Skip to content
jsonW Home GitHub

jsonW 1.4.2 · measured evidence

Test transparency

Wortic Tecnologia · updated November 19, 2026

Every number we publish about jsonW comes from a recorded measurement. Here is each one: what we measured, the result, the date, the machine, the method, the command to rerun it and the raw output.

Evidence
  1. 1. Speed and size
  2. 2. The real INPI dataset
  3. 3. Key lookup
  4. 4. Partial remote read
  5. 5. Custody
  6. 6. Permission views
  7. 7. Encryption at rest
  8. 8. Measures, money and documents
  9. 9. Protocol profiles
  10. 10. Post-quantum transport
  11. 11. SuperVendor on jsonW
  12. 12. Wortic AI models
  13. 13. Rerun everything

In numbers

6.6×smaller than JSONL, 100k records
1,886×faster to read one column (median of 10)
613×indexed key lookup vs. scan, 1 million rows
12.8×smaller than XML on the full INPI dataset
49 of 49protocol objects round-trip byte for byte
100%accuracy on 2 Wortic AI models

1Speed and size: 100,000 records

jsonW against the same set in JSONL: 100,000 knowledge-base cards with text, a decimal, a list and a 384-dimension vector. The .jsw writes 124 times faster, reads everything 48 times faster, reads one column 1,886 times faster and takes 6.6 times less space. On every run the tool checks that both reads counted 100,000 rows with the same sum.

Time. Median of 10 runs; ratio = JSONL ÷ jsonW on each run, with the median and the range of the 10.
OperationJSONLjsonW 1.4.2Ratio (range)
Write15,312 ms130 ms124× (94 to 136)
Read all, every field3,756 ms79 ms48× (35 to 51)
Read all, 8 threads (JSONL on one)3,756 ms8.4 ms453× (262 to 598)
Read one column3,756 ms2.00 ms1,886× (1,487 to 2,542)
Go straight to row N, one fieldread from the start0.075 µs
One message in row form5,157 B1,073 B4.8× (all 10)
Size. The same 100,000 records, in MB (106 bytes); ratio = JSONL ÷ the file.
FileSizeRatio
JSONL519.31.0×
.jsw, vf16 vector, default level79.16.6×
.jsw, vq8 vector41.112.6×
.jsw, vf16 vector, --max1.36381×
.jsw without the vector column, --max0.01752,185× (vs. 38.2 MB)
  • Date: 2026-11-19, 23:00 to 23:06 UTC (times); sizes at 23:06 UTC.
  • Version: public jsonW 1.4.2, bin/linux-x86_64/wjsw from tag v1.4.2, SHA-256 3e4e587b7153e69875c6e946a1596d99207ace61c41f016603f966a924e43e55.
  • Machine: virtual server, AMD EPYC 9355P, 8 vCPUs, 18 GB of memory, Linux 7.3; files on a tmpfs.
  • Method: wjsw medir with 100,000 rows, 10 runs in a row; sizes from gerar-cartoes.awk and wjsw de-texto, custody of all 6 files verified (ok: 100000 linhas on each). The JSON side is written and read by jsonW's own text reader.
  • Raw output: medir-01 to medir-10, sizes, custody and machine.
git clone --branch v1.4.2 https://github.com/developer-wortic/jsonw && cd jsonw
export JSONW_SEM_PING=1
for i in $(seq 1 10); do bin/linux-x86_64/wjsw medir /tmp 100000; done

2The real INPI dataset: 55.3 million rows

The seven files of the Brazilian INPI open trademark data, converted 1:1 with nothing cut: 55,278,212 rows and 12.53 GB of CSV. As .jsw the dataset is 4.9 times smaller than the CSV at the default level, 6.5 times at the maximum level and 12.8 times smaller than the XML, with the SHA-256 and CRC32C of every block of all seven files verified. Counting the in-force records, the .jsw answered in 43 ms, ahead of Parquet read by DuckDB, DuckDB reading the CSV, grep and the compressed CSV, with the right answer: 2,343,607.

Count the in-force records (codigo_situacao = IPAS023, 6.7 million filings), in ms, lowest of 5, all in the same hour.
Format and readerTimeAnswer
.jsw --max 1.3.11, wjsw calcular --onde432,343,607
.jsw 1.3.11, wjsw calcular --onde442,343,607
Parquet, DuckDB 2.0.0, 8 cores542,343,607
CSV, DuckDB 2.0.0, 8 cores3752,343,607
CSV, grep -c5192,343,607
CSV + zstd, zstd -dc | grep -c7982,343,607
  • Date: 2026-11-13, 06:21 UTC (fifth round); INPI data downloaded 2026-11-12.
  • Version: the 1.3.11 release wjsw; 1.4.2 reads the same files.
  • Machine: 8-core server, 31 GB of memory, Linux 7.3.
  • Method: a hand-written schema per file; the C converter checks every number and date against the CSV text; each query ran 5 times with the file cached and the lowest counts; Parquet written and read by DuckDB 2.0.0 with zstd.
  • Where: Tables 5C and 5D of the paper.
curl -O https://dadosabertos.inpi.gov.br/download/marcas/marcas.zip
# each CSV converted to .jsw with a hand-written schema, as in Section 5.4 of the paper
wjsw calcular MARCAS_DADOS_BIBLIOGRAFICOS.jsw --onde "codigo_situacao = IPAS023"
wjsw conferir MARCAS_DADOS_BIBLIOGRAFICOS.jsw

3Key lookup: 1 million rows

The ABI key lookup on a one-million-row file, median of 200 lookups by id: 77.4 ms scanning the column and 0.126 ms through the index, 613 times faster, with 0 wrong answers in 400 lookups. The index took 160 ms to build, once. The same program runs the proof cases (unique, repeated and missing key, wrong type and the tenant view) and ends with TUDO CERTO (0 falhas).

  • Date: 2026-11-19, 23:07 UTC.
  • Version: sdk/prova_chave.c from the repository, linked against the public 1.4.2 libjsonw.a.
  • Machine: the same as test 1.
  • Result: scan median 77.3590 ms, p95 88.8815 ms; index median 0.1263 ms, p95 0.2122 ms; writing the million rows, 535 ms.
  • Raw output: saidas-1.4.2.txt, section "busca por chave".
cc -O2 -Isdk -o prova_chave sdk/prova_chave.c bin/linux-x86_64/libjsonw.a -lpthread -lm
./prova_chave                 # the proof cases
./prova_chave --medir 1000000 # the measurement

4Partial remote read

wremoto opens a 79.1 MB .jsw by downloading 8.7 KB (header and footer). To read row 50,000 it downloads 0.98% of the file; for the filter id > 99990, 0.84%, because the zones rule out the other blocks before downloading. Read locally, the same filter skips 102 of 103 blocks and answers in 0.9 ms.

  • Date: 2026-11-19, 23:07 UTC. Version: wremoto and wjsw 1.4.2. Machine: the same as test 1.
  • Method: the 100,000-record file from test 1, read through a file:// address, with the same byte-range requests wremoto sends to an HTTP server, R2 or S3; wremoto itself counts the bytes downloaded.
  • Raw output: saidas-1.4.2.txt, "Listagem 6".
$ wremoto ver file://$PWD/com.jsw 50000 1
baixado no total: 777.8 KB em 5 pedidos (0.98% do arquivo)
$ wremoto filtrar file://$PWD/com.jsw 'id > 99990'
baixado no total: 667.7 KB em 5 pedidos (0.84% do arquivo)

5Custody: no byte changes unnoticed

One byte changed in the middle of the file fails conferir; a file cut short at the end, as in an interrupted write, is flagged and reparar makes it whole again; the chain works with SHA-256 and with BLAKE3.

$ wjsw conferir vendas.jsw
ok: 2 linhas, 1 blocos, sha256 e crc32c conferidos
$ wjsw conferir adulterado.jsw
FALHOU: bloco 0: nao veio ou nao confere
$ wjsw conferir cortado.jsw
FALHOU: sem rodape: rode wjsw reparar
$ wjsw reparar cortado.jsw && wjsw conferir cortado.jsw
ok
ok: 2 linhas, 1 blocos, sha256 e crc32c conferidos
$ wjsw regravar vendas.jsw vendas-b3.jsw --hash blake3 && wjsw conferir vendas-b3.jsw
2 linhas: 708 -> 708 bytes (100.0%), 0.0 s
ok: 2 linhas, 1 blocos, blake3 e crc32c conferidos
  • Date: 2026-11-19, 23:07 UTC. Version: wjsw 1.4.2. Machine: the same as test 1.
  • Method: adulterado.jsw = an X written a third of the way into the file with dd; cortado.jsw = the file without its last 20 bytes. Inputs at jsonw.org/paper/exemplos.
  • Raw output: saidas-1.4.2.txt, "Listagem 7".

6Permission views on every command

In a clinic file with the national ID marked +pessoal, the amount marked +escopo(financeiro) and a tenant column, each credential sees only what it may: front desk at unit 1 sees unit 1's schedule without ID or amount; finance sees the amount; nobody at unit 1 sees unit 2. Without scopes, para-texto returns ID and amount as null, and filtering by the ID is refused. In our command-line permission test, 1.4.1 passes all 37 checks.

$ WJSW_ESCOPOS="tenant:1" wjsw ver clin.jsw
{paciente:"Ana Souza",cpf:null,valor:null,data:t"2026-11-12T13:00:00.000000Z",tenant:1}
{paciente:"Carla Dias",cpf:null,valor:null,data:t"2026-11-12T15:30:00.000000Z",tenant:1}
$ WJSW_ESCOPOS="financeiro tenant:1" wjsw ver clin.jsw
{paciente:"Ana Souza",cpf:null,valor:180.00 BRL,data:t"2026-11-12T13:00:00.000000Z",tenant:1}
{paciente:"Carla Dias",cpf:null,valor:180.00 BRL,data:t"2026-11-12T15:30:00.000000Z",tenant:1}
$ wjsw filtrar clin.jsw "cpf = 012.345.678-90"
coluna sem permissao: cpf (exige escopo)
  • Date: 2026-11-19, 23:07 UTC, with wjsw 1.4.2 (the listings); 37 of 37 with 1.4.1 on macOS arm64, at the 1.4.1 release on 2026-11-15, recorded in the package README.
  • Machine: the same as test 1. The clinic data is fictitious, in clin.jsonl.
  • Raw output: saidas-1.4.2.txt, "Listagem 8".

7Encryption at rest

A .jsw encrypted with XChaCha20-Poly1305 and a 256-bit key kept outside the file does not open without the key, neither to verify nor to read; with the key, custody checks out and the rows come back. We changed one byte of the encrypted file at three positions: at all three, conferir and para-texto refused the file.

$ wjsw cifrar cifrado.jsw --chave chave.hex
$ wjsw conferir cifrado.jsw
FALHOU: arquivo cifrado: falta a chave do tenant (WJSW_CHAVE_ARQ ou credencial wjsw-chave)
$ wjsw conferir cifrado.jsw --chave chave.hex
ok: 1 linhas, 1 blocos, sha256 e crc32c conferidos
# one byte changed at position 238 of 716
$ wjsw conferir troca.jsw --chave chave.hex
FALHOU: bloco cifrado nao e o que o indice assinou
  • Date: 2026-11-19, 23:13 UTC. Version: wjsw 1.4.2. Machine: the same as test 1.
  • Method: a test key generated on the spot from /dev/urandom, never printed and deleted at the end; positions 238, 358 and 477 of 716 bytes.
  • Script and raw output: provar-cifra.sh and saidas-cifra-1.4.2.txt.

8Measures, money and documents

A measure column keeps the converted value with no loss: 1 lb becomes 0.45359237 kg in a medida[kg] column, and 1 mi becomes 1609.344 m in a medida[m]. calcular --valor sums money in whole cents and returns the currency: 12.50 + 0.25 = 12.75 BRL. A mascara(cpf) column accepts the Brazilian CPF with or without punctuation, stores the same canonical form and refuses a CPF with a wrong check digit.

$ wjsw de-texto "id:i64,peso:medida[kg],preco:dinheiro" m.jsonl m.jsw && wjsw ver m.jsw
2 linhas -> m.jsw
{id:1,peso:0.45359237 kg,preco:12.50 BRL}
{id:2,peso:0.90718474 kg,preco:0.25 BRL}
$ wjsw calcular m.jsw --valor preco
1 grupo(s) em 0.4 ms
total: 2 linha(s) · soma 12.75 BRL · min 0.25 · max 12.5 · media 6.375
$ wjsw de-texto "id:i64,cpf:mascara(cpf),cnpj:mascara(cnpj)" docs.jsonl docs.jsw && wjsw para-texto docs.jsw
2 linhas -> docs.jsw
{id:1,cpf:"12345678909",cnpj:"11222333000181"}
{id:2,cpf:"12345678909",cnpj:"11222333000181"}
  • Date: 2026-11-19, 23:07 UTC. Version: wjsw 1.4.2. Machine: the same as test 1.
  • Raw output: saidas-1.4.2.txt, sections "1.4.2" and "1.4.0: máscaras".

9Agent and commerce protocol profiles

49 objects from A2A, UCP, ACP, AP2, Verifiable Intent, x402, MPP, AG-UI, OpenTelemetry, L402 and Pix, each written to .jsw with the profile schema, verified by custody and returned as JSON byte for byte equal to the sample: 49 ok, 0 failures.

$ bash perfis/provar-perfis.sh bin/linux-x86_64/wjsw | tail -1
perfis: 49 ok, 0 falhas
  • Date: 2026-11-19, 23:07 UTC. Version: wjsw 1.4.2 and the perfis/ folder of tag v1.4.2. Machine: the same as test 1.

10Post-quantum transport

The connection to jsonw.org negotiates the hybrid X25519MLKEM768 key exchange (X25519 with FIPS 203 ML-KEM-768), over TLS 1.3 with the TLS_AES_256_GCM_SHA384 cipher.

$ openssl version
OpenSSL 4.1.0 23 Oct 2026 (Library: OpenSSL 4.1.0 23 Oct 2026)
$ echo | openssl s_client -connect jsonw.org:443 -servername jsonw.org 2>&1 | grep -E "Negotiated|Cipher is"
Negotiated TLS1.3 group: X25519MLKEM768
New, TLSv1.3, Cipher is TLS_AES_256_GCM_SHA384
  • Date: 2026-11-19, 23:04 UTC. Machine: Linux server in Brazil, OpenSSL 4.1.0.

11SuperVendor: company search in C on jsonW

SuperVendor's company search, rewritten in C on jsonW files, returned the same answer as the old database search in 18 of 18 requests, including 4 WhatsApp requests. Search time in C: 55 to 1,562 ms per request; in the database: 83 to 90,010 ms.

  • Date: 2026-11-13, 13:45. Machine: Wortic's production server, on a parallel instance whose process SHA-256 was checked before the test.
  • Method: the same request sent to both searches and the answers compared one by one; script prova-sv-busca.sh, recorded in the report of the 2026-11-13 session.

12Wortic AI models

Wortic's AI models answer from knowledge bases stored in jsonW. Each one is measured against its own answer key: a question counts as right when the model decides and its first answer is among the accepted ones; on a question outside its knowledge, it counts as right when the model says it does not know.

Accuracy per model. Benchmark of 2026-11-16; answer time per question in microseconds (median and p95).
ModelQuestionsRightAccuracyMedianp95
Adam ESG96795799.0%546 µs729 µs
SaaS Management22021999.5%191 µs436 µs
Roteiro181181100%378 µs593 µs
Site Builder16415795.7%233 µs415 µs
SuperVendor151151100%379 µs642 µs
  • Date: 2026-11-16, around 22:45. Machine: Wortic's build container on the 8-core server, gcc 15.3.
  • Method: the benchmark runs every question in each model's answer key and records questions, accuracy, top 3, top 10 and times; right answers are the percentage applied to the number of questions.
  • Record: the benchmark file regua/depois.jsw.txt, in jsonW text, dated 2026-11-16.

13Rerun everything

Tests 1 and 3 to 9 run with the public package on a Linux x86_64 machine, with two scripts:

git clone --branch v1.4.2 https://github.com/developer-wortic/jsonw jsonw
mkdir -p exemplos-142/entradas && cd exemplos-142
for f in clin dup hoje leite marcas ontem prod termos vendas; do curl -s -o entradas/$f.jsonl https://jsonw.org/paper/exemplos/$f.jsonl; done
curl -O https://jsonw.org/paper/exemplos/rodar-exemplos.sh -O https://jsonw.org/paper/exemplos/provar-cifra.sh
bash rodar-exemplos.sh ../jsonw saidas.txt
bash provar-cifra.sh ../jsonw saidas-cifra.txt

rodar-exemplos.sh reruns every listing in the paper and the new 1.4.0 to 1.4.2 tests, writing each command with its real output below. Times change with the machine; results, counts and messages stay the same. Corrections and third-party measurements are welcome through the public repository.