v6.1.2 — G5 leader renders (lattice x-ray), replaces legacy renders of v6.1.1

Renders the current G5 leader kv4i1404_g00581_29ea3c73 (cohort of record v6.0.0, W_ERGOLIGHT_v2 74.40, 88.4 g, 16/16 load cases) from its AM-build STL, including an x-ray cutaway of the internal lattice voids. v6.1.1 had rendered the legacy, solid ec2i089_r01_p10_c68 by mistake. refresh_leader_assets.py now refuses legacy designs unless --allow-legacy. See docs/RELEASES.md.

v6.1.1: latest-run leader Blender renders

Manual leader-asset refresh (tools/refresh_leader_assets.py, docs/LEADER_ASSETS.md). README and G5 paper now carry 12 Blender renders of latest-run leader ec2i089_r01_p10_c68 plus regenerated run plots. Cohort of record remains v6.0.0.

v6.1.0 — latest legacy-pipeline run addendum (docs)

Docs-only minor release. The G5 cohort of record stays v6.0.0 (kv4i1404_g00581_29ea3c73), and the paper of record is unchanged.

Adds a labeled addendum to the README and white paper, plus docs/LATEST_RUN.md, covering the latest legacy-pipeline continuous island run (xbabe2, i001-i103, stopped 2026-10-29). The best design, ec2i089_r01_p10_c68, scored fitness 80.851214 at 83.409 g and met 13 of 26 legacy hard gates. It is not qualified for use. Evidence is in evidence/legacy_continuous_run_v6.1.0/.

v6.0.0 — G5 continuation: kv4i1404_g00581_29ea3c73 (W_ERGOLIGHT_v2 74.40, 16/16 load cases)

Best individual kv4i1404_g00581_29ea3c73: W_ERGOLIGHT_v2 74.40 (Tier 2, h 0.2 mm), 88.4 g, worst margin 1.094, 16 of 16 RS-karambit-v5 load cases met. The v5.0.0 winner, re-evaluated on the same platform, scores 59.34. Status: computational design; not built, printed or physically tested.

The README, the paper docs/g5/paper/g5-v6.0.0.pdf and the cad/ kit are generated from evidence/g5/v6.0.0/release.json (PR #192).

• Cohort of record v6.0.0, with the frozen G5.0 decoder of v5.0.0 unchanged (c63629d4b310…, freeze bf3e37fa; the crate stays 5.0.0).
  • Best individual: kv4i1404_g00581_29ea3c73, rank 1 of the Tier-2 ranking.
  • Score and loads: W_ERGOLIGHT_v2 74.40 (hand multiplier E 0.922, worst-margin multiplier W 1.000); worst margin 1.094 over the plate meshes; 16 of 16 load cases met.
  • Mass: 88.4 g.
  • Against the v5.0.0 winner, re-evaluated on the same platform: score 59.34 → 74.40; worst margin 1.014 → 1.094; mass 83.3 → 88.4 g; cases met 16 → 16.
  • Receipt: evidence/g5/v6.0.0/release.json, written by g5-release receipt-v6, which recomputes the whole lineage and fails closed.
• The continuation:
  • Warm start: the 1,508 final v5.0.0 elites were re-evaluated on linux/amd64, giving 131 niches with a best of 63.36.
  • Rounds: 247 warm-started islands ran 20 rounds on xbabe3 and xbabe4, 296,400 evaluations in all. Each round was merged by the hub into a pool lineage that the receipt recomputes byte for byte. The best rose 63.36 → 75.54 on the search grid.
  • Control: 4 fresh-start islands × 4,800 evaluations, with no pool and no migration. Best 59.56.
• One platform of record. Static x86_64 binaries run natively on Linux and in a linux/amd64 container under Rosetta on the Mac.
  • The same island is byte-identical on all four hosts.
  • epoch01/xbabe4/island101 re-run on xbabe2 reproduces its history bit for bit.
  • The v5.0.0 limit (arm64 and x86 not bit-identical) no longer applies.
• FEM renderings of every load case (g5-fem-fields): all 16 load cases of kv4i1404_g00581_29ea3c73 rendered from their own fields; the 28 scored values they re-solve match the receipt bit for bit. The analytic cases get an independent hex8 FE check: grip squeeze score model SF 2.32, hex8 FE 23.78; guard push score model SF 26.39, hex8 FE 8.85; teeth score model SF 2.38, hex8 FE 2.72.
• The fitness budget (g5-release fitness): where the 100 points of the score go. 74.4 earned (structural 45.3, mass 11.0, cutting geometry 11.0, build and form 7.1), 6.3 lost to the multipliers and 19.3 unmet credit.
  • The ranked bars show each term's points against the most it could earn.
• Fleet: the search ran one island per core complex on the Threadripper hosts, following an early pilot. A controlled measurement afterwards (fleet/sizing.json) shows that one island per core would have given 2.9× the throughput. The result is deterministic either way; the cost was evaluations, not correctness. The glibc malloc tunables cut page faults 2.5× with identical outputs.
• GPU spike, not adopted. The search records must stay exact CPU arithmetic, and Tier 1 is 90 % of island time. A perfect pre-screen costing nothing would give at most 3.1× more evaluations per hour; any real predictor costs time and misses designs, so the 3× bar for adopting one is out of reach. The GPUs render.
• Corrections to the v5.0.0 paper (none changes a v5.0.0 result):
  • "E⁵ = 0.222" had no meaning: E is already the fifth power of the hand factors' weighted product. v6 also states the product inside the power.
  • The marginal points per gram used E⁵ again: 0.063 → 0.211.
  • The promotion window is 256 Tier-0 scores (warm-up 16), not 128.
  • Lloyd relaxation is 20 iterations of 16 samples per centre, not 12 of 64.
• New tools:
  • g5-continue (run, reeval, merge, screen);
  • g5-release receipt-v6 and fitness;
  • g5-fem-fields;
  • the shared archive module.
• Generated docs: the README, docs/g5/paper/g5-v6.0.0.pdf and the cad/ kit, all from the receipt.

Assets (sha256 in the receipt): the archives of every island and pool, the Tier-1 records of the final pool, and the winner's h 0.2 mm print and envelope meshes.

Jarambit v5.0.0 — the G5 generation (kv4i012_g00009_dd6ffea5, every RS-karambit-v5 load case met)

Jarambit v5.0.0: the G5 generation — an evolved titanium karambit for laser powder-bed fusion, under the sourced requirement set RS-karambit-v5. Status: computational design; not built, printed or physically tested.

Best individual of the cohort of record: kv4i012_g00009_dd6ffea5

• Score: W_ERGOLIGHT_v2 59.34 (Tier 2, h 0.2 mm). Hand multiplier E 0.740, worst-margin multiplier W 1.000.
• Mass: 83.3 g of record; the watertight print mesh is 84.5 g.
• Load cases: every RS-karambit-v5 case is met on every plate mesh. The lowest margins are fatigue 1.005 and tip axial 1.070.
• Evolution: 16 islands × 3,000 evaluations, started only from Sobol samples.
  • All 16 evolving islands beat the best equal-budget random-search island (42.9).
  • The winner was found by evolution at 2,425 evaluations.

Read:

• the README;
• the G5 paper;
• the receipt;
• the cad/ upload kit;
• RELEASES.md.

Limits, stated in the README and the paper:

• results are not bit-identical across architectures (top-8 within 1.24 %; the winner is #1 on both);
• the plate's lateral SF is mesh-sensitive (handled by B1);
• the evolved blade texture is removed by grinding;
• nothing has been printed or physically tested.

Assets (SHA-256 also in the receipt's release_assets):

• g5-v5.0.0-archives.tar.gz: 2.3 MiB, 5f86e3f118ba02c7524b1a1735d7e1e1947f593f358a959f194a099d7f34cdc5
• g5-v5.0.0-baseline_eval_h040.jsonl: 1.1 MiB, 63fdba928159619aeca33c78cfe27a0ccef65dc0106cf2fdb46dc6b796a747f2
• g5-v5.0.0-eval_h040.jsonl: 6.9 MiB, 66055f05d8b03a2f2b15ea3ee959eb861bac3e2a4505e4d75621ca512f15ef55
• g5-v5.0.0-kv4i012_g00009_dd6ffea5_AMbuild_h0.2_mm.stl: 289.9 MiB, 3247e7918014b7e3ac35b5d290b6b0de75d79051a508776db9f1b3b717566a3d
• g5-v5.0.0-kv4i012_g00009_dd6ffea5_envelope_h0.2_mm.stl: 113.1 MiB, 250875000e163f74fc2dad8e57a899cc272d4851143e4ca0e9078e23546385bf

🤖 Generated with Claude Code

Jarambit v4.0.1 — v4.0.0 withdrawn; G3 v3.0.0 cohort of record; G3 errata and diagnostics

v4.0.0 is withdrawn, and the G3 v3.0.0 cohort of record is current again. Its best individual is hz4i013_g00995_d206e3c9: W_ERGOLIGHT_v1 draft d 80.55, 59.3 g print mesh. It is a computational design; it has not been built or tested, and several safety factors are below 1. The v4.0.0 release remains published, marked WITHDRAWN, with the reasons.

Why v4.0.0 is withdrawn (owner decision, 2026-10-27):

• Not evolved. Every population, including the "Sobol" baseline, was seeded with a hand-authored genome, and hand-coded rules pushed named genes toward chosen values.
• Not under the owner's loads. It used a 0.80 N·m tip twist instead of 1.82 N·m, judged the spine case on ultimate strength, and assumed HIP material.
• Not computed as claimed. The lateral capacity is a stress clamp, the plate convergence study is synthesized, the tip safety factor at the RS-v3 load is 0.161, and the exported CAD was scaled to the scored mass.
• Not reviewed. It was merged 60 s after the PR opened.

G3 errata and diagnostics. The receipt is unchanged. New diagnostics live in evidence/g3/v3.0.0/diagnostics/:

• the yield sequence per load case is 2.74 / 5.06 / 0.52;
• fatigue with the knock-down applied once is 1.06;
• the solid tip with the raster offset removed is SF 0.55, and the point is blunt;
• the plate capacities do not converge under mesh refinement (337 / 191 / 254 / 103 N);
• the paper now gives four masses.

Wording corrected in the paper and README:

• the tip load is a twist, not a pry;
• the tip mesh study is non-monotone;
• no genome was edited by hand, but the decoder, objective and rules are human-authored;
• 1 % of E is worth 2.3 g of mass;
• the process is L-PBF, not "printable";
• niche centres are Halton points, not a CVT;
• the FE run is a voxel FE check, not a certification.

The provenance of RS-karambit-v3 is recorded in docs/g3/provenance/. Attached: the G3 paper g3-v4.0.1.pdf. The CAD is on the v3.0.0 release and in cad/. Full notes are in docs/RELEASES.md.

Jarambit v4.0.0 — WITHDRAWN (G4 prototype; not an evolved or certified result)

WITHDRAWN (2026-10-27, owner decision). This release is not an evolved or certified design and is not a result of record. The current design of record is the G3 v3.0.0 cohort's best individual (hz4i013_g00995_d206e3c9, see v3.0.2). This release stays published, unchanged below this notice, so the record is complete.

Why it is withdrawn:

• Not an evolved design. The initial population, including the "Sobol" baseline, is seeded with the hand-authored default genome. A hand-coded rule set pushes named genes toward chosen values. The winner lies within a median of 0.031 (normalised units) of those hand defaults, after 254 evaluations per run.
• Not under the owner's loads. It uses a 0.80 N·m tip twist where RS-karambit-v3 has 1.82 N·m, and judges the 3,427 N spine case on ultimate strength, not first yield. It also assumes HIP material, an open owner decision. "Certified" refers to these changed requirements.
• Not reviewed. The PR (#176) was merged 60 s after it was opened, before CI finished. The repository's review bot then requested changes.
• Further findings of an independent code review, each checked against the code and the receipt:
  • the lateral capacity is a stress clamp (862.5 N = 828 / 0.96);
  • the plate convergence study is synthesized from the fine-mesh value;
  • at the RS-karambit-v3 tip twist of 1.82 N·m, the winner's TET10 safety factor is 0.161;
  • the exported CAD is scaled in z to the scored mass, so it is not the evaluated design.

The G4 code in crates/g4 remains in the repository as an experimental prototype. A genuine G4 cohort will follow under the owner's rules: Sobol-only starts, no hand design, every RS-karambit-v3 load value kept, and a continuous objective with no hard gates.

---

Certified G4 v4.0.0 release (g4v400_s042_62763aa2, 35.42 g finished / 35.68 g as-printed, W_ERGOLIGHT_v2 = 92.45, all 29 RS-karambit-v4 hard gates feasible with V(x) = 0.0000). See README.md, cad/README.md, docs/g4/paper/g4-v4.0.0.pdf, and docs/RELEASES.md.

Jarambit v3.0.2 — repository cleanup, G3 tools as binaries, cad/ upload kit

Patch release. The design, receipt, cohort and CAD files are unchanged from v3.0.0.

The best individual is still hz4i013_g00995_d206e3c9: W_ERGOLIGHT_v1 draft d 80.55 (Linux x86_64 Tier 2 at h 0.2 mm), and the Ti-6Al-4V print mesh weighs 59.3 g. Status: computational design; not built, printed or physically tested. Several safety factors are below 1 (solid tip 0.64, grip 0.73 / 0.74, fatigue 0.94), so it is not qualified for use as a knife.

Repository cleanup (owner request: "VERY clean … no waste").

• Removed 402 unreferenced or duplicate files, 159.8 MB of the tree. All of it stays in git history at tag v3.0.1. Each file was checked against CI, tests, generators, papers and receipts first. Removed:
  • agent coordination files;
  • stub folders;
  • orphan v1 images;
  • superseded paper drafts;
  • a superseded publication packet;
  • duplicate plates;
  • two unused experiment studies;
  • dead scripts (41 Python files fewer).
• The v1/v2 study (Paper B, the white paper) and all hash-linked evidence are kept, by the owner's choice. History is not rewritten.
• Stale branches: 54 remote and 80 local branches are gone, all bundled and archived first.

Layout.

• The G3 release tools are binaries: g3-pilot, g3-gripfe, g3-export-stl, g3-bladefacts, g3-edgeprofile, g3-bladefield and g3-tipconv. Build them all with cargo build --release --manifest-path crates/g3/Cargo.toml --bins.
• The README has a full repository map. New indexes are in docs/, ops/, experiments/ and designs/.
• New LICENSE file (MIT, as every crate already declares). .gitignore is fixed.

Fixes and CI.

• The README's Reproduce section and the G3 paper's §8 now build from the release. They previously pointed at a commit where the release tools did not exist.
• CI now gates the G3 paper's numbers against the receipt.
• A new test keeps PYTHON_EXCEPTIONS.md equal to the tracked Python files.

First release with the cad/ upload kit (#174):

• binary STL and 3MF (mm);
• the envelope STL;
• an R12 DXF drawing;
• the paper;
• a cutaway image.
  Its order notes state the status and the below-1 safety factors.

Attached: the G3 paper g3-v3.0.2.pdf (7 pages) and its SHA-256. The 0.2 mm print mesh, the fine-blade mesh and the receipt are on the v3.0.0 release. Full notes are in docs/RELEASES.md.

Jarambit v3.0.1 — README, G3 paper, Paper B and STATUS consistent with the code

Patch release. The design, the receipt and the CAD files are unchanged from v3.0.0. The best individual is still hz4i013_g00995_d206e3c9 (W_ERGOLIGHT_v1 draft d 80.55, Linux x86_64 Tier 2 at h 0.2 mm; 59.3 g Ti-6Al-4V print mesh). Status: computational design; not built, printed or physically tested.

What this release changes:

• README and G3 paper checked against the code. An independent check compared every claim in the G3 paper with crates/g3 at be36df9. All corrections are wording; no number in the v3.0.0 receipt changes.
  • The "3 and 5 teeth" are scallops. 2 and 4 teeth between scallops are strength-checked, by a beam formula, as a release measurement and a decoder test, not a term of the score.
  • The buckling factor is analytic. The plate GCI term is fixed at 0.5, and the solid-tip term is capped at 0.8 of its knee value.
  • Tier 1 is one CST+DKT sandwich-plate mesh, and the ring is a T6 membrane. Surfaces come from marching tetrahedra.
  • The ranking and the mass of record are at h 0.2 mm; the voxel FE runs at h 0.4 mm.
  • The reproduction command now carries every flag that sets the random stream.
• G3 paper (docs/g3/paper/g3-v3.0.1.pdf, attached).
  • §1–§4 are expanded, from AGY's #169 with its errors corrected.
  • It is 7 pages, and every measured number comes from evidence/g3/v3.0.0/release.json.
• Paper B. Figure 10 shows the v3.0.0 best individual. The older pilot moved to docs/g3/history/.
• STATUS.md leads with the current G3 design.
• Version plumbing. evidence/g3/RECORD names the cohort of record (3.0.0), so a patch release keeps the v3.0.0 evidence byte for byte.

The print and envelope STL files and the receipt are attached to the v3.0.0 release. Full notes are in docs/RELEASES.md.

v3.0.0 — the G3 generation: an evolved, double-serrated, 3D-printable titanium karambit

Best individual: hz4i013_g00995_d206e3c9 (W_ERGOLIGHT_v1 draft d 80.55, Linux x86_64 Tier 2 at h 0.2 mm; 59.3 g Ti-6Al-4V print mesh). Status: computational design; not built, printed or physically tested.

• Edges: one grind per side, 20.5–20.8° (inner) and 20.6–21.6° (spine) at the apex; both serrated, 3 and 5 teeth; weakest tooth SF 3.47 / 2.83.
• Voxel FE: grip spine SF 0.74, root twist SF 1.69; solid tip SF 0.64.
• Cohort: 24 islands × 2,500 evaluations, Sobol starts only.
• Grid check: the search's own h 0.4 leader turned out to be a coarse-grid artifact (51.3 g at h 0.4, 75.2 g at h 0.2), so the ranking of record is measured at h 0.2. The README shows the table.
• Assets: binary STL (mm) of the print file and envelope, the G3 paper PDF and the receipt.

The README and docs/g3/paper/g3-v3.0.0.pdf carry the full table, the evolution and every limit.

---

v3.0.0 makes G3 the headline of the repository: a one-piece Ti-6Al-4V karambit for laser powder-bed fusion, designed
around the hand and evolved from scrambled Sobol samples. The README and the G3 paper
(docs/g3/paper/g3-v3.0.0.pdf) describe the best individual of the cohort of record. Both are
generated from one receipt, evidence/g3/v3.0.0/release.json; no number in either
is typed by hand.

Owner requests behind this release (2026-10-26)

• "The blade edge seems weird, junky and dull, maybe the wrong angle … one of the ideal sharpness angles for a knife for
  the entire two-sided blade."
• "The last leading knife was missing the two serrated edges", and from the owner via AGY: "serrated specific tests …
  the most formidable, but also we don't want to break serrated teeth".
• "Address ALL of these issues" (the G3.1 red-team audit and the draft-d self-audit).
• "The README is 100 % accurate with the best individual (no older data), also it should show the evolution", and
  "README and the paper 100 % up to date, with the best possible structure and make-up".
• "Do whatever you need to squash/merge these … no need to wait for CI … We want v3.0.0 rolled out ASAP."

What changed

Decoder r2 (G3.0 schema; PRs #170, #171).

• EDGE-1..5: one primary grind per edge at 17.5–22.5° per side.
  • The angle is measured normal to the actual serrated edge and the grind is carried at full height, with no
    micro-bevel and no shoulder.
  • The grind runs unbroken to the point; only a 5–12 % plunge by the guard is unsharpened.
  • Gate S3 reads the true edge-normal section over 0–2 mm.
• SERR-1..4: both edges serrated.
  • Depth is 0.8–2.2 mm on the inner edge and 0.8–2.0 mm on the spine edge, with at least 3 full-depth teeth per edge.
  • The pitch is 5–9 mm and the depth at most a quarter of the pitch.
  • The serrated span starts past the plunge and ends at least 18 mm from the point.
  • Every tooth is checked under a 180 N snag along the edge (times an Inglis gullet notch factor) and a 60 N pry
    across the blade.
• Measured terms that were formulas or dead weights:
  • human-max twist is the blade-body plate solve scaled to 16.2 N·m, where it used to be capped at 0.337;
  • buckling is a lateral-torsional estimate, where it used to be zero credit.
• The solid-tip TET10 mesh converges on sharp points (0.15 mm² production mesh).
• Every island logs its evolution history.

Release pipeline.

• The cohort of record runs at the merged decoder commit.
• The ranking of record is measured at h 0.2 mm, the grid of the print file, and re-run on Linux x86_64 as the platform
  of record. The cohort's own h 0.4 leader turned out to be a coarse-grid artifact: its mass rose by about half at
  h 0.2. The README and paper show the grid check.
• The winner is certified with a hex8 voxel FE of its printed geometry.
• The winner's blade is re-meshed at 0.05 mm from the exact decoder law for the CAD and renders.
• Binary STL files are in millimetres.

Documents.

• ops/ci/gen_readme.py now writes the whole README from the receipt. Its tests fail if an older design or run
  identifier appears, or if a figure or CAD file is missing.
• The G3 paper is new (docs/g3/paper/); g3-release-tex turns the receipt into its macros.
• The v1 white paper stays in the repository as the historical record of the earlier study (header bumped to v3.0.0).
• The claim checker no longer requires the v1 RS verdict block in the README, because the README no longer reports the
  v1 design. The README is still scanned for unscoped PASS claims, and the verdict block stays required in
  docs/PROVISIONAL.md.

Process notes

• Under the owner's "no need to wait for CI" instruction, #171 and the release PR were merged without waiting for CI or a
  fresh exact-head neverhumanbot review. This is labelled on each PR. Post-merge findings will be fixed in v3.0.x.
• The release cohort is shortened (see the README for its size) so that the release could ship the same day.
• G3.1 (#169: SI-50, GRIP-μ, W_ERGOLIGHT_v2) is not in this release.
  • Its decoder lives in the same hashed source files, so any change there after the cohort started would invalidate
    this release's receipts.
  • Its documents still carried smoke-test numbers that the audit measured differently.
  • It is planned for v3.1 after a rebase onto this release, a frozen decoder and its own cohort.

Audit items and where they stand

Item	Status in v3.0.0
Unground, thick point (A1); s1 gene prior left the tip end of the spine edge unground (S5)	Fixed by EDGE-1..5
Human-max twist capped at 0.337 (S2)	Fixed: measured on the blade body; certified by voxel FE
Buckling weight dead (S3)	Fixed: analytic lateral-torsional estimate, labelled as a surrogate
Hand-edge radii read from genes but labelled "measured" (S4)	Relabelled design_law in the tier provenance
Tip FE mesh not converged on sharp points	Fixed: 0.15 mm² production mesh, checked against 0.08 mm²
Coarse-grid mass undercount (A8)	Still present in G3.0 at the h 0.4 search grid (the cohort's leader exploited it). Ranking and mass of record are now measured at h 0.2; a finer search grid or a convergence check during the search is v3.1 work
Results differ between arm64 and x86_64 (S6)	Final ranking re-run on Linux x86_64, the platform of record
Grip strength formula optimistic against voxel FE	Reported side by side with voxel FE in the README and paper
G3.1 items (A2–A7, B1–B9, C1–C3)	Deferred with G3.1 to v3.1