# Bounded completion of issues 61 and 62 The owner selected Luna as the economical runtime and approved bounded implementation closure after the complete V23 evaluation: 233/320 useful answers, 27/40 complete retained sets and 15 unsupported published answers. The original numerical gate fails and is not redefined by closure. See the [full report](README.md) and [failed coordinates](failure-ledger.json). Unmet coverage and accepted semantic errors continue as regression cases under #66. This describes source fidelity within a declared language/discourse scope, not a guarantee of world truth or complete interpretation. ## Issue 61: knowledge language and cross-language retention | Original requirement | Delivered behavior and evidence | Boundary carried forward | | ----------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Owner-controlled knowledge language, stable default and discoverability | `knowledge_language: "en"` is explicit; unset adds no language-target policy; context exposes the effective setting. Config, model-language and production integration tests cover the contract. | English is the supported target. Other generated target languages require their own evaluation. | | Preserve authored bytes, exact quotes, identities and taxonomy | Original support/frame quotations and source identity remain separate from generated English prose. Existing paths/names are not translated. Whole-tree checks and live archived-support/rebuild audits cover the distinction. | Exact archived evidence does not prove completeness or semantic correctness of a generated paraphrase. | | Generated prose across retention and maintenance | Shared policy checks cover generated retained prose, labels, summaries, observation/reflection/curation prose and new titles. Revision-language tests protect authored material and exact authorized transfers. | The frozen language matrix disables model index summaries/facts, graph expansion and maintenance; their policy/authority assertions are deterministic coverage. | | Exact caller candidates and typed failures | Supplied exact candidates remain model-free; policy attestation is explicit. Wrong-language/unavailable validation produces typed holds/degradation. | A fallible verifier can accept an unsupported translation. Runtime acceptance is not independent source entailment. | | Preserve attribution, modality, disposition, polarity and time | Source-frame validation, bounded qualifiers, repair/verification and typed eligibility operate through retention/retrieval/answers. The full V23 review judges accepted outputs against originals. | Retention omissions, source-meaning changes, false holds and unavailable verification remain measured limitations. | | Queries and answer language independent from source/storage | English/Russian queries exercise inferred and explicit views; English/Russian answer requests are separate from English knowledge policy. Exact evidence remains in its source language. | Query inference supports bounded constructions rather than all phrasing. New recognition fixes have separate post-baseline evidence. | | Policy changes, restart/rebuild and replay | Receipts bind the effective policy; an existing revision replays its recorded result. Tests cover changes without retroactive translation; live runs audit source bytes/support archives through restart/rebuild/replay. | No bulk translation or vault reprocessing is authorized by a policy change. | | Reviewed live evaluation with separate error/coverage counts | V23 has reviewed development/held-out sources, two repetitions, frozen profiles/hashes, incremental results and an independent reviewer. Useful coverage, safe read-only holds, language errors, source errors and availability remain separate. | V23 varies the content/wording of eleven scenario structures. Accepted-proposal transitions, late corrections and unresolved-reference resolution lack live coverage here. Incomplete/oversized frames and broader long-distance discourse have deterministic controls, not a comprehensive live generalization claim. | ## Issue 62: qualification of ordinary Markdown | Original requirement | Delivered behavior and evidence | Boundary carried forward | | ---------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------- | | Keep nonfactual propositions out of factual downstream use | Rebuildable passage qualifications propagate into fact/graph derivation, recall/context/answers and maintenance eligibility. Ordinary live controls distinguish qualified inspection from factual exclusion. | The scanner recognizes bounded English/Russian conventions; arbitrary implicit discourse may escape classification. | | Preserve ordinary factual usefulness and deciding context | Mixed sections, ATX/setext headings, nested quotes, speaker turns and fenced examples have deterministic controls. Indented/empty headings preserve boundaries and exact frame bytes. | Bounded oversized frames remain inspectable but are withheld from answers with typed degradation. | | Invalidate stale qualification after edits and rebuild | Reads recompute from current bytes; projection upgrades invalidate derived state. Direct editor changes, restart/rebuild, index and fact/graph tests cover the path. | A derived reading remains separate from source authority. | | Leave authored Markdown and file set unchanged | Whole-tree snapshots and isolated ordinary/retained runs check index/rebuild/read stability. Read/inspect preserve original-language source text. | Opt-in writes remain separate owner-controlled features. | | English/Russian evaluation with English knowledge policy | The independently reviewed ordinary supplement has 47/48 useful answers, 24/24 useful contexts and 18/18 safe factual-negative controls across two repetitions. | One useful answer was withheld; four ordinary cases do not establish unlimited Markdown or multilingual coverage. | ## Follow-up ownership Issue #66 should retain the V23 failed coordinates and original sources, including extraction/verification omissions, source-unfaithful accepted output, safe but unavailable responses and false holds. The repaired rejected-proposal view cases remain routing regression controls. Compare future economical models under identical pipeline settings, keeping these source regressions separate from the new longitudinal stages #66 still needs to implement. Measure original error creation, later propagation and correction/retraction separately; this one-cycle baseline does not deliver that longitudinal work. Closing these issues records the agreed bounded delivery and remaining limits. PR #73 still requires its normal review/merge; issue closure does not merge or deploy code.