Skip to content

Cortext v1.0

Choose a tag to compare

View all tags
@gabewillen gabewillen released this 30 Jun 14:45
v1.0
e61f280

Cortext v1.0 Release Artifact

This file is the tracked release artifact for the v1.0 line. The v1.0 and
v1.0.0 tags should point at the commit that contains this artifact.

Release State

  • Version: 1.0.0
  • Runtime: hard-cut production engine, limited to embedding-backed memory,
    durable graph associations, bounded retrieval, and shallow consolidation.
  • Removed research stack: decoder, provider registry, semantic extractor,
    summarizer, static label bank, fact layer, label-bucket graph, and
    mode-selected deep consolidation.
  • Public API status: C++ facade, C ABI, and binding entry points remain the
    release surface for this tag.

Version Surfaces

The tracked package surfaces now report 1.0.0:

  • CMakeLists.txt
  • bindings/python/pyproject.toml
  • bindings/javascript/package.json
  • bindings/dart/pubspec.yaml
  • build.zig
  • build.zig.zon

cortext_version() and every FFI package should report the same version.

Current Evidence

The current v1.0 public evidence is the hosted Meta Multi-Session Chat eval on
the public nayohan/multi_session_chat mirror, plus the 128k RAG ablation and
post-optimization full replay verification.

Hosted Frontier Judge

  • Artifact:
    eval_runs/msc_frontier_late_200dlg_gpt55_20260630T053427Z/judge_openai_gpt55_four_system_clean.json
  • Replay: 9,130 text turns from 708 public MSC rows.
  • Protocol: daily source-time consolidation at 02:00 UTC, 5,000-event warmup,
    500-event probe stride, 9 probes, 3 hosted gpt-5.5 blind judgments per
    probe, 27/27 judgments complete.
  • Result: Cortext won 21/27, traditional chat+RAG won 0/27, full-history upper
    bound won 1/27, hosted compaction rollup won 5/27.
  • Mean context: Cortext 998 tokens versus traditional chat+RAG 49,196 tokens,
    a 97.97% reduction with probe-bootstrap 95% CI [97.77%, 98.17%].
  • Caveat: this supports token-reduction, noise-reduction, and blind-win claims;
    it is not a completed sufficiency-match claim.

128k RAG Ablation

  • Artifact:
    eval_runs/msc_rag_ablation_128k_gpt55_20260630T_actual/judge_openai_gpt55_rag_ablation_128k.json
  • Protocol: same 9 MSC probes, 3 hosted gpt-5.5 blind judgments per probe,
    128,000-token judge-context cap, 27/27 judgments complete.
  • Systems: Cortext native, semantic vector RAG, lexical keyword RAG,
    rolling-window chat, 16k hybrid chat+vector RAG, hosted compaction rollup.
  • Result: Cortext won 19/27, hosted compaction rollup won 8/27, and all four
    RAG-style packet variants won 0/27.
  • Prompt cap: max estimated judge prompt 116,425 tokens under the 128,000-token
    cap.
  • OpenAI usage: 1,871,994 prompt tokens and 39,001 completion tokens across 27
    judge requests.

Full MSC Latency Verification

  • Artifact:
    build/graph_profile/full_msc_verify_rerun_20260630T142345Z/summary.json
  • Replay: same 9,130-turn MSC slice, same default knobs, daily consolidation,
    5,000-event warmup, and 500-event probe stride.
  • Completion: 9 probes, 714 consolidations, 10,099 memories, 9,364 long-term
    memories, 21 working-memory rows, and 17,719 associations.
  • Behavior: recursive non-timing probe content matched the saved frontier
    artifact and the prior post-optimization replay.
  • Latency: judged probe total latency was 50.4-80.2 ms, mean 63.5 ms;
    GraphRetrieve.total was 22.2-29.2 ms, mean 25.9 ms.

Full protocol details are recorded in docs/paper/sections/9_experimental.qmd,
docs/paper/sections/11_optimization.qmd, and the generated manuscript at
docs/paper/_manuscript/index.md.

Build Gate

cmake -S . -B build/release \
  -DCMAKE_BUILD_TYPE=Release \
  -DCORTEXT_DISABLE_OPENTELEMETRY=ON
cmake --build build/release -j
./build/release/tests/cortext_tests '~[aist]' --reporter compact

Run AIST-tagged tests separately on a host with the model installed:

./build/release/tests/cortext_tests '[aist]' --reporter compact

FFI Gate

cmake --preset ffi-release-node
cmake --build --preset ffi-release-node --target cortext cortext_node -j
(cd bindings/go && go test)
(cd bindings/dart && dart pub get && dart analyze && dart test)
python3 -m py_compile bindings/python/cortext/__init__.py
node --check bindings/javascript/index.js bindings/wasm/cortext.js examples/web/main.js
node -e "const c=require('./bindings/javascript'); if (c.version() !== '1.0.0') process.exit(1)"

Browser WebAssembly Gate

./build-wasm.sh
test -s build-wasm/dist/wasm/cortext.js
test -s build-wasm/dist/wasm/cortext.wasm

CI Gate

The GitHub workflow must pass with optional OpenTelemetry disabled for
deterministic model-free CI:

  • Ubuntu native release build and non-AIST tests.
  • Arch Linux native release build and non-AIST tests.
  • Browser WebAssembly bundle build.
  • Zig host and Linux cross-build smoke checks.

Documentation Gate

When algorithm behavior or experiment conclusions change, update
docs/paper/sections/ and regenerate:

QUARTO_DISABLE_GIT=1 QUARTO_DISABLE_GITHUB=1 quarto render docs/paper

For packaging/API changes, update the root README and all affected binding
READMEs before tagging.

Assets 2
Loading