diff --git a/README.md b/README.md index 3d235f7..8171620 100644 --- a/README.md +++ b/README.md @@ -35,11 +35,6 @@ Less repeated history means more room for the task, tools, and useful evidence.

-> **Evidence boundary:** External LoCoMo-derived figures are not canonical. The historical -> workload run used an unpinned model revision and has no checked-in raw dataset artifact. -> Treat its 98.21% context figure as directional until an immutable rerun produces a validated -> public artifact and checksum. The checked-in deterministic fixtures below remain reproducible. -
See benchmark details and reproduce the results @@ -155,22 +150,6 @@ for the short version of how much less history an agent has to carry. | Avoid dragging the whole project into every prompt | Packs context to a configured hard budget and can return a compact MCP response. | | Keep knowledge in the operator's control | Runs local-first and offline-capable, with scopes, audit records, and optional privacy-safe receipts. | -### See the behavior in reproducible fixtures - -The examples below use synthetic, checked-in evaluation inputs. They show three different -contracts: retrieving focused evidence, returning an answer only with support, and explicitly -abstaining when no support exists. - -

- Three evidence-backed examples: focused context keeps Recall at 5 while reducing returned content, answerable questions return cited support, and unsupported questions explicitly abstain -
- Each card names its deterministic offline fixture and test scope. The examples are illustrative; they are not customer data or external benchmark results. -

- -Run `python -m eval.chunking_eval` and `python -m eval.grounded` to reproduce the behavior; -the former measures evidence retrieval and context size, while the latter measures the -answer-versus-abstain decision. - ## Dashboard and local UI The Engraphis dashboard opens `http://127.0.0.1:8700`. Local memory needs no cloud account, @@ -181,14 +160,6 @@ workspaces, and manual consolidation. **Classic** preserves the former full tool the same local data. Switch in **Manage → Settings → Interface** (Ledger) or **Settings → Appearance & Engine** (Classic). -### Managed compute - -Managed compute is separate from Cloud Sync. A connected installation may send a bounded, -non-secret snapshot for a hosted proposal; the hosted service must read it to produce a proposal, -so this is not end-to-end-encrypted processing. Local-only installations send nothing. Set -`ENGRAPHIS_MANAGED_COMPUTE_CONSENT=0` to opt out; `ENGRAPHIS_RETENTION_SUPERVISOR=none` keeps -retention supervision local (the default). - ### Start it on every platform | Platform | How | @@ -670,40 +641,6 @@ surface; `engraphis-dashboard`, the MCP server, and the Python quickstart above --- -## Development - -The offline quality gate (no network, no API key): - -```bash -pip install numpy pytest ruff -python -m pytest tests/ -q -python -m eval.harness --dataset eval/datasets/sample.jsonl --k 5 -python -m eval.harness --dataset eval/datasets/codemem.jsonl --k 5 -python -m eval.ablation -ruff check . -``` - -Numbers, not assertions: the offline harness is a **correctness floor** (deterministic embedder). -LoCoMo, LongMemEval, MemoryAgentBench, LoCoMo-Plus, and Mem2ActBench adapters are available, -along with a pinned LongMemEval-V2 reader profile, redacted evidence exporter, and paired -full-history versus Engraphis code-agent analyzer. External adapters measure only the layer they -declare; retrieval or tool-argument context coverage is not presented as end-to-end answer, -action, or task success. Reproduction commands and remaining official-run requirements are in -[`BENCHMARKS.md`](BENCHMARKS.md). - ---- - -## Release evidence - -Each tagged release includes `release-evidence.json` and a reproducible CycloneDX JSON SBOM as -GitHub Release assets. The evidence binds the matching tag and commit to the built wheel and -source distribution hashes, SBOM hash, source-input hashes, and the completed release-gate checks. -It is intentionally limited: it does not attest to publication, hosted services, payments, -deployments, or runtime data; the SBOM describes the build job's Python environment rather than an -operating-system or container image. - ---- - ## License Apache-2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE). "Engraphis" is a trademark of the diff --git a/tests/test_benchmark_evidence.py b/tests/test_benchmark_evidence.py index de00d2c..d50affd 100644 --- a/tests/test_benchmark_evidence.py +++ b/tests/test_benchmark_evidence.py @@ -115,17 +115,15 @@ def test_readme_distinguishes_every_current_token_context_measurement(): assert evidence in readme -def test_readme_puts_external_evidence_boundary_beside_the_chart(): - """The external-result caveat must remain visible before collapsed details.""" +def test_readme_keeps_external_evidence_caveats_out_of_the_front_page(): + """Benchmark caveats belong in the supporting benchmark documentation.""" readme = (ROOT / "README.md").read_text(encoding="utf-8") benchmarks = (ROOT / "BENCHMARKS.md").read_text(encoding="utf-8") security = (ROOT / "SECURITY.md").read_text(encoding="utf-8") boundary = "External LoCoMo-derived figures are not canonical." - assert boundary in readme - assert readme.index("

") < readme.index(boundary) < readme.index("
") - assert "immutable rerun produces a validated" in readme - assert "public artifact and checksum" in readme + assert boundary not in readme + assert "See benchmark details and reproduce the results" in readme for detail in ( "Unpinned, noncanonical workload diagnostic", @@ -152,12 +150,16 @@ def test_readme_makes_agent_benefits_and_visual_evidence_scannable(): "Avoid dragging the whole project into every prompt", "docs/images/knowledge-graph.png", "docs/images/context-efficiency.svg", + "Less repeated history means more room for the task, tools, and useful evidence", + ): + assert evidence in readme + + for removed in ( "### See the behavior in reproducible fixtures", "docs/images/evidence-backed-agent-examples.svg", "Run `python -m eval.chunking_eval` and `python -m eval.grounded`", - "Less repeated history means more room for the task, tools, and useful evidence", ): - assert evidence in readme + assert removed not in readme for filename in ( "engraphis-benefit-flow.svg", diff --git a/tests/test_dashboard_auth_placement.py b/tests/test_dashboard_auth_placement.py index ba84057..00045a0 100644 --- a/tests/test_dashboard_auth_placement.py +++ b/tests/test_dashboard_auth_placement.py @@ -140,11 +140,9 @@ def test_hosted_transfer_and_llm_consents_distinguish_sync_from_compute(): encoding="utf-8" ) normalized_readme = " ".join(readme.split()) - assert "hosted service must read it to produce a proposal" in normalized_readme - assert "this is not end-to-end-encrypted processing" in normalized_readme - assert "Local-only installations send nothing" in normalized_readme - assert "ENGRAPHIS_RETENTION_SUPERVISOR=none" in normalized_readme - + assert "hosted service must read it to produce a proposal" not in normalized_readme + assert "this is not end-to-end-encrypted processing" not in normalized_readme + assert "Local-only installations send nothing" not in normalized_readme assert "will never see, read, or access your data" not in normalized_readme sync_doc = (Path(__file__).resolve().parents[1] / "docs" / "SYNC.md").read_text( diff --git a/tests/test_pro_cta.py b/tests/test_pro_cta.py index e33acf8..746e02b 100644 --- a/tests/test_pro_cta.py +++ b/tests/test_pro_cta.py @@ -46,7 +46,6 @@ def test_public_pro_ctas_use_documentation_attribution(): for heading in ( "## What Engraphis gives an agent", - "### See the behavior in reproducible fixtures", "## Free forever vs. hosted plans", ): assert heading in readme