Jarambit 6.1.2: an evolved titanium karambit for laser powder-bed fusion

Every computed number and picture on this page describes one design, kv4i1404_g00581_29ea3c73, the best individual of the G5 v6.0.0 cohort of record. They are generated from evidence/g5/v6.0.0/release.json, its verification studies and the sourced requirement set requirements/RS-karambit-v5.toml (provenance). Status: computational design; not built, printed or physically tested. The only exception is the clearly labelled legacy-pipeline addendum at the end of the page, which describes an older, non-lattice design line.

Paper: docs/g5/paper/g5-v6.1.2.pdf, generated from the same receipt. 3D printing: the upload kit in cad/.

_{X-ray cutaway of the G5 leader kv4i1404_g00581_29ea3c73 (W_ERGOLIGHT_v2 74.40, 88.4 g): the internal 3D-printed lattice voids (void fraction 0.53 of the envelope). All renders · regenerate with python3 tools/refresh_leader_assets.py.}

The best individual

Titanium (Ti-6Al-4V, L-PBF, stress-relieved): a 75.9 mm grip of 99.4 mm perimeter with a 24.9 mm finger ring, a hollow core of sheet-TPMS lattice with powder ports, and a double-edged claw blade with serrations on both edges (6 and 5 teeth). It meets every RS-karambit-v5 load case: the lowest margin is tip axial, 1.094 (the minimum over the plate meshes 0.6 / 0.3 / 0.15 mm²).

Measure	Value	Requirement (RS-karambit-v5)	Reading
W_ERGOLIGHT_v2, Tier 2 at h 0.2 mm (rank 1 of 64)	74.40	higher is better	hand multiplier E 0.922 (raised to the 5th power), worst-margin multiplier W 1.000
Mass of record (partial volume at h 0.2 mm, × 0.00443 g/mm³)	88.4 g	ceiling 130 g; 50 g goal (a frontier, not a requirement)	meets target; watertight print mesh 90.3 g
Grip length / perimeter / height-to-thickness	75.9 mm / 99.4 mm / 1.63	≥ 69 mm / 84–100 mm / 1.5–1.8	meets target
Guard height / centre of mass to the index finger	15.8 mm / 7.3 mm	≥ 15 mm / ≤ 8 mm	meets target / meets target
Point: included angle at 1 mm / at 3 mm; width at 2 mm	65.5° / 60.4°; 2.05 mm	≤ 75° / ≤ 60°; ≤ 3 mm	meets target / above the limit; meets target (a continuous penalty, not a gate)
Edges A (inner) / B (spine): per-side grind angle	22.3° / 23.5°	20–25°	largest measured half-angle 23.5°, apex at most 0.113 mm thick
Serrations A / B: depth, teeth	1.09 mm, 6 / 1.30 mm, 5	0.8–1.5 mm; gullet radius ≥ 0.5 mm by construction	weakest tooth SF 2.38 / 3.22
Hand comfort factors (1 = within every limit; E is their weighted product)	contact pressure 0.90	NASA/MIL hand limits (contact pressure, edge radii, grip section)	the other factors are 1.00
Powder: sealed cavity / thin crevices (printed solid)	45 mm³ / 200 mm³	no closed voids; drainage passages ≥ 1 mm	sealed volume is penalised by the score; 4 ports of 3.9 mm (≥ 3 mm)

Every RS-karambit-v5 load case, as a safety factor (SF) against its requirement. The margin is v = SF / requirement; the smallest margin sets the worst-margin multiplier W once it falls below 1. SF is on the production plate mesh (0.3 mm²), as scored; the margin of record is the minimum over the plate meshes 0.6 / 0.3 / 0.15 mm² (cases not solved on the plate are mesh-independent):

Load case	SF	Requirement	Margin v (score mesh)	Margin v of record (min over meshes)
Lateral bending, limit: 12 N·m root moment, load 5 mm from the point (blade plate, grip)	2.01	≥ 1.2	1.678	1.636
Lateral bending, ultimate: 18 N·m (blade plate)	1.47	≥ 1.3	1.132	1.104
Body twist, limit: 16.2 N·m (blade plate, grip)	1.90	≥ 1.2	1.583	1.583
Tip, axial: 150 N along the tip bisector (TET10 solid)	1.31	≥ 1.2	1.094	1.094
Tip, side load: 10 N at the point (TET10 solid)	1.86	≥ 1.2	1.550	1.550
Tip, twist: 0.3 N·m, 3 mm from the point (TET10 solid)	2.50	≥ 1.2	2.083	2.083
Grip squeeze: 800 N	2.32	≥ 1.2	1.933	1.933
Guard push: 300 N	26.39	≥ 1.2	21.993	21.993
Ring bearing: 300 N on the handle side of the bore (T6 ring)	65.39	≥ 2	32.694	32.694
Ring back pull: 30 N (T6 ring)	68.46	≥ 2	34.232	34.232
Ring, any in-plane direction: 300 N (T6 ring sweep)	3.49	≥ 1.2	2.910	2.910
Axial pull: 600 N, held over the 20 mm nearest the point	1.83	≥ 1.2	1.523	1.399
Serration teeth: 180 N snag, 60 N pry shared by 3 teeth (every tooth)	2.38	≥ 1.2	1.983	1.983
Fatigue: 1,000,000 cycles, FKM mean stress (blade and hand zone)	1.17	≥ 1	1.171	1.171
Yield sequence: the blade yields before the hand zone, per load case (lateral / twist / pull)	1.91	≥ 1.2	1.592	1.592
Tip buckling under the 150 N axial load (Rayleigh–Ritz eigenvalue λ)	11.94	≥ 1.5	7.958	7.958

Plan view.	Edge-on: ring, grip and blade thickness.
The blade, reach 72.2 mm.	The point: 66° at 1 mm.
Inner edge (A), 6 teeth, as decoded (before grinding).	Spine edge (B), 5 teeth, as decoded (before grinding).
Grip, 99.4 mm perimeter.	Ring, 24.9 mm bore.
X-ray of the hollow core.	Lengthwise cut: skin, lattice and ports.
Section across the grip.	Section across the blade.
The edge at the point, as decoded (before grinding).

Edge sections at four stations per edge, true scale, from the exact decoder law (dashed: 20° per side, the smallest angle the requirement allows):

Where the score comes from

The score is 100 · E · W · base: the weighted credit of 22 terms (base, at most 1), times the hand multiplier E and the worst-margin multiplier W. Of its 100 points the winner earns 74.4: structural safety 45.3 (the 16 load-case margins), mass 11.0, cutting geometry 11.0 (edge, point and serrations) and build and form 7.1. It loses 6.3 to the multipliers (E = 0.922; the lowest hand factor is contact pressure, e = 0.897; W = 1.000), and 19.3 are unmet credit: terms short of full credit (base = 0.807).

FEM renderings of every load case

Each load case of kv4i1404_g00581_29ea3c73 re-solved by g5-fem-fields on the ranking grid (h 0.2 mm), with the field the scored solvers discard: von Mises stress, the governing point and the deformed shape. The 28 scored values these solves produce match the receipt bit for bit.

Lateral bending, limit load: SF 2.01	Lateral bending, ultimate load: SF 1.47
Body twist, limit load: SF 1.90	Pull: SF 1.83
Fatigue, fully reversed duty: SF 1.17	Point, axial: SF 1.31
Point, side load: SF 1.86	Point, twist: SF 2.50
Ring, bearing toward the handle: SF 65.39	Ring, back: SF 68.46
Ring, any in-plane direction: SF 3.49	Blade before hand (failure sequence): SF 1.91
Buckling of the point: SF 11.94	Grip squeeze: hex8 FE SF 23.78 (score model 2.32)
Guard push: hex8 FE SF 8.85 (score model 26.39)	Teeth: hex8 FE SF 2.72 (score model 2.38)

The score models grip squeeze, guard push and teeth analytically. The hex8 voxel FE of the same design is an independent check, not used by the score:

Case	Score model SF	Hex8 FE SF
Grip squeeze	2.32	23.78
Guard push	26.39	8.85
Teeth	2.38	2.72

How evolution produced it

v6.0.0 continues the v5.0.0 cohort with the same frozen decoder (G5.0, c63629d4b310) and objective, on one platform of record (linux/amd64: static x86_64 binaries, native on Linux and in a Rosetta container on macOS; Tier 2 at h 0.2 mm). The 1,508 final elites of v5.0.0 were re-evaluated on it and merged to the best record per niche: 131 niches, best 63.4 on the search grid (Tier 1, h 0.4 mm).

From that pool, 247 warm-started islands ran 20 rounds on 2 hosts (xbabe3: 7, xbabe4: 13), 296,400 evaluations in all. Whenever a host finished a round, the hub merged its islands into the latest pool and the host's next round started from that pool; every merge is recorded with its sources and recomputed by the receipt. Beside them, 4 fresh-start control islands ran 4,800 evaluations each from scrambled Sobol samples, with no pool and no migration.

The lineage's best rose from 63.4 to 75.5; the best control island reached 59.6.

The best design of the pool at points of the lineage, same camera and light (scores on the search grid):

epoch00/pool 63.4, 104 g warm start

pools/p04 67.8, 96 g

pools/p10 71.9, 96 g

pools/p14 72.8, 91 g

pools/p20 75.5, 89 g final pool

Host	Round	Islands × evaluations	Started from (best)	Merged into (best)
xbabe3	1	12 × 1,200	epoch00/pool (63.36)	pools/p02 (66.12)
xbabe3	2	12 × 1,200	pools/p02 (66.12)	pools/p05 (68.72)
xbabe3	3	12 × 1,200	pools/p05 (68.72)	pools/p08 (69.96)
xbabe3	4	12 × 1,200	pools/p08 (69.96)	pools/p11 (71.94)
xbabe3	5	12 × 1,200	pools/p11 (71.94)	pools/p14 (72.81)
xbabe3	6	12 × 1,200	pools/p14 (72.81)	pools/p17 (74.57)
xbabe3	7	13 × 1,200	pools/p17 (74.57)	pools/p20 (75.54)
xbabe4	1	12 × 1,200	epoch00/pool (63.36)	pools/p01 (65.56)
xbabe4	2	12 × 1,200	pools/p01 (65.56)	pools/p03 (66.85)
xbabe4	3	12 × 1,200	pools/p03 (66.85)	pools/p04 (67.81)
xbabe4	4	12 × 1,200	pools/p04 (67.81)	pools/p06 (68.79)
xbabe4	5	12 × 1,200	pools/p06 (68.79)	pools/p07 (69.85)
xbabe4	6	12 × 1,200	pools/p07 (69.85)	pools/p09 (71.58)
xbabe4	7	12 × 1,200	pools/p09 (71.58)	pools/p10 (71.94)
xbabe4	8	12 × 1,200	pools/p10 (71.94)	pools/p12 (72.23)
xbabe4	10	12 × 1,200	pools/p12 (72.23)	pools/p13 (72.81)
xbabe4	11	12 × 1,200	pools/p13 (72.81)	pools/p15 (73.83)
xbabe4	12	14 × 1,200	pools/p15 (73.83)	pools/p16 (74.57)
xbabe4	13	14 × 1,200	pools/p16 (74.57)	pools/p18 (75.15)
xbabe4	14	14 × 1,200	pools/p18 (75.15)	pools/p19 (75.54)

Round numbers skip 9, which the control islands use (islands 9xx).

Why the search leader is not the winner. The best design on the search grid, kv4i1406_g00009_528e4365 (75.5), ranks 2 once re-scored at Tier 2, h 0.2 mm. Every finalist is judged at Tier 2 (solid tip at 0.08 mm²); the top 16 also on five grids (partial-volume mass in grams at h = 0.4 / 0.3 / 0.25 / 0.2 / 0.15 mm; the h 0.2 and h 0.15 mm masses must agree within 2 %):

Rank	Design	Score	E	W	Mass (g)	Mass at the five grids (g)	h 0.2 vs 0.15
1	kv4i1404_g00581_29ea3c73	74.40	0.922	1.000	88.4	88.6 / 89.5 / 89.2 / 88.4 / 88.6	0.17 %
2	kv4i1406_g00009_528e4365	74.32	0.927	1.000	88.3	88.6 / 89.0 / 88.5 / 88.3 / 88.3	0.05 %
3	kv4i1408_g00735_e6796cbf	74.18	0.928	1.000	88.9	89.1 / 89.5 / 89.1 / 88.9 / 88.9	0.00 %
4	kv4i1405_g00335_c923c290	74.06	0.928	1.000	88.4	88.8 / 89.0 / 88.4 / 88.4 / 88.1	0.31 %
5	kv4i1414_g00435_0e2e651b	74.04	0.931	1.000	89.5	89.8 / 90.1 / 90.2 / 89.5 / 89.5	0.01 %
6	kv4i1410_g00279_290be921	74.04	0.926	1.000	89.0	89.3 / 89.7 / 89.2 / 89.0 / 88.9	0.10 %
7	kv4i1401_g00149_412d5153	74.03	0.926	1.000	88.6	88.9 / 89.4 / 88.7 / 88.6 / 88.5	0.06 %
8	kv4i1409_g00336_4a44463a	74.02	0.931	1.000	89.0	89.3 / 89.8 / 89.1 / 89.0 / 88.9	0.13 %
9	kv4i1401_g00281_b7d5a90f	73.99	0.928	1.000	89.1	89.6 / 89.9 / 89.7 / 89.1 / 89.1	0.08 %
10	kv4i1409_g00504_749c79d8	73.80	0.930	1.000	88.4	88.7 / 89.2 / 88.5 / 88.4 / 88.3	0.14 %
11	kv4i1414_g00999_42619e63	73.62	0.924	1.000	90.4	90.9 / 91.1 / 90.5 / 90.4 / 90.3	0.07 %
12	kv4i1410_g00280_5831a5e3	73.62	0.929	1.000	90.0	90.3 / 90.8 / 90.1 / 90.0 / 89.9	0.10 %
13	kv4i1405_g00205_a7f055b6	72.90	0.924	1.000	90.3	90.3 / 90.8 / 90.3 / 90.3 / 90.2	0.15 %
14	kv4i1411_g00490_71efc13e	72.83	0.931	1.000	90.0	90.3 / 90.7 / 90.0 / 90.0 / 89.8	0.30 %
15	kv4i1310_g00526_5127c223	71.78	0.904	1.000	90.5	90.9 / 91.5 / 90.7 / 90.5 / 91.0	0.51 %
16	kv4i1414_g00960_db1fd3ec	71.53	0.911	0.977	88.6	88.9 / 89.3 / 88.7 / 88.6 / 88.5	0.05 %

Verification and limits

• One platform of record. linux/amd64: static x86_64 binaries, native on Linux and in a Rosetta container on macOS; Tier 2 at h 0.2 mm. Every row of every archive re-scores bit for bit; every pool of the lineage is recomputed from its recorded sources byte for byte; re-evaluating all 282 elites of the final pool reproduces every Tier-1 row (172); and epoch01/xbabe4/island101 re-run on xbabe2 reproduces its history bit for bit (17 entries over 200 evaluations).
• Fleet sizing (fleet/sizing.json). The continuation ran one island per core complex on the Threadripper hosts, following an early pilot. A controlled measurement afterwards (12 Tier-1 evaluations: 5.6 s each at one per core complex vs 7.7 s at four) shows that one island per core would have given 2.89× the throughput. The search is deterministic either way, so this cost evaluations, not correctness; glibc malloc tunables cut minor page faults 2.5× with identical outputs.
• GPUs: renders, not search. A profile of the Tier-1 evaluation (spike/spike.json) puts 39 % of it in voxel kernels and 20 % in the plate solve. Records must stay exact CPU arithmetic, so a GPU could only screen children before Tier 1. A perfect screen costing nothing would give at most 3.1× more evaluations per hour, and the Tier-0 screen already keeps 17 of the 20 best children (rank correlation 0.90). A real predictor costs time and misses designs, so the 3× bar for adopting it is out of reach and the search stays on the CPUs.
• Plate mesh sensitivity (vv/plate_mesh.json). Across plate meshes of 1.2 / 0.6 / 0.3 / 0.15 mm² the lateral safety factor of the study designs varies by up to 20 % (not monotone) and the twist safety factor by up to 6 %; the search and ranking use 0.3 mm². So the winner was chosen on the minimum over 0.6 / 0.3 / 0.15 mm² (plate_meshes.jsonl, solved on the platform of record, whose 0.3 mm² rows reproduce the ranking bit for bit): in rank order, the first design meeting every load case on every mesh. It is the top-ranked design; its lowest margin over the meshes is tip axial 1.094.
• Numerical studies (evidence/g5/v5.0.0/vv/): grid agreement of mass, blade-plate and solid-tip mesh studies, the grip perimeter within 0.12 % of its gene, Rayleigh–Ritz against Euler within 5e-09, and Lloyd niches with 12.5 % less quantisation error than Halton.
• Texture on the blade. The frozen G5.0 decoder lets the evolved surface texture reach the blade, where its depth is the texture depth times the blade scope (0.22 mm × 0.07 = 0.014 mm here), so the blade and its bevels are nearly smooth. The edge figures and measurements use the exact bevel law, which grinding after printing restores: the requirement's finishing allowance (0.5 mm per ground face) exceeds the texture depth. Keeping texture off the bevels by construction is a decoder change for a later generation.
• Status: computational design; not built, printed or physically tested. Nothing here is a qualification for use. The safety factors are computed from models with the limits stated in the paper; the material values are specification minima, not measured coupons.
• Evidence gaps listed in requirements/PROVENANCE-v5.md need physical tests before any build: tooth loads, tip load shares, edge retention, thin-wall roughness, lattice homogenisation, the temperature derate and the drop test.

3D printing and CAD (cad/, millimetres)

The upload kit in cad/ holds six files for kv4i1404_g00581_29ea3c73, each under 50 MiB, generated from the winner's genome row (evidence/g5/v6.0.0/winner_genome.csv). Order notes are in cad/README.md.

File	What it is
cad/kv4i1404_g00581_29ea3c73_AMbuild_mm.stl	print file: skin, sheet-TPMS lattice and powder ports, 0.525 mm voxels, 1,030,568 triangles
cad/kv4i1404_g00581_29ea3c73_AMbuild_mm.3mf	the same print mesh as 3MF, with millimetre units
cad/kv4i1404_g00581_29ea3c73_envelope_mm.stl	outer envelope, 0.325 mm voxels
cad/kv4i1404_g00581_29ea3c73_drawing_mm.dxf	1:1 plan drawing (DXF, mm)
cad/kv4i1404_g00581_29ea3c73_spec_paper.pdf	the G5 paper
cad/kv4i1404_g00581_29ea3c73_xray_cutaway.jpg	cutaway of the hollow core

Finer meshes and the full evaluation records are release assets of v6.1.2, each with its SHA-256 in the receipt:

Asset	Size	SHA-256
g5-v6.0.0-archives.tar.gz	167.2 MiB	fe1651beaa485639…
g5-v6.0.0-eval_h040.jsonl	1.3 MiB	64027768be0aea9d…
g5-v6.0.0-kv4i1404_g00581_29ea3c73_AMbuild_h0.2_mm.stl	361.3 MiB	7054688acce5fd7e…
g5-v6.0.0-kv4i1404_g00581_29ea3c73_envelope_h0.2_mm.stl	116.9 MiB	26987a58000432d0…

Reproduce

# static x86_64 binaries, as on the platform of record (RUSTFLAGS="-C target-feature=+crt-static" cargo build --release --locked --target x86_64-unknown-linux-gnu (Rust 1.85.1))
RUSTFLAGS="-C target-feature=+crt-static" cargo build --release --locked --target x86_64-unknown-linux-gnu --manifest-path crates/g5/Cargo.toml --bins
B=crates/g5/target/x86_64-unknown-linux-gnu/release
# one island: the arguments of any evidence/g5/v6.0.0/launch/<round>_<host>_<island>.json
$B/g5-continue run --out ~/jarambit_runs/evo_g5_v6/<round>/<host>/<island> --pool <pool> --evals <evals> --seed <raw_seed> \
  --host <host> --island <island> --commit bf3e37fa --h 0.4 --tier1-top 0.4 --tier2-top 0 --log-every 25
# a pool of the lineage: the merge of its recorded sources, in order (evidence/g5/v6.0.0/pools/*.json)
$B/g5-continue merge --out <pool> --from <previous pool>,<island>,<island>,...
$B/g5-pilot eval <final_root> eval_h040.jsonl 0.4 4 && $B/g5-pilot tier2 <final_root> eval_h040.jsonl 64 tier2_h020.jsonl 0.2 4
$B/g5-release receipt-v6 . <run_root> <cert_dir>

Every island's launch record and history, every pool's sources and the final pool are in evidence/g5/v6.0.0/; the islands' archives are a release asset.

Repository

• crates/g5/: the G5 genome (G5.0), decoder, materials, load models, objective, search and release tools.
• requirements/: RS-karambit-v5 and its provenance.
• evidence/g5/v6.0.0/: the receipt, the winner's genome, every island's history, the ranking, the grid check and vv/.
• docs/g5/: the paper, the images and the CAD archive. docs/RELEASES.md: every release.
• Earlier generations, kept as their own record: G3 (docs/g3/, crates/g3/), the withdrawn G4 prototype (crates/g4/WITHDRAWN.md) and the v1/v2 study (docs/paper/, docs/WHITEPAPER.md). The v1 command-line pipeline (cargo run -p jarambit) is still in the repository; its plate level is opt-in with JARAMBIT_FEM=1.

G5 leader renders: kv4i1404_g00581_29ea3c73 (current pipeline)

Blender renders of the current G5 pipeline's leader kv4i1404_g00581_29ea3c73 (W_ERGOLIGHT_v2 74.40 on the h 0.2 mm ranking grid, mass 88.4 g, lattice porosity 0.58, internal void fraction of the AM-build mesh 0.53) from its AM-build STL (1,030,568 triangles). Regenerate with python3 tools/refresh_leader_assets.py (how). Finishes are illustrative; cuts are open so the printed internal lattice shows.

Reverse oblique.	Plan view.
Edge-on.	Ring.
Grip and pockets.	Blade.
The point.	Edge at the point.
Mid-plane cut (z = 0).	Section across the grip.
Section across the blade.

Legacy-pipeline run (v6.1.0 addendum — older, solid, non-lattice design line; not G5, not the cohort of record)

The most recent legacy continuous island run (legacy v1 W_ULTRALIGHT_v1 pipeline, 103 islands on xbabe2, stopped 2026-10-29) found ec2i089_r01_p10_c68: fitness 80.851214, mass 83.409 g (3D 83.3333 g), 13 of 26 legacy hard gates met. It misses the hunter 260 N load (201–202 N), tip twist and solid-tip safety factors (0.52–0.71 vs 1.0), fatigue (0.56), human maximum, sharpness law and tip profile, so it is a research result, not a qualified design. Full gate table and plots: docs/LATEST_RUN.md; evidence: evidence/legacy_continuous_run_v6.1.0/.