Status: Living Document Date: 2026-09-30 (includes twelve-exchange full-cycle optimization) Applies to: the whole repository Depends on: Agentic Technique Master.md — the style guide governing this tree

This tree follows the folder system in the style guide: each numbered folder is a prerequisite for the folders after it. A change is only "real" when backed by at least one of the three lanes — tests, documentation, code.

Local reader screen: Llama 3.2 3B Instruct Q4_K_M on the RTX 2070 SUPER answered 23/24 saved streaming packets correctly at a mean 1.68 s, compared with the historical Haiku control's 24/24 at 3.95 s. Six summary requests averaged 0.63 s, but manual review found factual corruption. This is a resident replay, not a full-pipeline benchmark; cold startup plus first generation took 98.74 s. Model defaults remain unchanged. See the local Llama section of Research Log 262.

Shared chat I/O: New chat exchanges and newly prepared engineering runs capture inputs, outputs, and tool results through one durable interface. Recall packets link their triggering input to exact original memory spans; completed exchanges feed those original chunks into the existing Hebbian co-access graph. See Research Log 260 for the interface, tested lifecycle, and integration limits. Every new input, output, and tool result is saved immediately. The live chat binding now starts background preparation after each completed exchange, with three bounded jobs and ordered publication. Recall still runs for each question against the last committed index. The reader receives up to twelve recent exchanges plus the current input, with an 8,192-token recent-context budget. Whole committed exchanges are trimmed first; uncommitted messages stay available even if they exceed that budget. Attention-guided groups seal at eight exchanges, preserving their parent summaries across later appends. Explicit batch settings retain the prior six-exchange path for comparisons. New fragments under 192 tokens (1.5 times the 128-token raw-summary limit) retain their exact text without a raw summarizer call. Flush and close drain pending work and finalize partial hierarchy groups; original-source learning waits for publication. The durable queue survives restart, and eager ingestion remains available for diagnostics. See the streaming section of Research Log 262 below.

Streaming continuation check: On one existing 1.13M-token memory, 24 consecutive exchanges scored 24/24, with memory at most three completed exchanges behind inside the twelve-exchange context window. All 96 new events and 24 learning updates persisted and passed cold-reopen checks. Mean reply was 4.83 s, maximum 9.38 s; the final drain was 7.87 s, for 125.55 s across all 24 exchanges, excluding startup. The queue kept up, but this sample does not establish faster replies than the prior batch run. It used the Haiku reader override and C: active storage. 142 regression tests pass; see Research Log 262 for the changed grouping policy, timing limits, and complete evidence.

Chat I/O million-token regression: One existing 1.115M-token history and 100 fresh answers scored 94/100, matching the cap-8 baseline. All source pointers, input/output links, and 100 Hebbian updates survived reopening. Median answer time was 4.670 s, or 5.557 s including the remaining indexing and learning drain. Retrieval used the fixed historical snapshot; continuous native hierarchy refresh was not exercised. See Research Log 261.

Live refresh latency: Exercising the full summary/hierarchy pipeline exposed large overhead in the subprocess adapter. A resident replacement reduced an analogous follow-up from 103.43 s to 31.72 s, with original input/output/tool recall and restart equality verified. Full refresh is still above the five-second target. The clean million-token re-ingest control was canceled; no equivalence claim is made for that unfinished run. See Research Log 262 for timing boundaries, the interrupted worker, tests, and remaining costs.

The follow-up now reuses historical index entries and writes changed native and parent rows in one transaction. On a cached 1.126M-token update, combined index work fell from 2.377 s to 0.264 s; publication took 2.556 s with fresh-build and separate-process receipt equality. Qwen's embedding table stays in CPU RAM while its transformer layers and BGE stay on the GPU, removing whole-model transfers in the bounded placement test. A subsequent complete eager turn on 1.13M tokens returned the correct answer in 16.78 s and completed ingestion and learning in 49.41 s, excluding startup. The nonblocking six-exchange queue and short-text bypass pass 52 focused tests. A revised 12-exchange run returned 12/12 correct answers (eight recalls, four acknowledgements) in 65.93 s, averaging 5.46 s each. Both ingestion batches and all twelve learning updates completed; new original-source facts were recalled after the drain. Background summarization took 653.97 s, and the full cycle 757.57 s in that baseline. Authenticated reuse of recalled evidence and three concurrent compiler calls first reduced that full cycle to 125.26 s. The latest retained configuration uses shorter generated parents, exact attributed summary reuse, parallel hierarchy compilation, concurrent raw indexing, and committed raw snapshots for nonblocking recall. The last push prepares reusable summaries after exchanges three and five while keeping publication and learning at six exchanges. A bounded exact token-count cache reduces repeated validation work, and optional preparation yields when a full batch is ready. With the authorized Haiku reader override, the best measured twelve-exchange cycle is 52.27 s; the latest configuration takes 54.93 s, with 12/12 correct answers, all twelve learning updates, and original-source hydration verified. The latest run uses C: temporary storage for its active memory and chat journal, following large write stalls on F: project storage; the normal project store was not relocated. Startup is excluded; the complete final ingestion/learning drain is included. The sub-50 target remains unmet. Latest mean answer latency is 3.68 s, maximum 4.95 s; answers finish in 44.64 s, followed by 10.29 s of draining. Merge calls fall from 28 to seven; repeated raw summarization remains at zero. 135 focused tests pass. Exact JSON fences are accepted for scoring and captured outputs remain unchanged. This is one bounded continuation, not a new 100-question or engineering-quality result, and the default reader remains Sol outside the evaluation override. Two local quantized LFM summarizers were tested and not promoted because of factual-association errors or missing routing details; the current raw-summary model remains unchanged. Details and artifacts are in Research Log 262.

Ten-session battery complete: Ten additional histories of 1.06–1.16M eligible raw tokens, with 100 new questions each, scored 913/1,000 (91.3%) under the unchanged grader. Warm median was 4.783 s, mean 5.238 s, p95 8.935 s; mean answer input was 1,500 tokens, and retrieval/prompt preparation averaged 0.282 s. All 1,000 exact raw-evidence checks passed. Every history used fresh application ingestion and a separate-process reopen with cached summary/attention compilation. Two sessions reached 95%; the aggregate target remains unmet. All workers are closed. See Research Log 244.

The status report records the handoff; Analysis 35 holds the failure findings, source examples, and repair priorities.

Single-history routing ablation: Removing the active retrieval additions reduced the same 100-question score from 94 to 73 on a 1.115M-token history, while saving 14.57% of input tokens. See Research Log 246. Removing downstream assistant filtering and user-first layout from both variants still yields 93 versus 73; see the compensation check, Log 247.

Priority-1 trace and repair: The three named missing-turn cases were traced to routing (H8 Q83, H6 Q51) and to hydration order (H1 Q66), where late user additions lost the shared budget to assistant sections the projection then discarded. A new sealed stage, user completion routing, hydrates user routes before assistant context and completes routed conversations' user turns under a bounded cap. A provider-free replay of all 1,000 sealed packets raises recorded-support coverage from 972 to 989 at cap 8 with every prior span preserved. A bounded answer run with the unchanged reader, model and grader recovers 27 of the 87 misses while 85 of 87 matched controls still pass, at about 20% more prompt tokens. The full-campaign effect is not yet measured: the sampled control loss rate would cost roughly 21 of the 913 passing answers if it held. See Research Log 245.

Cap-8 full-history check: All 100 questions on the same 1.115M-token comparison history score 94/100, unchanged from baseline, while recorded-support coverage rises from 97 to 100. Warm median is 4.345 s and mean input is 1,762 tokens (+18.69%). See Research Log 248.

Cap-8 integrated: Public native retrieval now uses the tested cap-8 defaults when its parent-summary index is installed. A reader-only comparison on the same 100 saved packets scores 93/100 versus 94/100, with 190 more input tokens per answer; v7 remains the reader. See Research Log 249.

Answer-model comparison: The available Claude Code Opus route scores 91/100 versus Sol's 94/100 on identical cap-8/v7 prompts. Both omit Strava; some grade changes penalize source-supported details. The gateway exceeds the requested output limit on one passing answer. Sol remains the default. See Research Log 250.

Metadata-hint soft test: Twelve fresh paired questions receive 9/12 plain versus 10/12 hinted source-review labels; all six controls pass in both arms. The hinted answer recovers Strava, with about 58% more input tokens. This remains an isolated prototype; see Research Log 251.

DSPy hint optimization: A bounded COPRO search followed by twelve held-out question pairs yields 11/12 plain versus 12/12 in the DSPy arm, including two fallbacks for invalid hints. Reader input rises about 8%, but generating hints adds a separate model call. The lone graded gain is a small wording distinction; see Research Log 252.

Expanded DSPy test: All 100 saved questions now have fresh paired answers. Among the 82 newly added questions, both arms score 74/78 assessable pairs, with two recoveries and two regressions. Across all 94 questions outside training, labels are 83/90 plain versus 86/90 hinted; three invalid reviews and one ambiguous question remain unresolved. All guides validate after the status-alias fix, but the extra generation call has not demonstrated a consistent benefit. See Research Log 253.

Original-context control: Twenty selected diagnostic questions receive fresh answers using memory or complete original source conversations, with the same Sol/v7 reader. Labels are 16 correct / 3 incorrect / 1 ambiguous for memory versus 13 / 6 / 1 for original context. All three shared failure labels persist; no memory failure is rescued. Several judgments remain debatable. This uses complete selected conversations, not the entire 1M-token history. Mean answer input is 1,865 versus 10,669 tokens. See Research Log 254.

Full 100-question context control: The same Sol/v7 reader scores 95 correct, 4 incorrect, 1 ambiguous with memory versus 87 / 12 / 1 with complete selected raw conversations. Three of four fresh memory misses (75%) also fail with raw context. Following the earlier plain-memory run's eight misses instead gives four of eight (50%); the denominators are different. The control retains the previous twenty pairs and adds eighty, with no new ingestion. Mean answer input is 1,762 versus 9,288 tokens. This is a source-conversation comparison, not a full-1M prompt, and several failure labels remain debatable. See Research Log 255.

All 87 campaign misses checked: The raw-context control now covers every miss across all ten histories. Raw answers receive 57 correct, 21 incorrect (24.1% of the 87) and 9 unresolved source-review labels. The original memory answers are reclassified as 40 correct, 38 incorrect and 9 unresolved; 13 of those 38 (34.2%) also fail with raw context, while 25 recover. Eighteen of the 25 recoveries come from packets with incomplete recorded support. Some review judgments are flawed or debatable, and these diagnostics do not replace the 913/1,000 benchmark. All 213 recorded support quotes are present in the raw inputs. See Research Log 256. The updated Analysis 35 classifies all 87 individually: among the 38 still rejected, 24 have material evidence-delivery gaps, 12 are reader-error candidates, and two have identified grading/question defects. The classification includes an auditable 87-row CSV.

Lightweight hints: A fresh saved-packet comparison improves 12 reader-error candidates from 2/12 to 5/12, with three recoveries and no binary regressions; all 18 historical passing controls pass in both arms. Four evidence-gap cases and two grading diagnostics were separated before calls. Compact labels from stored summaries add 245 input tokens (16.07%) and no hint-generation call. This selected diagnostic uses original pre-cap-8 packets and remains outside production. See Research Log 257.

Scoring bands saved: The ten-history QA campaign averages 95.3% after source-review corrections (913 original passes plus 40 accepted flagged answers). The original automated score remains 91.3%; nine reviews remain unresolved. See the scoring-bands note for the per-history arithmetic and scope.

Engineering/research battery complete: The local notes archive supplied 20 artifact tasks from ten session families: ten engineering and ten research, with four development and sixteen validation cases. Exact chronological prefixes span 4,189–98,809 tokens. All twenty source boundaries and eighty rubric criteria pass audit. All 14 generated code implementations pass their own tests, and all 12 fixed-interface implementations pass independent checks. Memory uses 85.7% fewer actor input tokens across eleven completed matched pairs, excluding ingestion and grading. Both arms produce artifacts on 15/20 cases; memory finishes 14, full context 15. Research exposes ingestion, hydration and definition-routing gaps; seven actor requests time out. See Research Log 258 for results and the canonical battery guide for the protocol.

Browsing follow-up: Public search and page opening are now optional matched tools for both evaluation arms. A separate three-case research follow-up checks whether they help the completed memory omissions. All six actors finished but none browsed or recalled; all three observed memory omissions persisted. The original twenty-pair results remain unchanged. See Research Log 259.

Engineering artifact comparison: The original longer session's implementation and the preserved memory-generated code were checked with the same behavioral assertions: 8/8 original, 7/8 memory, with only the MCP energy display failing in the memory artifact. It plausibly substitutes for the core implementation and fixes an evaluation-side reheating issue. It has not reproduced the original session's real-data measurements, which the replay's tool surface did not expose. Latency is assumed equal at the earlier seconds-level rate. No new model calls were needed. See Research Log 243.

The separately attempted fresh full-context control failed at the gateway before any new engineering answer and is not used to judge code quality. That attempt is closed; the user's subsequent request was to compare the saved artifacts. See Research Log 242.

Real engineering-session result: All 8 original prompts were replayed on the historical checkout, producing 25 changed files. Final validation found 237 regression checks passed, 73 additional checks passed, and 6/8 frozen independent checks passed. Earlier completed work was recovered through memory, with a bounded current-user working conversation. The final memory save failed: a separate process reopened 481 of 524 journaled turns, leaving 43 final events outside application memory. Live retrieval averaged 2.460 s, but successful ingestion batches took 176–554 s. This demonstrates useful engineering work; reliable, interactive context replacement remains unproven. All workers are closed. See Research Log 240.

Engineering token savings: Recorded coding and summary generation used 4.616M input tokens, versus an estimated 49.821M for resending the same observed history at every response, excluding internal memory receipts: 90.73% fewer input tokens. This includes recorded summary overhead and retries; it is a local-token estimate, with no full-context control. See Research Log 241.

Previous engineering result: The bounded replay stopped with 3/8 prompts completed, zero code changes and zero test runs. A repaired retrieval policy was checked for five more coding actions; all remained reads/searches. Both workers are closed. This setup has not demonstrated replacement of context during real engineering work. The five-action check required 88 generation calls, with mean per-action ingestion of 152.88 s, retrieval of 1.84 s and coding generation of 19.70 s. That failure motivated the bounded working conversation evaluated above. See Research Log 239.

Completed question-answer evaluation: Five additional 100-question tests scored 93, 92, 90, 96 and 91, totaling 462/500 (92.4%). Each history contained over 1.1M eligible raw tokens and followed normal ingestion plus a separate-process reopen. Combined warm median was 4.469 s, p95 6.893 s, and mean input 1,537 tokens; all 500 exact raw evidence checks passed. The subsequent eight-prompt raw-transcript build replay changed 16 isolated files: 240 regression checks passed, while four of eight independent behavioral checks failed. It exposed a user-requirement continuity gap, missing reinforcement during rapid turns, and an MCP energy-display bug. Those runs are complete and their workers closed. See Research Log 238.

Previous completed result: 96/100 on the original 100 questions over one normally ingested and reopened 1,098,417-token history. User-spine routing and summary-only Qwen artifacts are unchanged; the answer context retains exact user sections and omits whole assistant-only sections. Warm median is 4.947 s versus 4.356 s for matched direct API calls; p95 is 7.638 s and 52/100 memory answers finish under five seconds. Cold setup is 30.308 s. All 100 packets pass raw reconstruction, and provider-free grading replay is identical. The original fixed-benchmark threshold is met. This exposed development result has known semantic-grader limitations and does not establish held-out accuracy. See Research Log 237. The chronological record below retains earlier, superseded results.

Prior completed scope: exactly 100 questions on one already ingested ~1M-token history. The user explicitly rejected 100 histories. The r3 100-history controller and admission worker are stopped; no benchmark answers were released by that run. Do not resume them. The corrected run reuses the existing 1,098,417-token memory, with 100 source-grounded questions locked before answers. It loads one namespace once and uses the frozen retrieval method. Questions and reference answers do not enter ingestion. Follow Research Log 218 for the completed run and audit. The candidate scored 71/100, versus 64/100 for user-first retrieval. Warm median total was 3.700 s, versus 3.677 s for the matched API control; p95 was 6.478 s and 83/100 answers finished below 5 s. All 200 memory packets passed exact raw reconstruction. The 95% target remains unmet. These are generated questions over real transcripts, not an official LongMemEval score. Cold load and ingestion are excluded from warm latency. The completed reader-completeness comparison scored 81/100, a ten-point gain on the same questions and identical retrieved packets under unchanged grading. Its warm median total was 3.695 s, versus 3.780 s for fresh matched API controls; all 100 memory packets passed raw reconstruction. The goal remains unmet. Follow Research Log 219. A separate prediction-blind review of all 100 evaluation items is complete. It flags question/reference defects without changing either score; two questions have confirmed ambiguity among real source statements. Follow Research Log 220 for those limitations and the next summary-routing work on this same history. The completed semantic-seed/two-level-parent comparison scored 84/100, using a 2,048-token raw packet with the v6 reader and grading unchanged. Warm median total was 3.985 s, versus 3.869 s for the matched API. All 100 packets passed raw reconstruction; 97 contain every recorded support quote. That coverage is not an accuracy score, and the 95% goal remains unmet. No new ingestion occurred. Follow Research Log 221 for the result, sealed policy, failures and next work on the same history. Lifecycle boundary of the earlier scores: Those runs exercised cached ingestion artifacts loaded into the native memory components, fresh question retrieval, exact hydration, and answer grading. They do not exercise MemoryCondenser.ingest or the normal application's persistence/restart path. They establish cached-memory answer accuracy, not full application lifecycle validation. The latest audit reproduced all 100 packets and 2,656 original raw spans with zero model calls. See Research Log 222. Application integration is now verified: normal ingestion persisted all 5,357 raw turns, and a new process reopened the application and reproduced all 100 baseline packets and 2,656 exact spans. Eleven lifecycle tests, 138 existing condenser tests and 17 admission/presentation checks pass. Follow Research Log 223. The completed conversation-order comparison scored 85/100 through that reopened application, with unchanged evidence and reader. Warm median was 3.964 s versus 3.688 s for matched API controls; p95 was 6.331 s and 79/100 finished below five seconds. All 100 packets and 2,656 raw spans passed audit. Five gains and four losses do not establish a reliable layout improvement. The 95% target remains unmet. See Research Log 224. The completed v7 reader comparison scored 87/100 on that same application memory, changing only system instructions. Warm median was 4.465 s versus 3.962 s for matched API controls; p95 was 7.015 s and 71/100 finished below five seconds. All 100 packets and 2,656 raw spans passed audit. The 95% goal remains unmet. Follow Research Log 225. A provider-free packet-width diagnostic retained all recorded support on the same 97 questions while reducing median conversations from 14 to 4. All 200 candidate packets passed raw reconstruction. This is coverage, not accuracy; the next answer comparison tests 8 direct matches with the reader unchanged. See Research Log 226. The completed width-8 comparison scored 93/100 through the reopened application, with the same reader/model and unchanged grading. Warm median was 4.298 s versus 4.251 s for matched API controls; p95 was 6.741 s and 73/100 finished below five seconds. All 100 packets and 1,430 exact spans passed audit. The seven misses include two clear reader omissions, one routing miss and evaluation defects. The 95% target remains unmet. Follow Research Log 227. The completed Sol comparison also scored 93/100 on identical user-spine packets, with three gains and three losses. Warm median was 4.615 s, versus 4.383 s for matched direct Sol; p95 was 8.500 s. All 100 packets and 1,430 spans passed raw audit. The model switch did not improve the total score. Follow Research Log 228. Two complete lexical-summary diagnostics recovered the remaining empty-support question but exposed competition in parent expansion. Neither is promoted. Next work appends one lexical address after the unchanged semantic/parent routes, preserving the existing packet before using spare capacity. No reingestion is needed. Follow Research Log 229. The append-only implementation now preserves every prior route and raw section across all 100 questions. Recorded-support coverage rises to 98/100, with no losses; all 100 packets and 1,439 spans pass independent raw audit. Coverage is not accuracy. The completed 100-answer Terra comparison scored 87/100, with median 4.945 s versus 4.383 s for matched API controls; p95 was 7.347 s. Seven of eight regressions occurred on byte-identical prompts to the prior run, exposing answer/grading variability. All 100 packets passed raw audit. The change is not promoted on accuracy; the best completed result remains 93/100. Follow Research Log 230. The next comparison groups exact user statements before assistant context within each conversation, preserving all evidence, retrieval and reader instructions. The completed comparison scored 93/100, with median 4.651 s versus 4.581 s for matched API controls, and p95 7.084 s. All 100 packets and 1,439 spans passed independent raw verification. It ties the best score; 95% remains unmet. Two remaining support gaps trace to parent-seed selection and a parent group that omits an earlier user statement. A Sonnet readiness request was rejected by the upstream account's insufficient credit, so this comparison uses the available Terra model. Follow Research Log 231. A new persisted index of 522 complete parent user summaries now supplies up to two additional exact user atoms after the unchanged route. All original evidence survives; recorded-support coverage rises to 99/100, with all 100 packets and 1,459 spans passing raw audit. The completed answer comparison scores 93/100, with warm median 4.246 s versus 3.872 s for matched API controls, and p95 6.518 s. The saved-grade replay confirms the result with zero new calls. Coverage improved but answer accuracy still ties the best result; 95% remains unmet. Remaining misses include reader omissions despite complete user-spine evidence, conversation-scope errors and documented question/reference defects. No history was reingested and no new Qwen pass was needed. Follow Research Log 232. The shorter extractive reader completed at 90/100, versus 93/100 on identical v7 packets, and is not promoted. Its warm median is 4.025 s versus 3.624 s for matched API controls, with p95 7.406 s. All 100 packets and 1,459 spans pass raw audit; the saved-grade replay is identical. Twenty-seven checks pass. A separate audit verifies all 520 source-body hierarchies against 525 stored user-summary attention windows and exact atomic addresses. The best accuracy remains 93/100; 95% is unmet. Follow Research Log 233. The completed paired-answer diagnostic scores the saved API controls at 93/100, matching memory, with three different passes and three different failures. All 35 identical predictions receive matching grades; inspection of the six disagreements finds both reader omissions and semantic grading defects. No score is revised or combined. Follow Research Log 234. Sol on the current user-first packets and v7 reader scores 94/100, a new best under unchanged grading. Warm median is 3.854 s versus 3.532 s for matched direct Sol, with p95 5.021 s; 93/100 finish under five seconds. All 100 packets and 1,459 spans pass raw audit; the saved-grade replay is identical. The sole new failed grade contradicts an explicit user statement about prior open-mic attendance. Its grade remains unchanged after the separate source review; the original 95% acceptance gate is still unmet. Twenty-eight checks pass. Follow Research Log 235. The separate source-grounded review of all 100 saved Sol answers is complete and reproduced with 100 cache hits and zero new calls. It returned 87 correct, seven incorrect, two ambiguous and four invalid reviews. These are diagnostic labels, not a replacement score. Source inspection finds both grading defects and answer attribution/qualification errors that the original grader passed; the new reviewer also makes mistakes. All 100 items and original grades remain intact. The acceptance gate remains unchanged pending the user's preference. Follow Research Log 236. The completed assistant-context ablation scores 96/100 with the same Sol reader, routing, hydration, questions and original grader. All user sections remain exact; 314 whole assistant-only sections are omitted. Warm median is 4.947 s versus 4.356 s for matched API calls; p95 is 7.638 s and 52/100 finish under five seconds. Twenty-nine checks and the full raw/requirement audit pass. The original benchmark gate is met; grader limitations and lack of held-out generalization remain explicit. Follow Research Log 237.

Superseded 100-history preparation — historical record only: The user rejected full100-scale preparation during design iteration. That iteration is now complete on one cached 1,098,417-token history: the latest parent-context candidate scored 8/8 on the small development set, including denial, preference update and cross-conversation ordering. This establishes no population accuracy. Its median total was 4.783 s versus 4.117 s for the matched API control. A separate offline profile measured 0.172 s median packet preparation with identical packets. The candidate fixes the reader, 1,024-token cap, zero protected prefix and bounded parent context. No further pilot expansion is required before broad evaluation. The user accepted the observed sub-five-second response time and asked to focus on accuracy. Stop serving-latency tuning; retain accuracy and timing measurement together in the broad run. The reported 4.783 s is a median, not a tail guarantee. The original controller remains stopped: it binds the old router and repeatedly replays compilation ancestry. The active continuation completes the remaining corpus from existing caches and already binds the broad runner to the frozen candidate. Source summaries cover all 31,166 bodies; existing manifests contain parent trees for 13,768, leaving 17,398. This is a cache inventory, not a new full admission check. See Research Log 212. The 51 missing exchange bodies are complete, with all 707 raw spans independently verified. All 17,642 remaining attention windows are now cached. The parent compiler has completed and validated the combined 31,166-body cache, including all 17,398 missing parent hierarchies. Vector encoding also finished successfully: 323,124 rows, including 142,846 reused and 180,278 newly encoded summaries. The controller has started full100 population admission and packet preparation. Follow Research Log 217 for the current process and completed cache bindings. The first continuation stopped after 15,616 of 17,398 missing parent bodies. Its saved checkpoints and journals were authenticated, and the replacement controller found one missing checksum in the unfinished batch. That body was reproduced byte-for-byte from cached summaries and attention; only its missing checksum was restored, with zero new model calls. The r3 controller now resumes the remaining work with the same frozen evaluation. Research Log 216 records that recovery; the original interruption cause is unknown. The new broad runner binds the frozen candidate: 100 questions, two scored methods and matched candidate API controls. Eight focused runner checks pass, in addition to ten remaining-parent compiler and reader checks. The real answer run still awaits population admission and packet preparation. Follow Research Log 214. A separate result audit is prepared to recompute final scores and timings and verify every served raw excerpt against its original occurrence. Its nine checks pass; run it after the real report exists. See Research Log 215. The 129-token label repair remains verified at 22 tokens; its 74 software checks establish no answer-accuracy score. No native full100 result exists. See Research Log 208.

Historical first single-history design check: One history contains 1,098,417 actual eligible text tokens across 520 unique bodies. Its exchange, attention, parent and vector data are now cached for further iteration. Five fresh answers and two judgments completed. Flat scored 1/1 and the candidate 0/1, but all four evidence-bearing prompts were identical. The router consulted exchange leaves whose atoms were already in the direct shortlist, so it added no candidates before hydration. The packet also used 3,065 of 3,072 context tokens. The identical-prompt controls returned the opposite answers. This exposes ineffective context expansion and answer variability; it demonstrates no hierarchy advantage or population accuracy. Those findings led to the later cached-history checks and candidate freeze above. The original 100-history controller remains stopped; the current continuation is described above. See Research Log 209.

Follow-up on that same cached history: User-first packet ordering and bounded parent context each answered correctly on both repetitions; the old packet answered correctly once out of two. Median total times were 4.41 s, 5.30 s and 4.91 s respectively. These are repetitions of one exposed question, not population accuracy. Both revised packets retained 20 user sections versus 13 previously; parent context added no new evidence beyond user-first ordering on this question. Keep that hierarchy option experimental. Fourteen routing checks pass, and all eight streams and six judgments completed. See Research Log 210.

Prior packet-budget finding: On six frozen questions in that same history, user-first ordering scored 6/6 at 3,072 context tokens, while parent context scored 5/6. Reducing the cap to 1,024 lost a field-guide fact because a protected 561-token assistant reply crowded it out. The router now permits a zero-length protected prefix so all user evidence can be scheduled first. That repair passed a focused two-question check in both methods; the field-guide answer took 3.41 s versus 3.54 s for its matched API control. Twenty-three checks pass. These are one exposed official question plus manually authored design probes; neither 95% benchmark accuracy nor the joint latency target is established. The fresh complete design-set run and temporal/update coverage were subsequently completed in Research Log 212. See Research Log 211.

Active target: at least 95% answer accuracy over 1M-token memories, with end-to-end latency close to a matched direct API conversation. Accuracy and latency must pass together on the same implementation. The historical slow 95/100 and fast 73/100 results cannot be combined into a target pass. The earlier 95/100 already covered approximately 1M-token memories; it used the cumulative retrieval and answer-repair pipeline, without demonstrating API-like end-to-end query latency. See Research Log 132.

Latest full100: reject hierarchical routing, 8/100 versus flat 84/100. All 400 fresh answer streams completed; 200 logical judgments and the joint report replay with zero new calls. Hierarchy median total latency is 6.46 s versus 4.49 s for its identical-evidence API control and 3.94 s for short chat. Both target gates fail. All ten compiled hierarchies replay independently; the compiler and evaluation handoff have exited successfully. Saved-route tracing finds annotated support in 94/97 root shortlists, 62/97 root selections and only 9/97 final leaf selections. Exact hydration preserves those nine; the major loss occurs during attention pruning. Retain the flat packet while investigating the Qwen scoring policy. See Research Log 181.

Historical source preparation (broad execution now deferred). Further inspection confirms the previously documented pooled-history conflicts in flat misses. The existing M+S source bank keeps the 100 histories separate. All exceed 1M body tokens even after excluding generated source boundaries; the minimum through the question day is 1,007,016. The date-independent cache, batch runner and exact raw hydration integration pass 181 focused checks. Larger Terra batches produced all 50 summaries for the same probe fragments in three calls, with 54% less summed request time than the eight-call format. This is ingest work, not serving latency. Haiku returned no completions because its provider credit balance was too low. Full preparation is complete: 13,812 requests cover 323,143 fragments from all 31,166 bodies. The live handoff has released bounded Terra execution. The first Terra execution stopped after 2,735 completed batches and six transport failures. Its continuation accepted the first fresh batch and has resumed the remaining never-sent batches. Six interrupted batches now have verified recovery outputs, with their original uncertain requests preserved. The new summary-body store replays identically; real integration preserves all 76 tested occurrence-bound spans through the existing hydrator. Seven failed batches now have exact source subdivision repairs, preserving all 153 originally valid summaries. The initial admission snapshot incorporates those repairs into 1,669 complete bodies and 17,190 sections; incomplete bodies remain excluded. All 261 tested raw spans across the 23 bodies touched by the repairs hydrate exactly, with zero new model calls. Those 1,669 bodies now have 8,650 user-led exchanges and 1,695 cached local-Qwen attention windows. Both stages replay without model calls; the 14 bodies needing Qwen merges preserve all 104 tested raw spans and actual dates through hydration. Separate parent budgets published 1,526 body trees without new generation; another 147 local jobs completed all 1,669. The new serving adapter preserves all original attention partitions and raw addresses. Real occurrence rebinding preserves 940 tested raw spans, and all six oversized-exchange probes recover their selected original atom within budget. The native serving candidate retains direct atomic matches and adds context from attention-defined chunks. All 100 separate namespace checks pass with exact hydration of 400 selected spans; every incomplete namespace correctly rejects full-evaluation admission. All 17,146 unique summary embeddings are now complete. A real BGE resident probe preserves all baseline evidence and exactly hydrates 5,584 spans across both arms. Retrieval median is 0.105 s direct and 0.107 s with added chunk context, on partial histories of only 34k–91k tokens; this is not a 1M or API comparison. All 79 selected failed batches are now repaired, retaining 1,696 valid summaries. The prior body snapshot admits 5,481 complete bodies and 57,217 sections, preserving the earlier cache exactly. A further 40 batches are now repaired, retaining 875 valid original summaries. The successor store includes those repairs and six recovered batches, admitting 7,121 complete bodies and 74,327 sections; 10,641 batches remain pending in that snapshot. All 142 recovered sections hydrate exactly through actual namespaces. The 74,059-summary vector population is complete, reusing 57,020 vectors and computing 17,039 new ones. Expanded user-spine inputs are prepared; main ingestion continues. The expanded 57,020-summary embedding stage is complete. A diagnostic using all 100 benchmark questions finds annotated support available for only 21 of 97 annotated cases; direct retrieval and exact hydration retain it for 20. Attention context preserves that evidence but does not improve this count. These are partial 166k–260k-token histories, not an accuracy or latency pass. The old Qwen compiler stopped cleanly. The successor completed all initial trees; real checks across all 100 partial namespaces preserve baseline evidence and exactly hydrate 3,556 spans. Expanded exchange compilation stopped on two over-length summaries; separately recorded local recovery now admits both under the unchanged 128-token limit. Subsequent local compilation completed all 7,121 prepared body exchange sets, containing 37,226 exchanges. Attention is also complete for those bodies, with 7,217 cached summary windows; parent compilation is released. Another 71 failed source batches are fully repaired, retaining 1,541 valid original summaries. The next admitted store contains 11,438 bodies and 119,144 sections, preserving the previous store. Its 118,693-summary vector stage is prepared and queued. The native full100 joint runner passes 43 focused checks, including synthetic stream/judge orchestration. Actual admission rejects the incomplete corpus before any model or provider construction; no native full100 accuracy or API-latency result is available yet. Another 54 failed batches are fully repaired, preserving 1,161 valid original summaries. A malformed JSON batch separately recovered all 21 fragments in three calls. The successor assembler preserves that original failure record; 15 recovery/admission checks pass. Real serving verification of 6,725 expanded body hierarchies passes across all 100 partial histories, preserving baseline evidence and exactly hydrating 3,450 raw spans. The fresh corpus snapshot is complete with 13,468 bodies and 140,256 sections, preserving the earlier store. Its expanded exchange inputs are complete and reuse 317 authenticated Qwen summaries without new generation. A zero-generation pass completed 13,316 of its body exchange sets; 152 bodies still need merges. Matching attention, parent and vector stages are queued. Parent compilation on the earlier snapshot completed all 7,121 body hierarchies; 663 parent summaries are verified reusable. Another 34 rejected batches are now repaired, preserving 718 valid summaries; those repairs await a later body snapshot. The full100 loader supports the new producers and passes 44 focused checks. Real admission still rejects the incomplete corpus before model construction; the joint target remains unverified. The 118,693-summary vector build has completed and released expanded local-Qwen generation. The next vector population contains 139,694 summaries. A new live coordinator automatically repairs newly rejected public source batches and will assemble the full corpus after all original requests and repairs complete. It authenticated 286 existing repairs and six transport recoveries; its first cohort contains 25 further failures. Fourteen new orchestration checks pass, including both repair types followed by full-store assembly. The actual final corpus and native full100 evaluation remain incomplete. That first repair cohort has now completed all 25 batches in 19 calls. The complete-corpus handoff is live: it waits for the six existing producers, reuses their finished caches, compiles the full 31,166-body store in separate model processes, then runs and replays the unchanged joint full100 comparison. Fourteen additional pipeline checks pass. No full-corpus model stage or native full100 answer has been released yet. Exchanges are now complete for all 13,468 R6 bodies: 70,218 exchanges, with 437 new local jobs and exact atomic coverage. That process exited successfully and released attention preparation. The stage prepared 13,643 summary windows and completed all of them with summary-only local Qwen. The attention worker exited successfully and released expanded parent compilation. The first parent pass completed 12,794 of 13,468 body trees without generation; the remaining 674 bodies require summary merges in the bounded continuation. Further automatic cohorts repaired 12 batches in five calls, 15 batches in eight calls, and ten batches in six calls. Full source ingestion and the joint native evaluation remain incomplete. A read-only verification of the new expanded reader and exact hydration is queued after the R6 parent/vector jobs. It uses existing summaries and cached vectors across all 100 partial namespaces, with no model or answer calls. An early check now loads all 12,794 initial parent templates through the new reader and constructs one real partial namespace, exactly hydrating all four raw spans of a selected leaf without model calls. Full admission still rejects it. Further source compilation, repairs, complete hierarchy construction and a fresh joint full100 comparison remain pending. See Research Log 182 and Research Log 183 and Research Log 184 and Research Log 185 and Research Log 186 and Research Log 187 and Research Log 188 and Research Log 189 and Research Log 190 and Research Log 191 and Research Log 192 and Research Log 193 and Research Log 194 and Research Log 195 and Research Log 196 and Research Log 197 and Research Log 198 and Research Log 199 and Research Log 200 and Research Log 201 and Research Log 202.

Hierarchy restoration and bounded Qwen traversal implemented and evaluated. All ten memories have validated plans for 22,257 parents, preserving all 27,062 existing leaves. The resident routing/hydration integration passes 32 focused tests. The first real parent-summary run failed because the local gateway's qwen3-8b alias points to an unavailable qwen3-8b-gguf backend. The local fallback below has since completed all ten memories. The full100 result above rejects the current traversal policy. See Research Log 173.

Earlier evaluator admission checks. Sixteen additional checks include a complete synthetic 400-stream execution, 200 logical judgments and zero-call replay. Real admission correctly rejects the unfinished parents. A fresh readiness request confirms that the Qwen backend is still unavailable. Existing cached Qwen merges can recover some parent trees without new calls; provenance-bound cache reuse is the next compilation step. No real quality or latency gain is claimed. See Research Log 174.

Earlier cached parent recovery completed and replayed. It preserves 450 parents across 489 complete source trees with zero new calls. The remaining 4,316 sources need summary merges; no complete hierarchy is ready yet. Five new cache checks pass. The full local-Qwen fit test completed: the real summary passed validation in 12.18 seconds with a 4.70 GiB peak allocation. This is ingest generation, not query latency or an accuracy result. Query-time attention remains unchanged. See Research Log 175.

Local parent compiler added. Nine focused checks cover completion, replay, interrupted execution, attribution and bounded recovery. The eight-job real run completed: seven outputs passed immediately, one exceeded its summary budget, and 12 additional parents were published. Replay preserves the same progress with zero model calls. Its bounded repair passed, and the subsequent run supplied 90 reusable local summaries before a deliberate batch-policy switch. See Research Log 176.

Four-summary ingest batch measured 1.82x faster on the same four jobs. All outputs passed validation at 5.07 GiB peak allocation. That successor reused 90 local summaries and started offset 000 from 159 completed source trees and 345 parents. Session 65191 later stopped on the JSON error above. All 4,805 source trees fit the query depth limit. No new answer-accuracy or query-latency result is claimed. See Research Log 177. All ten hierarchies are complete and independently replayed. Offset 000 contains all 499 source trees, 2,245 parents and 2,744 unchanged original leaves; offset 010 contains all 464 source trees, 2,161 parents and 2,625 unchanged original leaves; offset 020 contains all 479 source trees, 2,214 parents and 2,693 unchanged original leaves; offset 030 contains all 473 source trees, 2,278 parents and 2,751 unchanged original leaves; offset 040 contains all 481 source trees, 2,256 parents and 2,737 unchanged original leaves; offset 050 contains all 480 source trees, 2,232 parents and 2,712 unchanged original leaves; offset 060 contains all 457 source trees, 2,155 parents and 2,612 unchanged original leaves; offset 070 contains all 488 source trees, 2,198 parents and 2,686 unchanged original leaves; offset 080 contains all 488 source trees, 2,220 parents and 2,708 unchanged original leaves; offset 090 contains all 496 source trees, 2,298 parents and 2,794 unchanged original leaves. All ten complete hierarchies reconstruct identically from saved outputs with zero model calls, using the same compilation method. The tenth replay ran after the timed evaluation exited. Both full100 target gates failed. A small manual check found a parent that flattened conflicting event status and another that dropped named entities. Exact leaf evidence is preserved; the impact of these summary losses on routing accuracy is still unmeasured. A separate recount verified every memory exceeds one million cl100k_base BPE tokens in complete raw turns, excluding chat framing; the minimum is 1,039,792 tokens. This check made no model calls.

Original full100 handoff stopped safely. Session 61096 stopped when its compiler terminated without a complete ten-memory population. It sent no evaluation calls. The successor handoff above retains the same complete-parent admission, idle-workspace check, 400 fresh answer streams, 200 logical judgments and zero-provider replay. Nine handoff checks pass. See Research Log 178.

Four-summary Terra generation comparison completed. All four outputs passed the same summary contract, but the concurrent batch took 15.57 s versus 13.64 s for the saved Qwen batch. This small sample does not demonstrate an ingest speed gain; Qwen remained the compiler backend. Replay verified the four responses with zero new calls. See Research Log 179.

Previous timed full100: control 81/100, grouped 77/100, supplemented 76/100. Both candidates are rejected with the same v2 reader throughout. All600 fresh streams completed and 151 Sol judge calls replay with zero new calls. Grouped median/p95 total time is 5.055/9.814 s; supplemented is 5.346/9.350 s. Both fail their matched API allowance and short-chat allowance. Retain the flat packet and v2 reader. Twenty-four focused checks pass.

Earlier implementation gap, now addressed: the serving indexes were leaf-only. They contained 27,062 leaves and zero parent summaries. Qwen shaped ingest partitions, while those query paths used BGE leaf retrieval. Parent compilation and the full100 hierarchical evaluation above now address that missing work; completing the architecture did not establish useful routing quality. See Research Log 172.

Previous timed full100: current control 80/100; grouped evidence/v4 reader 78/100. Reject the combined change. It preserves all 2,680 selected raw spans and cuts median context from 2,959 to 2,454 tokens, but loses more answers than it gains. Candidate median/p95 total latency is 4.729/8.247 s, versus 4.379/11.583 s for identical-evidence API and 3.221/4.556 s for short chat. Accuracy and short-chat latency fail. All400 streamed responses are verified; uniform judging recovered from a TLS failure and replays with 145 hits and zero calls. Retain the prior reader and retrieval path. See Research Log 171.

Earlier timed full100 comparison: as-of date cutoff 80/100, semantic seeds 74/100. All ten memories contain approximately 1.04M token proxies. All 500 fresh Terra responses sealed before 200 logical Sol judgments using 136 physical calls. Source admission and judgments replayed without new calls. The cutoff's median/p95 total latency was 6.580/9.366 s, versus 6.039/9.734 s for its identical-evidence API control and 5.083/6.353 s for short API chat. Both methods pass the provisional matched-evidence latency allowance, and both fail accuracy and short-chat latency. Neither joint target gate passes. Qwen still receives summaries only; query-time routing uses BGE summary addresses over Qwen attention-partitioned leaves, followed by exact raw hydration. See Research Log 166.

New bounded routing fix, fresh accuracy screen: 78/100 versus 76/100. Relative-day user evidence reservation changes three packets, recovering the smoker purchase and first-client contract answers with no paired regressions. All100 questions were answered afresh; 97 identical packets share fresh responses across arms. The 103 Terra answers and 103 Sol judgments replay with no calls. Thirty-six focused checks pass. This batch does not measure serving latency or replace the earlier independently timed 80/100 result. The prepared native corpus remains parked. See Research Log 169.

Global fine user-summary retrieval failed its full100 comparison. The fresh control scores 81/100, versus 72/100 at 3,072 context tokens and 71/100 at 2,048. The candidates recover some facts but lose more correct answers; neither is promoted. All 300 Terra responses and 158 Sol judgments replay with zero calls, and sixteen focused checks pass. Retain the prior retrieval path. These are batched accuracy results; the latest independently timed full100 at that point remained 80/100 and the joint target is still unmet. See Research Log 170.

Historical95 versus current80 is now paired: 78 both correct, 17 old-only, 2 current-only and 3 both wrong. Twelve of the seventeen losses are multi-session or temporal questions. The old cumulative answer-repair result did not prove matched API-like latency; the current independently timed pipeline has not regained its accuracy. See Research Log 167.

Reader and corpus audit complete. Sol on the same twenty miss packets gets four judge accepts, including one inconsistent judgment of essentially the same university answer. This is not an 84/100 result or a latency measurement. All100 native-history checks find every original turn in the pooled memory; twelve wrong answers already have every annotated support turn in their packet. Ninety-six packets also contain user statements from outside the question's original history. One verified case adds an incompatible Sophia meeting to the packet. Foreign membership alone does not prove a contradiction. The audit replays identically with no calls and does not change scores or production routes. The native-data continuation below prepares a separate complete corpus; the old scores and their source population remain preserved.

Complete same-history M+S corpus prepared and verified. The 100 separate memories contain 113.27M tokens in total, at least 1.076M each and at least 1.019M through each question day. The construction preserves native M and adds absent sessions from the same record's S history. It removes the known cross-history Sophia conflict without pooling different question histories. Repeated session IDs are preserved as separate occurrences. The verifier checks all 52,207 occurrences, 31,166 shared text bodies, raw-source identities and the separate evaluation plane. Earlier grouped-ID counts are superseded. Twenty-four focused checks pass. No model was called and no new score is claimed. This corpus is prepared and parked. Following the user's request for concrete progress, the immediate work returns to a bounded retrieval or reader improvement on the existing 80/100 population before further ingestion. See Research Log 168.

Full100 preparation: ten complete memories, 500 control requests prepared. All ten admissions verify under the common v11 method. The final memory has 5,624 admitted fragments, 2,794 attention leaves across 496 sources, and all three summary indexes, including 6,007 passage addresses. Both raw ingestion and compilation schedulers finished successfully. The token-accounting repair preserves every raw byte and distinguishes fragment counts from whole-turn counts. Earlier failed executions and their reservations remain preserved. The old r3 controls supplied the completed comparison below; their original answer campaign remains unexecuted.

Completed timed full100 comparison: the frozen experiment compares semantic seeds with the as-of date cutoff, keeping the same reader and raw budgets. It requires 500 fresh requests, both API latency controls, and all answers sealed before judging. All ten memories supply 500 prepared requests. The first namespace and offsets 050–090 used live query encoding; offsets 010–040 reuse authenticated live-diagnostic prompts for preflight only. Every timed memory request still recomputes embedding, routing, and hydration. Thirty-one evaluation checks pass. The actual runner validated all 500 requests, both fresh synthetic readiness probes passed, and all 500 responses completed. The final report verifies the same predictions and timings across all ten complete memories. See Research Log 166.

Automatic evaluation handoff completed the comparison. It verified both completed dependencies, prepared offsets 080/090, and released the unchanged experiment after readiness passed. Twenty focused checks pass. Failures preserve reservations and stop without automatic retries. All 500 answers sealed before any judging. Session 13811 exited successfully; its completion explicitly records that the target gate failed. Preserve the frozen runtime. The handoff protocol is documented in Research Log 163. Ninth-memory admission is recorded in Research Log 164.

Earlier date-aware routing diagnostic: the separate as-of candidate reproduces all fifty semantic-seed controls, removes all 240 future raw spans, and adds 273 eligible spans. The missing April 21 tomato-planting statement now reaches the gardening packet. All fourteen packets without future spans are unchanged; the other 36 change, displacing eleven previously included eligible user spans. A separate relative-date hint changes only one packet and is not needed for the planting-statement recovery. Mixed-date sources retain older evidence; exact hydration still enforces the same raw budgets. Twenty-two focused checks pass. No answer accuracy or API-relative latency gain is claimed, and the frozen full100 comparison is unchanged. See Research Log 159.

Earlier passage-routing diagnostic: additional addresses from exact passages within stored user summaries recover all three third-memory witness statements while retaining every previously hydrated user statement across all 30 development questions. All three memories have compiled passage indexes. The fresh answers above recover two of those questions but lose the painting recommendation, showing that recovered evidence alone does not ensure answer synthesis. See Research Log 142. The full100 passage report and support-syntax recovery now pass 54 focused checks. The fourth raw namespace is complete; its new full audit has six oversized summaries and no unresolved schema failures. All six compacted in one summary-only Qwen call; complete source admission now reproduces all 5,516 fragments without calls. Its 2,751 attention-guided leaves are complete, along with semantic vectors and 5,849 passage addresses. All four memories now share the replayed version-6 admission method. See Research Log 143.

Completed reader experiment (preparation history): 240 requests were prepared for all 40 questions in the four complete memories. The shorter policy adds precise entity/event qualification and direct preference-to-recommendation binding. All control prompts match the existing passage prompts byte for byte, with identical raw evidence for both readers. Eighty-two focused checks pass. The serial runner stopped before sending any answers when the fifth-memory ingest timed out. The prepared reader comparison remains unchanged. No new accuracy result is claimed. See Research Log 145. Both local model readiness probes subsequently returned HTTP 500. A separate recovery stage preserves all 99 completed fifth-memory responses and lists 733 first attempts plus five bounded additional attempts, without clearing original reservations or making recovery calls. Both original processes are terminal. See Research Log 146. Continuation: the 240-request comparison completed with the results above. Fresh readiness passed again, and the offset-40 recovery executor is now running after timed evaluation finished. A source-scoped summary-term supplement recovers the named schedule while preserving prior user evidence in all 40 diagnostic packets; its answer accuracy remains unmeasured. All four complete memories replay under one version-7 admission method with unchanged atom bytes; 120 focused recovery, admission, gate and scheduling checks pass. See Research Log 147.

The first two memories have complete admitted atoms, attention-guided leaves, semantic vectors, and user-summary addresses. The third is now fully admitted at 5,357 atoms and 1.045M token proxies. Its 2,693 attention-guided leaves, semantic vectors, and user-summary addresses are complete. Nine oversized source summaries were compacted in two Qwen batches plus one bounded recovery call. All three memories now share the same replayed version-4 admission method; 28 focused tests pass. See Research Log 140. The fourth raw namespace has finished all 838 requests and complete source admission. The fifth stopped with 99 completed responses and five unresolved requests; five more namespaces are prepared but unstarted. Gateway inference had recovered from the earlier outage before the renewed failure above. Original failed reservations and completed responses from that earlier outage remain preserved through explicit transport accounting. See Research Log 138. Full100 evaluation and untouched confirmation remain outstanding.

Completed real-data pilot: user-spine hierarchy and exact raw hydration. Forty Terra answers across five routing variants have now been judged by Sol. Every variant scored 7/7 on source-derived diagnostic questions and 0/1 on the separate benchmark chronology question. This is a small development comparison over 39 turns in three already selected conversations, not a full-corpus validation. User-only Qwen routing reduced input tokens by 35% and observed provider time by 24%, without improving accuracy. All answer and judge phases replay with 61 authenticated hits and zero new calls. The local integration suite passes 115 tests. Qwen receives summaries only. No router is promoted. The user's subsequent continuation instruction resolved the execution review block. See Research Log 131 for measured results, failure analysis, and reproducible artifacts.

Completed conventional packet comparisons. The fresh r9 reduced30 control scored 12/30. Source grouping scored 8/30, BGE ordering 10/30, MiniLM ordering 12/30, raw-only 12/30 and verbatim highlights 11/30. A compact packet reduced prompt tokens by 64% but scored 10/30 and did not improve observed answer latency. None was promoted; answer reliability remains unresolved. All seven arms are sealed and replayed, with 189 tests passing. See Research Log 128 for results, exact call accounting and the remaining failure mechanisms.

Conventional routing remains a control. Conventional routing with the fast packet has not been ruled out. On the same 48 synthetic summary queries presented in two orders, BGE-M3 selected correctly in 96/96 presentations; full Qwen returned 90 valid selections and six over-cap label lists. This is routing-only evidence, not fast-packet answer accuracy. Keep the Qwen routers experimental pending broader matched evidence on the new shared hierarchy. See Research Log 124.

Earlier experimental Qwen summary hierarchy. Attention-guided hierarchical chunking and summary routing are implemented as an opt-in path. Qwen sees only summaries and the query; exact raw sections are authenticated and hydrated afterward. The latest integration suite passes 282 tests. A real local Qwen smoke verified zero raw model inputs and exact hydration, but selected the wrong semantic branch. Corpus-scale compilation, timing and matched answer accuracy remain unmeasured. See Research Log 122 for the API, model-input boundary, local Qwen smoke and separate r9 lineage.

Authorized temporal-reference evaluation. The user confirmed authorization for the local gateways. The previously prepared successor completed 30 Terra answer calls and 30 Sol judge calls, scoring 12/30 versus r9's 11/30. Both artifacts replayed with zero calls. The changed-prompt group improved from 0/3 to 1/3; unchanged prompts had two gains and two losses. This does not evaluate the Qwen summary hierarchy or establish a causal improvement. See Research Log 123.

Fast-packet r9 handoff (separate lineage). The additive v7/r9 packet construction preserves all protected parent evidence and passes its focused 107-test contract suite, but the sealed answer result is only 11/30 on the exact v6/r3 failure cohort. The corresponding 81/100 number is a conditional non-regression projection, not a measured full100 result; no r9 full100 run was promoted. Gold-open inspection found sufficient supporting material in 18/19 remaining incorrect packets, localizing most residuals to source/event binding, temporal/operator semantics, preference synthesis, and final answer policy rather than raw-window absence. The implementation/docs are currently untracked and the sealed artifacts live under ignored eval_results/, so the operational continuation and durability warning are in Research Log 121, with mechanism details in Research Log 120.

The locked validation100 promotion gate now passes at 95/100; confirmation remains open. The proof-carrying policy-v5-r3 successor combines the frozen terminal P/R/L/G memory stack with an exhaustive numeric frontier and reducer-observable post-admission state equivalence. Its provider-free frontier closed Q28, Q53, Q67, and Q69 while correctly leaving Q14, Q40, and Q77 open. The final differential judge reused 97 authenticated Sol judgments, made exactly three new zero-retry calls, and sealed 95/100 with five remaining misses. This is still analysis-used validation evidence: the policy must be source-frozen before the disjoint confirmation200 treatment is opened, and confirmation must report both full200 and the predeclared non-exposed185 sensitivity slice. See Research Log 101 and Analysis 30. git log --oneline and the machine-readable artifacts remain the authority over prose.

Current 1M synthesis result. On the original ten-question development concatenation, the repaired v3 Terra policy held retrieval fixed and reached 6/10 exact match, 0.901019 mean F1, and 10/10 independent Sol semantic accuracy at S1, S2, and S3. The exact replay made zero provider calls and reproduced identical artifact bytes. This is diagnostic development evidence, not a target-gate result: the population is only ten and the structured synthesis call allowed 4,096 output tokens rather than the frozen answer-stage allowance of 256. The >=95% gate remains unpassed; see Research Log 25.

Earlier fixed-stage development status. The 8,000/256-token fixed-S1 path first ran on the original ten development questions. That lineage is protocol-ineligible because a sandbox-blocked root duplicated its first Terra reservation; its sealed Sol score remains a diagnostic 9/10. A later clean 100-question validation lineage completed at 56/100 and failed the target. See Research Log 27 for the development audit and Research Log 45 for the formal result.

Locked cumulative validation result. All ten retrieval shards are sealed and merged at e36b54ec...22007f. The exact 100-call fixed-S1 Terra responder completed and replayed at d7fc47b8...2a38cd. Independent Sol then scored the sealed predictions 56/100 semantically, versus 33/100 normalized exact match. The judge used exactly 100 live calls with zero retries; its zero-call replay reproduced 56/100 and byte-identical SHA 5dc56a24...ec77df. The preregistered >=95%/100Q gate therefore failed. See Research Log 45.

Routed numeric EM repair. A question-only numeric_reduce route held the sealed retrieval fixed and treated 32/100 questions at answer time. Thirty-two Terra compression calls yielded 19 valid fact packets and 13 baseline fallbacks; 19 answer calls changed 11 predictions. Independent Sol found three rescues and two regressions, moving 56/100 to 57/100. Compression, answer, and judge artifacts replay byte-identically with zero calls. This is analysis-used development evidence, not retrieval recall or a new target pass; see Research Log 47.

Isolated routed mechanism matrix. All six question-only mechanisms now have separate compression, answer, fallback, and changed-prediction judge budgets. Only numeric_reduce is positive (+1); direct extraction is -1, temporal timeline is -3, and set/state/synthesis are flat. The positive-only composer therefore admitted numeric for 32 routed questions and preserved baseline behavior for 68, sealing 57/100 with zero additional provider calls. All 18 compression/run/judge replays are byte-identical; see Research Log 48.

Matched retrieval matrix (closure v9 complete; broader matrix active). The common-renderer S0-v2 control scored 53/100. All 79 eligible representative-bridge/artifact-global retrieval artifacts and ten shards sealed; their 100-row matched answer and judge planes replayed byte-identically with zero calls. Representative bridge and artifact global each scored 52/100 (two rescues, three regressions, net -1) and are rejected from positive-only composition. Earlier 57/100 S0, 60/100 EM-fact, and 53/100 CAV observations used historical renderers and are not common-renderer marginals. The 84/100 number was an earlier 56 + 28 counterfactual ceiling, never an observed score. The >=95% gate remains unpassed; see Research Logs 49 and 54--59 and Analysis 13.

Matched evaluation spine (live control and gated repairs). Decision 2 is implemented under tools/matched_eval/. The fresh common-renderer S0-v2 control made exactly 100 Terra answer and 100 independent Sol judge calls, replayed both planes byte-identically without calls, and scored 53/100 semantic versus legacy S0's different-renderer 57/100. On the ten paired verdict flips, compact v3 scored 4/10, compact question-sandwich v4 scored 5/10, and gold-blind synthesis over v4 evidence plus both sealed answer hypotheses scored 3/10. The sealed candidates have a posthoc oracle union of 10/10 on that slice and 60/100 full, but no oracle-free resolver achieved it. All three promotion gates failed, so no full-100 repair campaign ran. See Analysis 13 and Research Logs 51–53.

Earlier CAV-ordering diagnostic. The original 1,039,203-token retrieval artifact can be answered downstream without rebuilding a corpus or opening the store. The fixed learned-CAV extraction/reinjection router ran over its globally deduplicated question/evidence features: on the same S1 evidence membership, the CAV-steered text-order arm reached 6/10 normalized exact match and 0.843171 mean F1, versus 5/10 and 0.811906 for matched original order. One encoder API call comprised 67 internal Qwen forward batches; 22 unique feature packets produced tensor-free order receipts. This is a ten-question X/X1 ordering proxy, not CAV linking or responder-side direct activation injection. See Research Logs 35--36 for that historical diagnostic.

Earlier genuine CAV-link dev10 result. CAV is the linking/fusion layer after cumulative S3 evidence, not another retrieval arm. A v2 artifact now seals the real rectangular extraction [K,N] and reinjection [N,K] links without an evidence-pair matrix or retained transformer token state. Matched S3 prompts held evidence membership and order fixed and changed only the link-guide slot. Terra exact scores were 7/10 unlinked versus 6/10 linked, but independent Sol judged both arms 10/10: the entire exact-score difference was 190 versus semantically equivalent 190 pages. This is no causal semantic-accuracy gain on dev10; the later raw fixed-S1 locked gate failed at 56/100, and the fair Mem0 gate remains open. See Research Logs 38 and 45.

Current causal Hebbian result. The old mechanism was implemented but absent from the benchmark lineage: all three Hebbian tables in the sealed combined source store had zero rows. A new causal replay learned 2,379 prior-access events into an isolated 5,978-node/51,072-edge graph, then compared sealed S0 against one budget-neutral tail-replacement arm. Only 3/10 memberships changed; base reached 6/10 EM and 0.836009 F1, while h1 fell to 5/10 and 0.736009. All three admitted edges had support 1 and co-access count 1; one replaced the decisive sixth-museum excerpt with unrelated recent evidence. Exact 13-call Terra execution and 13-hit/zero-call replay passed. This is a negative ten-question development diagnostic with no independent judge, not the >=100Q gate. See Research Log 37.

Current EM fact-memory result. The sealed S1 stage runs without rebuilding retrieval: S0 anchors episode selection, then S0-selected rows are excluded from the answer-time EM delta. The original facts arm tied raw at 6/10 exact match, improved F1 from 0.805372 to 0.827558, and cut prompt tokens by 39.45%. Independent Sol now judges both raw and facts 10/10; appending the complete raw tail scores 9/10 semantic because it contaminates one correct answer. Facts-only is therefore the measured operational default. See Research Logs 41 and 43.

EM v2 result. The facts-only candidate completed exactly 20 Terra calls with no retries: 171 selected EM rows became 17 facts citing 15 unique rows, mean prompt length was 3,190.6 tokens, exact match rose to 7/10, and mean F1 rose to 0.914065. Independent Sol judged all ten answers correct, and the zero-call replay reproduced 10/10. Protected S0 still supplied substantial evidence, so this proves representation preservation rather than an isolated EM recall gain. See Research Log 44.

Code/evidence boundary. The organized source tree is implementation epoch v4. Frozen validation-v3 evidence still certifies commit bfa5b6daf6a5e61881ac10f0555e5d9972f9e1c2 and implementation SHA 452be3bfa7524bb81676c7abcb032529a32a480311d24d1e17f8513c783ecd83. Because the implementation digest includes package-relative paths, v3 caches cannot be relabeled for v4; see 03 - Architecture/03 - Code Package Layout.md.

docs/
├── 00 - Theory/           Retrieval, proof-carrying coverage, associative memory, and episodic closure
├── 01 - Design/           The original architecture plan + eval design rationale
├── 02 - Implementation/   Setup, eval modes, Qwen labs, frozen contracts, and as-built mathematics
├── 03 - Architecture/     The as-built system map, package layout, and proposed hypergraph memory plane
├── 04 - Reference/        External landscape (SimpleMem, Mem0, MemDelta…) + vocabulary
├── 05 - Standards/        Normative data contracts (SQLite v2, embedding, memory provenance, formats)
├── 06 - Roadmaps/         Gap analysis: designed vs. built vs. measured, tiered next steps
├── 07 - Status Reports/   Dated snapshots (session handoffs)
├── 08 - Analysis/         Measured results, corrections, and evidence-backed engineering decisions
├── 09 - Archived/         Superseded material (append-only)
├── 10 - Research Log/     Dated experiment entries with data/ artifacts; baselines of record
└── 11 - Codex Workstream/ The Codex transcript decomposed: 9-chapter dev guide + 40 ADRs

• Resuming the user-spine hierarchy? → Research Log 129. The builder and summary-only Qwen smoke are implemented; 93 tests pass. The smoke exposes narrow-branch routing loss, and the real-source compiler is prepared pending explicit raw-payload approval.
• Reviewing the conventional fast-packet work? → Research Log 128. Seven fixed-pool packet controls are complete; none improved the fresh 12/30 baseline. Admission lost source/story witnesses, while temporal interpretation also failed with evidence present. The user has since resumed hierarchical attention.
• Resuming the requested summary-routing task? → Research Log 122. Qwen attention operates on hierarchical summaries; raw hydration happens only after selection.
• Resuming cold? → 06 - Roadmaps/01 - Delivering the Specified System.md first — it explains why every memory-arm number on record is void and what order the remaining work has to happen in. Then 07 - Status Reports/2026-08-15_retrieval-measurement-session.md for the retrieval half, which still stands.
• "What does the system do?" → 03 - Architecture/00 - System Overview.md.
• "How did factual retrieval improve, and what remains unproven?" → 00 - Theory/04 - From Top-K Recall to Proof-Carrying Factual Retrieval.md.
• "How will diffuse evidence use EM-LLM?" → 00 - Theory/05 - EM-LLM Episodic Discourse Closure for Diffuse Retrieval.md — the general closure mechanics and a bounded transient Qwen OV-transport episode signal are implemented provider-free. The latter is an attention-head semantic-change adaptation, not paper-exact token-NLL surprise, and still requires matched evaluation against fixed and embedding-change controls.
• "Where does the code live, and which imports are supported?" → 03 - Architecture/03 - Code Package Layout.md.
• "How would a native hypergraph interact with live memory?" → 03 - Architecture/01 - Native Hypergraph Memory Plane.md — canonical higher-order observations, pairwise serving projections, bounded traversal, and event-aware pruning.
• "How are complete sets deduplicated without losing distinct events?" → 03 - Architecture/02 - Query-Conditioned Bayesian Coverage Loop.md — a primary six-layer Qwen3-8B QK/OV affinity arm plus a secondary compact-INI classifier, followed by recall-safe representative-first packing; the locked baseline isolates a 100%-raw versus 94.7%-packed gap and the prefix treatment is pending measurement.
• "What's left to build?" → 06 - Roadmaps/00 - Gap Analysis and Roadmap.md.
• "How do we cut 1M-memory retrieval latency with minimal online compute?" → Analysis 31, Analysis 32, and Research Log 117 — source-balanced admission and binary packing provide the fast raw path; the online graph and typed temporal frontier raise strict source coverage to 99/100 with warm processing at 139.164 ms mean and 337.139 ms p95. This is evidence coverage, not judged answer accuracy.
• "How does user-led conversational ingest become episodic recall?" → Analysis 33 and Research Log 118 — user turns own stable envelopes, assistant/system payloads remain exact evidence, and macro structure links rather than replaces microepisodes. The opt-in production read API now expands selected raw hits under a separate companion budget; the eight-question pilot is positive and provider-free, while the full100 shadow gate remains open.
• "How do I run it?" → 02 - Implementation/01 - Running the Eval Harness.md (start with the free --compare mode).
• "How do I run the Qwen attention-prefix experiment?" → 02 - Implementation/03 - Qwen3 Prefix Attention Lab.md.
• "How will episode retrieval feed the K-latent attention fusion stage?" → 02 - Implementation/04 - Episode-Primary Latent Evidence Fusion.md — a design-frozen, query-conditioned GPU feature-to-router contract with no trained or measured fusion claim yet.
• "What exact math does the working implementation execute?" → 02 - Implementation/05 - As-Built Mathematical Reference.md — the code-aligned equations, defaults, edge cases, tie-breaks, and focused test map for working retrieval paths that were previously implicit.
• "How do I actually use it day to day?" → 02 - Implementation/02 - MCP Integration.md — the memory system is exposed to Claude Code as an MCP server.
• "Is this competitive?" → no evidence of that yet. The protected typed-final result is 73/100 on this already analysis-used locked population, still well below the preregistered 95% target. Reduced controls rule out batch memory pressure; the exact-ten treatment matrix shows that coverage-aware selection and CAV provenance activate but are non-monotonic when they compete inside one replacement lane. See Research Log 68 and Research Log 76. The same-budget Mem0 arm is also still open.
• "What is the large-model attention-head memory idea?" → 00 - Theory/01 - Extracted Attention Heads as Recursive Associative Memory.md — a DRAFT whose first CAV/live-head prototype is implemented, including the full-teacher J-Space implication.
• "How do later prompts consolidate connected memory partitions?" → 00 - Theory/03 - Prompt-Driven Systems Consolidation.md — schema-v9 causal binding plus repeated co-activation across semantic memories and evidence, including bounded iterative reads, the transient Qwen hyperplane seam, and anti-self-reinforcement rules.
• "Did the downloaded Qwen prefix produce usable CAVs?" → 10 - Research Log/02 - 2026-08-16 - Qwen3 prefix CAV gate.md — yes on the first controlled local probe; this is not yet a retrieval result.
• "Did extracted heads improve live-memory retrieval?" → 10 - Research Log/03 - 2026-08-16 - Live Qwen head memory smokes.md — calibrated layer-1 heads and temporal direction reached 1.000 R@1/R@3 on four development links; direct QK/OV failed and blind replication remains open.
• "Did persistent CAV/QK memory save tokens on unseen source families?" → 10 - Research Log/04 - 2026-08-16 - Safe associative memory confirmation.md — yes locally, without lowering baseline recall; a fresh recall gain and public-benchmark result remain unconfirmed.
• "Can attention heat control how much memory each source contributes?" → 10 - Research Log/05 - 2026-08-16 - Source heat diffusion development.md — implemented as a bounded external scalar walk plus source-aware packing; the selected dual QK/heat policy is posthoc development evidence awaiting a new locked split.
• "What exactly does 95% long-chat accuracy mean?" → 10 - Research Log/06 - 2026-08-16 - 95 percent long-chat target.md — answer-stage judge accuracy, minimum sample size, 8k hard prompt cap, locked LongMemEval partitions, and the experiment ladder.
• "Has live schema-v8 consolidation run through the real Qwen checkpoint?" → 10 - Research Log/08 - 2026-08-16 - Real Qwen consolidation path.md — yes on a temporary store copy; it validates execution and memory bounds, not a recall gain.
• "Did causal Qwen consolidation improve the operational long-chat probe?" → 10 - Research Log/09 - 2026-08-16 - Causal binding reaches 97.4 percent evidence recall.md — yes on the locked local literal-evidence test (38/39, no regressions); answer-stage judged LongMemEval remains the primary gate.
• "Did it answer LongMemEval questions from a 1M-token chat?" → 10 - Research Log/15 - 2026-08-18 - Policy-locked 1M-context answer pilot.md — its campaign artifacts report 10/10 positive judge verdicts with a mean 2,342-token legacy local prompt proxy; structural bindings are verified, but provider/judge execution and factual correctness are not independently authenticated.
• "What is the frozen v3 treatment and validation plan?" → 10 - Research Log/16 - 2026-08-18 - V3 retrieval freeze and validation campaign.md — the final no-provider replay, exact artifact/cache identities, ten-shard held-out plan, prompt-proxy semantics, and corrected Mem0 comparison boundary.
• "What is operational now, what is still unmeasured, and what comes next?" → 10 - Research Log/17 - 2026-08-18 - Locked treatment handoff and discourse closure frontier.md — the current readiness matrix, exact invariants and hashes, Mem0 production NO-GO, authorization gates, and the general-purpose diffuse-retrieval design.
• "Did the frozen v3 treatment generalize to all 100 validation questions?" → 10 - Research Log/18 - 2026-08-18 - Validation v3 provider-free retrieval audit.md — not at the retrieval-admission level; it records the exact metrics, identities, post-run cache audit, and why no provider accuracy claim follows.
• "How are new retrieval methods added without replacing the strongest prior packet, and what happened at 1M?" → 10 - Research Log/22 - 2026-08-21 - Recall-guarded cumulative retrieval.md — four provider-ready stages preserve a frozen-v3-compatible protected root, then add direct episodes, representative episodes, and artifact-global closure. On the exact original 1,039,203-token development concatenation, only direct episodes improved a scored retrieval metric; later stages preserved evidence but added no further gain under the cap.
• "What happened to the Hebbian arm?" → Research Log 37 — the mechanism existed but was never wired into the cumulative benchmark, whose sealed source had zero Hebbian events, nodes, and edges. The restored causal S0/H1 test works end to end but is negative: three single-support replacements changed 6/10 EM and 0.836009 F1 to 5/10 and 0.736009 because one discarded decisive evidence.
• "Was CAV supposed to be a linking layer, and did the genuine-link arm help?" → Research Log 38 — yes: CAV links/fuses canonical S3 evidence through rectangular concept extraction and reinjection before synthesis. The genuine matched arm stayed within the 8k/256 budget and both arms reached 10/10 under independent Sol, so this development test shows no causal semantic-accuracy gain; its one exact-score loss is only 190 pages versus 190.
• "Did a larger synthesizer answer the cumulative 1M contexts and score the episodic additions?" → 10 - Research Log/24 - 2026-08-21 - LiteLLM Terra episodic synthesis and rescoring.md — yes on the ten-question development concatenation: strict Terra synthesis reached 5/10 exact match and 0.718433 F1 at S1 with exact-quote citations, while both semantic and local numeric scoring found no strong S2-only evidence. This is not a held-out or independently judged result.
• "Did an independent judge verify the 1M answers, and did the synthesis repair work?" → yes. Research Log 25 records the 10Q development diagnostic; Research Log 45 records the clean fixed-S1 validation result. Sol accepted 56/100 sealed Terra answers, so the formal >=95% gate failed.
• "Did the fixed-stage 100Q memory test run?" → Research Log 45 — yes. The responder and judge each made exactly 100 live calls with zero retries and then replayed offline. The semantic result was 56/100, not a pass; the fair Mem0 arm remains outstanding.
• "Which retrieval styles are failing inside a method versus between methods?" → Research Log 46 — the sealed 100Q result is now split by evidence topology and answer operator. Sixteen errors have incomplete S1 source coverage, while 28 occur after nominal full-source acquisition; temporal ordering is retrieval-boundary-heavy and numeric aggregation is representation/reasoning-heavy.
• "How do we distinguish a failed method from one that was never applicable, and what did the first cleanup find?" → Analysis 26 and Research Log 97 — define the joined per-method outcome contract, remove repeated immutable work, record the 60.6% focused-slice speedup, and separate query latency from SQLite-heavy fixture construction.
• "Are provider prompts leaking local data or wasting slots, and how much tree memory was duplicated?" → Analysis 27 and Research Log 98 — no gold/raw locator/model-state leak was found, but stable internal IDs and repeated provider metadata were removed behind compact v2; sealed-ten chat tokens fall 29.32% and retained descendant tuple storage falls 95.2% while compact v1 stays exact.
• "Were retrieval stages excluding evidence only when combined, and is that repaired?" → Analysis 28 and Research Log 99 — reproduced and repaired fail-closed heuristic pruning, premature lane termination, incomplete specialist closure, false frontier closure, pre-admission dedup loss, and dedup-without-refill. The complete matched-eval suite passes 896 tests with one skip; no provider call or new accuracy claim was made, and witnesses outside the bounded first tree frontier remain the next explicit recall gap.
• "Did the new fixed-stage answer and judge paths work live before the 100Q run?" → 10 - Research Log/27 - 2026-08-22 - Fixed-stage S1 LiteLLM development diagnostic.md — operationally yes, but that ten-question lineage is protocol-ineligible because of one duplicate reservation. The later clean validation lineage is separate and completed at 56/100; see Research Log 45.
• "Has the locked 100Q retrieval campaign completed?" → yes. All ten shards merged at e36b54ec...22007f, the fixed-S1 answers sealed at d7fc47b8...2a38cd, and the independent judge completed at 56/100. Research Logs 43--45 contain the merge, answer, and judge records.
• "Why is the cumulative apparatus slow, and what changes after the frozen campaign?" → Research Log 34 — the completed offset-50 build made 963 snapshots for 480 sources and 12,891,681 source-row hash visits; across the then-sealed 60 questions S2 added evidence on 4/60 and S3 on 0/60. It separates code-backed cost candidates from the unproven cause of the q058 outlier and defines a six-step, receipt-safe optimization and benchmark sequence. Log 39 records the later exact frozen-source resume; no optimization is being relabeled as part of the historical campaign.
• "Can the intended H2 then CAV stack execute without another evaluation framework?" → Research Log 40 — yes in a provider-free integration path: exact H2 evidence now feeds genuine CAV links and the actual final 8k synthesis preflight through one transient digest-bound adapter. This is not yet a locked-100 answer-quality result; the locked shard adapter, final-stage-only routing, provider answers/judge, and official Mem0 arm remain open.
• "Can EM be converted to facts without duplicating selected evidence?" → Research Log 41 — yes: S0 first selects S1, then the protected S0 rows are excluded from the EM delta. On dev10, cited facts retained raw's 6/10 exact score, improved F1 to 0.827558, and used 39.45% fewer final-prompt tokens. Reattaching every raw row was worse, so raw inclusion remains an optional fallback rather than the default.
• "Did the consolidated test/eval path run live, and did the renderer repair work?" → Research Log 53 — the S0-v2 control ran and replayed at 53/100, but the follow-up gated repairs did not qualify. V3 scored 4/10, v4 5/10, and dual-answer synthesis 3/10 on the selected verdict flips. Their candidate union is only a posthoc oracle ceiling; no full-100 repair run was made.
• "Did independent representative/global closure improve the matched control?" → Research Log 59 — no. Both sealed descendants scored 52/100 against S0-v2's 53/100 and are excluded from positive-only composition. The 84/100 figure was a counterfactual ceiling, not an observed run, and 95/100 remains unpassed.
• "Can a provider-free full-store scan recover sources that S0 and both closure pools never constructed?" → Research Log 60 — partly. Question/S0 lexical routing plus complete scans of four semantic partitions reaches 19/27 missing sources and admits 14/27 under 2,048 tokens, with no question-ID prefix filtering or provider calls. This is source-history reach, not proof of answer-bearing excerpt relevance or answer accuracy.
• "Did balanced allocation fix partition-scan packing, and do routed EM facts improve the matched control?" → Research Log 61 — partly. Allocation v2 admits all 19 sources already reachable inside the top four partitions, up from 14, while eight sources remain outside the router and answer value is unmeasured. Separately, a parent-preserving facts-only gate scores 54/100 versus 53/100 with one rescue and no regressions; the two results are not additive and 95/100 remains open.
• "Does LLM query expansion add sources beyond the balanced scan, and where are they lost?" → Research Log 62 — yes, but mostly before packing. Query expansion reaches 20/27 missing candidates and admits 14; its admitted union with partition v2 reaches 180/188 total source targets (95.74%). This is structural source-ID coverage, not answer accuracy. The measured losses motivate source-balanced repacking and query-routed exhaustive partition scans before fact conversion.
• "Did the query-era answer arms improve matched accuracy, and can their union reach 95?" → Research Log 63 and Analysis 14 — direct payload improves S0-v2 from 53/100 to 71/100, while facts, partition, and guided payload score 64, 59, and 58. Their posthoc five-arm oracle is only 74/100, so recombination of existing predictions cannot reach 95; operator, answer-shape, sufficiency, and remaining construction failures require new work. Mem0 is still open.
• "Did reducing the workload show a memory-management problem?" → Research Logs 71--76 — no. The exact-ten v3 resident run held seven approximately 1M-token indexes (7,208,302 tokens total) and was then replayed one namespace per child process. Streaming cut simultaneous index residency by 85.66% and peak worker memory to about 0.822 GiB, but reproduced 7/7 namespaces, 10/10 questions, 70/70 method/question outputs, and the canonical construction bytes exactly. The four fact treatments show a technique problem instead: coverage and CAV provenance both activate, but replacement inside one shared lane is non-monotonic. The baseline fact reread remains strongest at a structural parent union of 14/23 targets and 3/10 complete source sets; no union is terminally repacked or provider-ready, and no new answer score is claimed. The next step is separate protected mechanism budgets followed by post-selection deduplication and one terminal fair repack.
• "What happens when every callback-selected fact fits and only the ten misses are answered?" → Research Log 77 — the sealed delta-only treatment retains all 75--109 unique callback rows per nonempty question, stays within 7,443/8,000 tokens, and scores 2/10 under ten Terra plus ten independent Sol calls. It proves later admission/composition losses on q31 and q81, but most failures remain upstream or relational: missing callback targets, source hits without decisive local spans, the q72 insufficiency case, and q86 distractor-sensitive temporal synthesis. The post-hoc known-miss result is not a locked-100 score.
• "Does the exact-ten failure persist when the answer workload is only a few thousand tokens?" → Research Log 78 — yes. Ten 2,772--5,832-token prompts lose no selected target sources during final fitting but score 1/10 after validation. Terra's raw outputs already contain the correct q31, q43, q61, and q86 answers; the global validator rejects all four and restores wrong parents. q7/q81 are upstream domain-admission gaps. This isolates technique and scope errors rather than process or prompt-memory pressure.
• "Can scoped specialist repairs recover those same ten misses?" → Research Log 79 — yes on the post-hoc development set. The canonical construction reaches 23/23 labelled sources with 4,213--5,849-token complete envelopes; ten Terra answers and ten independent Sol judgments replay byte-identically and score 10/10 semantic correct. This strongly rejects memory pressure for these misses and validates specialist-local proof scopes, but it is not a locked-100 score and cannot be added to 73/100. Full-population generalization and the same-budget Mem0 arm remain open.

Built ≠ measured, and locally measured ≠ externally competitive. Passing tests establish implementation behavior. The local analyses and research logs establish only their stated datasets, splits, and metrics. The QK/CAV result currently supports token saving with recall non-regression on one locked fresh split; it does not support a general recall-gain claim. Any broader claim without a public benchmark is a bug — report it.

Dated status reports, analyses, research logs, and frozen scripts may cite the flat module paths that existed when their artifacts were produced. Active code examples use the canonical v4 package paths; historical evidence is not mechanically rewritten.

Verification block: run

git log --oneline -1
git status --short
pixi run --frozen -e dev pytest -q

If the suite is green, this tree is accurate as far as it goes: the core and association and source-heat paths are implemented and locally measured, while external competitiveness remains unknown. The next evidence gates are improved write-time association coverage and --answer-recall or another public/common benchmark—not a larger transformer context or a third retrieval hop.