Recover joint planning dependencies and preserve valid hypothesis siblings

This commit is contained in:
Nico committed 2026-10-20 08:12:23 +02:00
1 parent 83ef9e6d6c
commit c3a4690a55
15 files changed
+237 -10

No files matched your search

+25
View File
@@ -97,3 +97,28 @@ Sperren erhaltener Vorschläge, falsch geformte Modellfelder, Abbruch/Wiederaufn
Import in einen neuen Workspace ab. Ein Fall des unveränderten kuratierten Fehlerkorpus
prüft die Lokalisierung der tatsächlich beobachteten Fehler. Kein bezahlter API-Test
wurde ausgeführt; Kosten und semantische Qualität sind im nächsten Realtest zu messen.
## Dependency recovery revision 1
New joint contexts freeze `joint_generation.recovery_version = 1`. Contexts without
this marker retain the earlier projection and request behavior for replay.
Valid explicit premise references are unioned into their proposal dependency lists
on a copied projection, never by editing the saved LLM answer. Only existing readable
source IDs and exact existing interpretation statement pointers are eligible. Unknown
references and grounded unread references still fail validation. No premise is promoted
from unverified to grounded, and no claim is deleted or rewritten.
Each candidate is checked independently before freezing `joint_retained` in the
hypothesis context. Initial candidates based on an already accepted interpretation
can be retained directly; after interpretation repair, explicit dependency review is
still required. Invalid candidates remain unlocked. Duplicate candidates are not locked
against a necessary diversity correction. A followup corrects only unlocked positions;
locked proposals are enforced with schema constants and local equality checks. Final
publication still validates the full set for count, grounding, primary focus and diversity.
The joint prompt repeats the required `described_behavior.verification` value. Missing
verification is not silently filled: the scoped statement correction must provide it.
The original 2026-10-20 replies and a bounded validation context are preserved in the
curated rejection corpus. Offline regression checks hashes, reference union, sibling
preservation and rejection of invented references. No paid API request was needed.