Skip to content

World's first open-source enterprise world model.

The design pages say what a call to the endpoint survives (#758)

WaylandYangca4678084da4
diff --git a/docs/design/chat-and-mcp.md b/docs/design/chat-and-mcp.md
index 57115e76..4ec9461f 100644
--- a/docs/design/chat-and-mcp.md
+++ b/docs/design/chat-and-mcp.md
@@ -13,7 +13,8 @@ as before; a tool's UI step goes to the stream; an empty turn is asked again onc
turn from an endpoint that ignored `required` is sent back once [0042]. A turn cannot end before a
tool has run; `no_evidence_needed` is the exit for a greeting or "make it shorter"; `STALL_NUDGE`
and `DONE` are gone [0042 d3]. The wire stays `LlmClient` behind `RigModel`: the read timeout,
-error bodies, the out-of-credit against rate-limit classification and cache logging; earlier
+error bodies, the classification of a failure as out of credit, rate limited, unavailable or nothing
+of the kind, and cache logging; earlier
entities become a `system` message right before the question; degradation to one-shot RAG happens
only on a 400 or 422 to the first request with tools [0042 d2].
diff --git a/docs/design/extraction.md b/docs/design/extraction.md
index 32ed79ae..a719b98e 100644
--- a/docs/design/extraction.md
+++ b/docs/design/extraction.md
@@ -48,6 +48,22 @@ a fifth of the chunks of a filing, tables included, and the bullets of a release
company as their subject without it. These are contract rules, not server checks: the shape
checks below look at structure, never at vocabulary.
+**What a call to the endpoint survives** [#757]. Extraction, both aligners and time resolution
+share one retry: a failure that fixes itself is waited out (five tries, backing off to a minute,
+honouring `Retry-After`), anything else is raised. Four kinds fix themselves: the endpoint rate
+limits, it answers 408, 502, 503 or 504, the request never reaches it, or its stream stops in the
+middle of an answer. A 500 does not, because that can be the endpoint's own bug, and neither does a
+read timeout: the ceiling is 300 seconds of silence and retrying that five times is 25 minutes.
+
+The call is streamed and the answer assembled, so that the timeout measures silence rather than
+thinking. Not streamed, the first byte waits for the whole answer, so with reasoning on the model's
+thinking counts as silence: one dense passage of a filing crossed 300 seconds three times and was
+declared dead, while the same passage streamed has bytes at 2.6 seconds and finishes at 231. It is
+not faster, it is only honest about what silence means. A stream that ends without `[DONE]` or a
+`finish_reason` is a failure and not a short answer, because half an answer is a valid string and
+the statements it lost would go unmentioned. A chunk that still cannot be extracted is a drop row
+(`chunk_unextracted`), so the gap survives the document being marked done.
+
**Server checks, each a drop reason in `extraction_drops`.** Every quote must occur in the chunk
(`quote_not_in_chunk`); every name in its quote (`name_not_in_text`); a time mention only when its
words occur in the statement's own quote (`time_not_in_quote`), because the model otherwise attaches
diff --git a/docs/design/ontology.md b/docs/design/ontology.md
index 85171c0d..07fec107 100644
--- a/docs/design/ontology.md
+++ b/docs/design/ontology.md
@@ -67,7 +67,8 @@ statement's subject is the property's subject, reverse when its object is) for t
bind. A signature the votes disagree on is `undecided` for the alignment queue of #725; one with no
fitting property is `none`, its statements stay in the open graph and it counts toward the
workbench's suggestions. Bindings live in `phrase_bindings` and go stale when the property they
-bound to changes or a property is added; a person's decision is never overwritten by the agent. On
+bound to changes or a property is added; a burst of ontology edits debounces into one run rather
+than one run each [#757]; a person's decision is never overwritten by the agent. On
the 25-document batch with a hand-written ontology of 14 classes and 28 properties, and 60 of
400 kind words bound, 423 signatures cover 861 statements: 36 bind (184 statements), 110 bind to
nothing, 1 splits the votes and 276 have no admissible property because an end is unbound; a