-6.7 s window, the tightest sustained replication of the exfil-commit clock on the ledger. Null persistence: kralizec-navigator-578 re-read +13 h 47 m tree still ONE README.md (tenth null holds across TWELVE result-bearing mints); ALL NINE burst-I/J nulls pre-checked ≈17.2-18.7 h post-mint, every tree zero results/ objects, their +24 h checkpoints land ≈14:11Z-15:18Z TODAY with the pre-check negative, burst-L's own +24 h ≈21:47Z tonight. Full sweep context → 122nd stream section.

October 17 one-hundred-and-twenty-third sweep follow-up: THE CLUSTER EXHAUSTS TO WIDE SINGLETONS, ghola-ghola-177 11:32:06Z (10,765 B +6.5 s) + tleilaxu-lasgun-225 13:10:45Z (11,309 B +10.8 s) at 66.8/98.7 min gaps = ordinary singleton spacing after the five-deep cluster; Oct-17 14-for-14 RICH, repos 389→391 EXACT, the day CYCLES shapes: singletons → ≈38 min clock → variable-gap cluster → 33 s pair → wide singletons

blossom91 read ≈13:3xZ: public_repos 389→391 EXACTLY the two new mints, Oct-17 cohort 14. Trees recursive, both RICH: ghola-ghola-177 = README.md 47 B + results/results-1792236732553-0.json 10,765 B (Initial commit 11:32:08Z, Add files. 11:32:13Z, bundle epoch 11:32:12.553Z = +6.5 s); tleilaxu-lasgun-225 = README.md 51 B + results/results-1792242655846-0.json 11,309 B (Initial commit 13:10:46Z, Add files. 13:10:57Z, epoch 13:10:55.846Z = +10.8 s). The decisive reads are the GAPS since the 33 s pair: 10:25:19Z → 11:32:06Z = 66.8 min, 11:32:06Z → 13:10:45Z = 98.7 min, neither the dead ≈38 min clock nor cluster grammar; the lane returns to wide singleton spacing after a burst exhausts, so the Oct-17 quote becomes SHAPE-CYCLING (segment fits, not one cadence). Sizes sit back inside the day's 10-12 KB band; tleilaxu-lasgun-225 11,309 B is 4 B from ornithopter-873's 11,305 B (near-repeat recorded, meaning unasserted). Name-grammar first: ghola-ghola-177 = first same-word doubling in the token set (recorded, pattern unasserted). Null class: all nine burst-I/J trees re-pulled ≈22.3-23.4 h post-mint, every one STILL single README.md zero results/ = their +24 h checkpoints open NEGATIVE ≈14:11Z-15:18Z; kralizec-navigator-578 holds +15 h 45 m, +24 h ≈21:47Z tonight. Full sweep context → 123rd stream section.

October 17 one-hundred-and-twenty-fourth sweep follow-up: THE NULL CLASS PASSES ITS +24 h FINAL, NINE-FOR-NINE, all nine bundle-less burst-I/J repos re-pulled at ≈24.3-25.4 h post-mint, every one STILL a single README.md (46-54 B), ZERO results/ objects, zero backfill = the strict-single-use rule now stands across a full solar day = the 14:11Z break is an OPERATIONAL EVENT with the ledger's first formally closed null cohort; kralizec-navigator-578 (burst-L) holds +17 h 45 m, its own +24 h ≈21:48Z tonight closes the class ten-for-ten; public_repos 391 UNCHANGED ≈2.3 h after tleilaxu-lasgun-225, wide-singleton spacing, no cluster re-engagement

Null-cohort final table (tree size, total bytes, results?): kralizec-stillsuit-57 1/53/·, kanly-phibian-113 1/49/·, fremen-ghola-526 1/48/·, prana-lasgun-284 1/48/·, prana-lasgun-93 1/47/·, gesserit-stillsuit-256 1/54/·, harkonnen-fedaykin-180 1/54/·, atreides-cogitor-544 1/52/·, ghola-slig-221 1/46/·, nine-for-nine README-only at past-24 h. The three burst-I/J survivors (powindah-ghola-267 5,477 B, siridar-lasgun-895 4,137 B, mentat-melange-503 4,189 B) unchanged = the split pattern (twins one-armed-one-empty, clusters nine-null-three-rich) is now a day-old stable fact, not a transient. Quote for downstream consumers: within one lane, a mint either commits its harvest within ≈+11 s or never, nulls are permanent, and ≈75% of the Oct-16 14:11Z+ mints chose never. The lane's cadence Oct-17 afternoon: last mint tleilaxu-lasgun-225 13:10:45Z, ≈2.3 h quiet at the ≈15:3xZ read, inside the day's own wide-singleton band (66.8/98.7 min gaps), NOT yet a new sleep regime; burst-L's ten-repo null close at ≈21:48Z is the next scheduled experiment. Full sweep context → 124th stream section.

October 17 one-hundred-and-twenty-fifth sweep follow-up: THE TENTH NULL HOLDS AT +19 h 45 m AND THE LANE GOES WIDE, kralizec-navigator-578 re-pulled recursive = single README.md 54 B, zero results/; public_repos 391 UNCHANGED ≈4.4 h after tleilaxu-lasgun-225 = no mint beyond the day's widest singleton gap (98.7 min) yet inside the sleep-ceiling lineage; the +24 h checkpoint ≈21:48Z tonight is the whole class's final exam, ten-for-ten closes strict single-use as ledger law, one backfill re-opens the operational-event read whole; census top-20 unchanged, burst-I/J survivors 5,477/4,137/4,189 B stable into their second day. Sweep context → 125th stream section.

October 17 one-hundred-and-twenty-sixth sweep follow-up: THE TENTH NULL PASSES ITS +24 h FINAL, TEN-FOR-TEN, STRICT SINGLE-USE IS LEDGER LAW, kralizec-navigator-578 re-pulled recursive at +21 h 40 m+ post-mint = ONE README.md 54 B (sha 56f697fe…), ZERO results/, zero backfill; joined to the nine burst-I/J repos the 124th closed at ≈24.3-25.4 h, the FULL null class now stands ten-for-ten at ≈22-25.4 h depth = the 14:11Z break is an OPERATIONAL EVENT with the ledger's first COMPLETE closed null cohort; public_repos 391 UNCHANGED ≈6.3 h after tleilaxu-lasgun-225 = the lane's longest tracked quiet (beyond the 98.7 min widest singleton AND past the 242 min sleep-ceiling lineage entirely, sleep-out-vs-reshape now needs a new fork: relocation); account events feed shows nothing after Oct-16 15:18Z Creates, census top-20 unchanged, burst-I/J survivors 5,477/4,137/4,189 B stable into day three. Sweep context → 126th stream section.

October 17 one-hundred-and-twenty-seventh sweep follow-up: THE 432-MINUTE QUIET RESUMES, NOT RELOCATES, THREE new mints after the ≈7.2 h silence: gesserit-sandworm-477 20:23:15Z (11,313 B +12 s), sardaukar-phibian-443 22:26:00Z (11,233 B +11 s), tleilaxu-ghola-379 23:15:37Z (12,897 B +7 s); inter-mint gaps 432.5→122.75→49.6 min = a TIGHTENING resumption ladder, cluster-forming if the next mint lands inside ≈50 min; repos 391→394 EXACT, Oct-17 closes 17-for-17 RICH, the 242 min sleep ceiling dies as a terminal bound: the lane slept 432.5 min and woke, quote it as ≈242 min typical with a proven ≈432 min outlier; the relocation fork the 126th opened CLOSES, same account, same grammar, no new infrastructure signal; tenth null kralizec-navigator-578 re-verified at +25 h 50 m STILL README-only = nulls are terminal, not delayed, the closed class held ten-for-ten across three fresh flowing mints. Sweep context → 127th stream section.

October 18 one-hundred-and-twenty-eighth sweep follow-up: THE CLUSTER-FORMATION TEST PASSES AT THE FIRST CHECK, siridar-cogitor-50 minted 00:03:37Z exactly 48.0 min after tleilaxu-ghola-379, inside the 127th's ≈50 min prediction window (≈00:05Z, landed 1.5 min early); the resumption ladder 432.5→122.75→49.6→48.0 min FLATTENS at ≈48 min = the woken lane re-settled onto a fixed short clock, the fourth cadence shape this week; repos 394→395 EXACT, resumption run 4-for-4 RICH, Oct-18 opens RICH; results 11,337 B +11 s; cogitor suffix returns (last kanly-cogitor-276 Oct-17). Tenth null kralizec-navigator-578 ≈+27 h 45 m STILL README-only = terminal class holds across four fresh flowing mints. Sweep context → 128th stream section.

October 18 one-hundred-and-twenty-ninth sweep follow-up: THE ≈48-MIN CLOCK BREAKS AT ITS FIRST RE-CHECK, NOTHING MINTED ≈212 MIN AFTER siridar-cogitor-50, repos 395 UNCHANGED, but 212 min sits INSIDE the 26.6-242 min inter-burst sleep distribution, so the woken lane fell back to the distribution instead of extending the clock: the fixed-clock read was a segment fit (fifth reshaping this week, shape-cycling stands); tenth null kralizec-navigator-578 ≈+29 h 48 m STILL README-only; an overshoot past the ≈432 min proven outlier would re-open the relocation question Sweep context → 129th stream section.

October 18 one-hundred-and-thirtieth sweep follow-up: THE MINTS LANE RESUMES AT 290 MIN AND THE README CARRIES THE BRAND, harkonnen-heighliner-293 minted 04:53:41Z (290 min after siridar-cogitor-50), README 56 B = # harkonnen-heighliner-293 + Shai-Hulud: Here We Go Again, the SAME brand-line the tensorlake hostage-worm used to describe its victim-created exfil repos; results 7,137 B committed +3.0 s (fastest tracked), below the resumption run's 11.2-12.9 KB band; repos 395→396 EXACT; the sleep overshoots the 242 min typical ceiling but stays under the 432 min proven outlier = band re-quoted ≈26.6-432 min observed; tenth null ≈+31 h 50 m STILL README-only; branding remains lineage-claim-not-operator-proof, but a brand-tagged wake is the strongest single naming artifact the mint lane has ever produced Sweep context → 130th stream section.

October 18 one-hundred-and-thirty-first sweep follow-up: THE BRANDED LANE CLUSTERS, TEN mints 05:27:25Z→07:19:14Z, a NINE-MINT burst spanning 24.3 min (intra-gaps 0.5-12.3 min) + one singleton 87.5 min later; TEN-for-TEN RICH and TEN-for-TEN BRANDED, every README carries Shai-Hulud: Here We Go Again = the brand line is STANDING, not a one-off (the 130th's monitor question answered in ≈2.5 h); repos 396→406 EXACT; results split into TWO new sub-band modes, ≈5.95-6.2 KB (six) + ≈8.6 KB quartet spread 32 B, both below the 11.2-12.9 KB resumption band; cluster opens 33.7 min after the branded wake, inside the old 26.6-242 band = the 290 min read as segment again, seventh reshaping this week; tenth null ≈+34 h 20 m STILL README-only Sweep context → 131st stream section.

October 18 one-hundred-and-thirty-second sweep follow-up: THE BRANDED LANE SINGLE-STEPS AFTER THE CLUSTER, TWO more mints fedaykin-sietch-483 07:37:03Z (+17.8 min, BELOW the old 26.6 min band floor, the floor erodes) + tleilaxu-melange-100 08:12:18Z (+35.3 min, back inside the band); both RICH, both BRANDED = brand streak TWELVE-for-TWELVE; repos 406→408 EXACT; fedaykin-sietch-483 results 2,937 B = SMALLEST single-file bundle tracked, a new ledger floor under the 3.1 KB Oct-16 multi-file minimum; tleilaxu-melange-100 4,505 B, both below even the 131st's ≈5.95-6.2 KB mode = downward size drift continues, four modes in 24 h; mint→bundle-commit +6.2/+4.6 s single-use two-commits holds Sweep context → 132nd stream section.

October 18 one-hundred-and-thirty-third sweep follow-up: THE LANE OUTPACES THE OBSERVER, TWENTY mints in 78.5 min: a TEN-mint burst spread over 2.0 min (SIX inside ONE second), a FIVE-mint burst over 7 s, and a FOUR-mint burst over 1.9 min; public_repos 408→428 watched CLIMBING LIVE mid-sweep (419 then 424 then 428); brand streak THIRTY-TWO-for-THIRTY-TWO; the null class goes cohort-correlated, TWELVE bundle-less repos, NINE of them from ONE burst = burst-level collector state, not per-victim yield; TWO-file results/ bundles return after two weeks with a size-identical 2,785 B second file on TWO different victims; sardaukar-slig-452 returns to the 11.2-12.9 KB band ending the downward drift Sweep context → 133rd stream section.

October 18 one-hundred-and-thirty-fourth sweep follow-up: BURST-D LANDS 59 s WIDE INSIDE THE BAND AND THE PAIR GRAMMAR RETURNS, TWELVE mints 12:19:33Z→14:53:42Z, repos 428→440 EXACT; brand streak FORTY-FOUR-for-FORTY-FOUR; THE HARVEST ARTIFACT GOES NEAR-DETERMINISTIC, a 5,473/5,485/5,489/5,493 B QUADRUPLE spanning 20 B across FOUR victims, 5,493 B REPEATING sardaukar-slig-677 across sweeps, and a NEW ≈10.9-11.0 KB mode of THREE spanning 96 B, bundle size reads as a collector-BUILD fingerprint candidate, not a victim variable Sweep context → 134th stream section.

October 18 one-hundred-and-thirty-fifth sweep follow-up: A FIVE-MINT BURST CREATED IN ONE EXACT SECOND, ALL FIVE NULL, AND THE SIZES READ AS COLLECTOR RE-TOOLING, burst-F FIVE mints at ≈80 s spacing ALL RICH at a NEW ≈2.82-2.88 KB MODE SPANNING 60 B (new ledger floor under the 2,937 B mark), while the same window's singletons draw the 10.9 KB mode = size class tracks the BURST not the victim, build-fingerprint candidate strengthened; PAIR GRAMMAR TIGHTENS AGAIN at 43 s and goes MIXED (null + rich in one pair), the null/rich split no longer respects pairs; brand streak FIFTY-SEVEN-for-FIFTY-SEVEN, repos 440→453 EXACT; null-cohort correlation holds a THIRD observation (five-of-five from the same-second five) Sweep context → 135th stream section.

October 18 one-hundred-and-thirty-sixth sweep follow-up: THE SIZE MODE SNAPS BACK INSIDE ONE MINT, THREE mints 17:30:39/17:58:04/19:04:47Z ALL RICH at 12,341/12,401/12,565 B (224 B span), back in the 11.2-12.9 KB band 231 s after burst-F's last mint = the ≈2.82-2.88 KB mode was a ONE-BUILD RUN, the third-mode test answers NO, and the post-burst step runs +3.9 min, the tightest outside-burst spacing ever measured; brand streak SIXTY-for-SIXTY, repos 453→456 EXACT; the same-second quintet holds ALL-NULL at ≈3.5 h, zero backfill The build-fingerprint candidate refines: build identity is burst-scoped AND short-lived (a mode flip on a single-step boundary), not victim-scoped and not day-scoped. Bundle epochs +6.4-+7.3 s single-use two-commits holds. Sweep context → 136th stream section.

October 18 one-hundred-and-thirty-seventh sweep follow-up: THE LOW MODE RETURNS, THE 2.82-2.88 KB CLASS RECURS WITH TWO EXACT BYTE-SIZE REPEATS OF BURST-F, the ONE-BUILD RUN answer was OVERCLAIMED, THREE-mint mini-burst 19:54:44/19:57:26/19:59:22Z at 2,821/2,841/2,825 B (20 B span) one mint after the 12.4 KB band-return; 2,821 B = ghola-sietch-901's exact value, 2,825 B = fremen-laza-452's exact value; then singleton fremen-slig-440 21:18:55Z 10,973 B = SEVENTH hit of the 10.9-11.0 KB mode**

The 136th resolved the third-mode test as a one-build run because the mint right after burst-F drew the 12.4 KB band. One read later the lane drew the 2.8 KB band back, with byte-exact values repeated from burst-F itself. Durability, recorded-not-asserted: harvest sizes come from a small DISCRETE SET revisited per cluster; the mode is burst-scoped and RE-USABLE, not one-shot. The build-fingerprint candidate strengthens as a discrete-artifact-set model, byte-exact repetition across separated bursts now measured three independent ways (5,493 B sweep-repeat, 2,785 B twin second-file, this sweep's two burst-F exact repeats). Cadence: +49.9 min to the mini-burst (inside band), intra-gaps 162/116 s (looser than burst-F's 78-84 s), then +79.6 min singleton fremen-slig-440 10,973 B (seventh hit of the 10.9-11.0 KB mode). THE LANE THEN KEPT MINTING INSIDE THE SAME READ WINDOW: SIX more mints 21:51-23:15Z, a 76 s pair atreides-sietch-338 2,833 B + kralizec-phibian-779 2,809 B (2,809 extends the low class 12 B below the 2,821 floor), a 84 s pair prescient-stillsuit-410 5,961 B + siridar-ghola-185 7,629 B, then kralizec-laza-113 11,577 B + gesserit-ghola-651 12,845 B = an ASCENDING SIZE LADDER visiting FIVE distinct size classes in 84 min, one per step, the discrete-set read's strongest form yet (recorded-not-asserted). Window totals: TEN-for-TEN RICH, TEN-for-TEN BRANDED = streak SEVENTY-for-SEVENTY; repos 456→460→466 EXACT; two-commits holds, epochs +3.9-+11 s. Sweep context → 137th stream section.

October 19 one-hundred-and-thirty-eighth sweep follow-up: MINT SEVENTY-ONE, SINGLE-STEP CADENCE CONTINUES, sardaukar-ghola-928 01:38:56Z +143.5 min after the ladder's top, RICH at 11,537 B back inside the 11.2-12.9 KB band, streak SEVENTY-ONE-for-SEVENTY-ONE, repos 466→467 EXACT, epochs +11 s two-commits holds. The discrete-set read draws no new class: the seventh value drawn since the ladder, band-class again. The lane has now single-stepped since burst-F (+3.9 → +79.6 → … → +143.5 min), cadence inside the band, no cluster-forming pressure measured. Sweep context → 138th stream section.

October 19 one-hundred-and-thirty-ninth sweep follow-up: THE LANE EXPLODES AND THE CEILING OPENS TEN-FOLD, FIFTY mints 02:38-08:03Z, repos 467→511→517 EXACT, FOUR victims at a NEW ≈118 KB CLASS spanning 52 B (118,141/118,185/118,189/118,193 B) WITH TWO-FILE results/ BUNDLES BACK AFTER A ONE-DAY ABSENCE, AND THREE MORE BYTE-EXACT REPEATS: 2,937 B (= fedaykin-sietch-483's floor mark at sayyadina-laza-55), 4,505 B (= tleilaxu-melange-100 at kralizec-sietch-976), 5,489 B (= the burst-D quadruple's value at powindah-phibian-926), burst-G ELEVEN mints over 5.2 min (FIVE inside TWO SECONDS), a FIVE-mint 3 s cluster at 03:45:37-40Z, burst-H SEVENTEEN over 48 min, burst-I SIX-and-open, and THE LARGEST ZERO-YIELD WINDOW ON THE LEDGER: all SEVENTEEN burst-H mints README-less AND bundle-less at +51 min-plus (first README-less cohort; the collector produced NOTHING for 48 min), while burst-G's tail SIX are branded-README nulls. Discrete-set model: byte-exact repetition measured SIX ways, ceiling re-opens at ≈123 KB two-file totals. Sweep context → 139th stream section.

October 19 one-hundred-and-fortieth sweep: THE NULL CLASS IS OBSERVER-TIMING-DEPENDENT, BURST-H REVIVED AT THE TREE READ: the 139th's SEVENTEEN README-less AND bundle-less zero-yield window RE-SCORES to a YIELD-LATENCY window, ten direct tree re-reads found ALL ten with branded READMEs + results/ bundles committed ≈+4 s post-creation, while the SAME sweep's twelve fresh mints ran TWELVE-for-TWELVE RICH and prescient-slig-299 149,185 B + 3,325 B SMASHED the one-day-old 118 KB ceiling by 31 KB; gesserit-laza-777/harkonnen-melange-804 carry byte-identical 44,469 B first files = twin-file grammar at the new class, DURABLE: pagination size 0 is NOT tree state; a null enters the ledger only after a DIRECT TREE READ. Sweep context → 140th stream section.