chore: restore full history with private data purged

Author: Engraphis Maintainers

BENCHMARKS.md

@@ -11,8 +11,8 @@ and stated everywhere the numbers appear (`eval/external.py`).
 
 - **Correctness gate** — `eval/harness.py` over `eval/datasets/sample.jsonl` and
   `codemem.jsonl` (conflict resolution) and `graph_multihop.jsonl` (multi-hop graph recall).
-  Runs on the deterministic embedder, so it is a plumbing/regression floor, not a competitive
-  number. This is the gate CI enforces.
+  Runs on the deterministic embedder, so it is a plumbing/regression floor, not a public
+  performance claim. This is the gate CI enforces.
 - **Ablation** — `eval/ablation.py`: vector-only vs. 1-hop graph vs. Personalized-PageRank arm,
   to show the graph arm actually earns its place.
 - **External benchmarks** — `eval/external.py` loads **LoCoMo** and **LongMemEval** and pushes

README.md

@@ -180,6 +180,12 @@ structured-extraction entries.
 
 ---
 
+## Evidence-backed capabilities
+
+Engraphis publishes capability descriptions and reproducible benchmark protocols without narrative vendor comparisons. See [`BENCHMARKS.md`](BENCHMARKS.md) for the evidence and claim policy.
+
+---
+
 ## Hosted Pro and Team
 
 Pro and Team are services, not alternate modes hidden in the public image. The official cloud

engraphis/backends/extractor.py

@@ -1,6 +1,6 @@
 """Fact extractors — implementations of the ``core.interfaces.Extractor`` protocol.
 
-Every SOTA memory system (external memory system, A-Mem, external memory system) auto-distills raw text into discrete
+Modern memory systems often auto-distill raw text into discrete
 facts before storage; Engraphis makes that step *pluggable and optional* so the core
 stays offline-capable (AGENTS.md §3.8):
 

engraphis/classic_assets/dashboard.js

@@ -336,6 +336,6 @@
      graph view. dashboard.js fetches both on demand from graphRender(); see loadForceGraph()
      and loadGraphEngine(). scripts/externalize_dashboard_assets.py enforces both halves:
      they stay out of this file, and the lazy references still have to resolve. -->
-<script src="/classic-assets/dashboard.js?v=20260729-hub-materials"></script>
+<script src="/classic-assets/dashboard.js?v=20260728-reference-materials"></script>
 </body>
 </html>

engraphis/classic_assets/index.html

@@ -102,7 +102,7 @@ def _validated_token_subject(value: object) -> str:
     return subject
 
 
-def _token_subject(saved: dict, *, uses_persisted_credential: bool = True) -> str:
+def _token_subject(saved: dict) -> str:
     """Return the immutable subject bound to the current credential family.
 
     ``ENGRAPHIS_CLOUD_TOKEN_SUBJECT`` selects the subject for an environment-only
@@ -111,13 +111,11 @@ def _token_subject(saved: dict, *, uses_persisted_credential: bool = True) -> st
     environment override win causes the next refresh to present (for example) a
     member credential as ``device``; the service must reject that mismatch, and the
     client then has to retire a credential that was never actually spent.  Persisted
-    state therefore wins only while the selected credential itself is persisted.
-    A retired saved credential leaves its subject behind for auditability; a replacement
-    environment bootstrap credential must instead use its explicitly configured subject.
+    state therefore wins whenever it carries a subject.
     """
 
     persisted = saved.get("token_subject")
-    if uses_persisted_credential and persisted is not None and str(persisted).strip():
+    if persisted is not None and str(persisted).strip():
         return _validated_token_subject(persisted)
     configured = os.environ.get("ENGRAPHIS_CLOUD_TOKEN_SUBJECT", "").strip()
     return _validated_token_subject(configured or "member")
@@ -796,15 +794,15 @@ def configured(*, require_compute: bool = True) -> bool:
     # A configured environment value is bootstrap material. After its first successful
     # use, the server-returned rotation is persisted and must take precedence; otherwise
     # every subsequent call would replay the now-invalid bootstrap credential.
-    saved_refresh = str(saved.get("refresh_credential") or "").strip()
-    refresh = saved_refresh or os.environ.get("ENGRAPHIS_CLOUD_REFRESH_CREDENTIAL", "").strip()
+    refresh = str(saved.get("refresh_credential") or "").strip()
+    refresh = refresh or os.environ.get("ENGRAPHIS_CLOUD_REFRESH_CREDENTIAL", "").strip()
     if _refresh_is_unusable(saved, refresh):
         refresh = ""
     control = os.environ.get("ENGRAPHIS_CLOUD_CONTROL_URL", "").strip()
     control = control or str(saved.get("control_url") or "").strip()
     compute = direct_compute or str(saved.get("compute_url") or "").strip()
     if refresh and control:
-        _token_subject(saved, uses_persisted_credential=bool(saved_refresh))
+        _token_subject(saved)
     return bool(refresh and control and (compute or not require_compute))
 
 
@@ -839,8 +837,8 @@ def access_for_workspace(
         # single-use credential; reading it before the lock lets two workers spend the
         # same value and causes one request to fail as a replay.
         saved = _load()
-        saved_refresh = str(saved.get("refresh_credential") or "").strip()
-        refresh = saved_refresh or os.environ.get(
+        refresh = str(saved.get("refresh_credential") or "").strip()
+        refresh = refresh or os.environ.get(
             "ENGRAPHIS_CLOUD_REFRESH_CREDENTIAL", ""
         ).strip()
         if _refresh_is_unusable(saved, refresh):
@@ -858,9 +856,7 @@ def access_for_workspace(
             )
         control = _reachable_cloud_base_url(control)
         compute = _reachable_cloud_base_url(compute) if compute else ""
-        token_subject = _token_subject(
-            saved, uses_persisted_credential=bool(saved_refresh)
-        )
+        token_subject = _token_subject(saved)
         try:
             body = _post_refresh(control, refresh, workspace_id, token_subject)
         except CloudSessionError as exc:

engraphis/cloud_session.py

@@ -1,6 +1,6 @@
 """Sleep-time consolidation (episodic→semantic distillation).
 
-Some hosted memory systems offer "sleep-time compute"; the local-first equivalent is a
+Some systems ship "sleep-time compute" as a cloud service; the local-first equivalent is a
 background job the *user* schedules (cron / Windows Task Scheduler / a session hook):
 
     python -m scripts.consolidate --db engraphis.db --workspace acme

engraphis/core/consolidate.py

@@ -336,6 +336,6 @@
      graph view. dashboard.js fetches both on demand from graphRender(); see loadForceGraph()
      and loadGraphEngine(). scripts/externalize_dashboard_assets.py enforces both halves:
      they stay out of this file, and the lazy references still have to resolve. -->
-<script src="/static/dashboard.js?v=20260729-hub-materials"></script>
+<script src="/static/dashboard.js?v=20260728-reference-materials"></script>
 </body>
 </html>

engraphis/static/dashboard.js

@@ -7,7 +7,7 @@
 ships — so the number you get is about the product, not a bare index.
 
 What it measures — honestly: **retrieval** (evidence recall@k / hit@k), not
-end-to-end QA accuracy. Published LoCoMo/LongMemEval scores from external memory system/external memory system also hinge
+end-to-end QA accuracy. Published LoCoMo/LongMemEval scores from other systems also hinge
 on an answering LLM + judge; this harness isolates the part Engraphis owns, needs no
 API key, and states exactly that in the report. Add an answering model on top for a
 QA-accuracy number when you want one.