Work In Progress (WIP) Documentation

This folder holds early-stage notes, backlog plans, and reference material that has not been promoted to a PRD, RFC, ADR, runbook, or release note. Content here is not authoritative — when an item is ready, it moves to its proper home and the WIP file is removed.

Promotion targets

• Long-term feature ideas → PRD in docs/prd/ and the relevant RFC in docs/rfc/
• Architectural decisions → ADR in docs/adr/
• Operator-facing procedures → guide in docs/guides/
• Post-release retrospective → fold the substance into the corresponding release note in docs/releases/ (do NOT leave it in WIP and link to it — see the rule below)

References are ONE-WAY: permanent docs must never point back into docs/wip/

A WIP doc is deleted the moment its work lands, so anything permanent that links to it becomes a dead reference. No ADR / RFC / PRD / release note / guide / docs/api/ doc, and no code / test / README / comment / docstring, may reference a docs/wip/… path. (WIP↔WIP is fine — a WIP set travels together; and a WIP doc pointing at a permanent artifact is just naming its promotion target.) When a permanent doc needs WIP content, promote it — copy the substance inline or cite the durable source (shipped code, commit hash, issue/PR, or the ADR/RFC/PRD that superseded it). This is the rule the 2026-08-02 hygiene pass had to retrofit; the templates and AGENTS.md carry it.

Start here — the live handovers

Two threads are in flight and each keeps its own handover. Read the one for your thread; they do not cover each other.

• App / UI — BRANCH-HANDOVER-ui-followups-2026-09-27.md is the entry point for PR #2127 (110 commits, 2026-09-18 → 09-27). It links down to DEVICE-TIERS-HANDOVER-2026-09-27.md for the device-tier detail, and says which issues close on merge, which four are deliberately open, and what is NOT done.
• Backend / processing — that thread works from a separate worktree and lands its own dated handover here. A second HANDOVER-2026-09-* file is not a duplicate of the above — it is the other thread's, and equally live.

Because this directory is shared by every branch, do not retire a WIP doc that belongs to the other thread's domain — pipeline, corpus, enrichment, DGX, prod infra — without that thread's say. A doc that looks stale from one side is often the other side's working plan.

Every other HANDOVER-* / *-handover.md / *-HANDOFF-* file below belongs to an earlier, unrelated arc — 17 of them as of 2026-09-27. They are not a reading list. Several describe work that has since shipped and none has been re-checked; retiring them is the pending hygiene pass (DOCS-HYGIENE-AUDIT-PLAN.md, untracked). Open one only when its topic is the topic at hand.

Current contents

File	Description	Status
BRANCH-HANDOVER-ui-followups-2026-09-27.md	TWO SESSIONS — read to the end. Session 2 (october-fixed, 2026-09-27 evening, 21 commits, UNPUSHED) is appended below the original: nine device-only defects the operator found on an iPhone against 1,714 passing tests, with an equal-weight NOT-done half — the scroll artifact still undiagnosed and its speculative lever deliberately unshipped, Ask's timestamps/duplicates index-side, #2164 recorded not repaired. Ends with the session's failure pattern, four instances of which were mine. Branch-level view of PR #2127 — 110 commits, 2026-09-18 → 09-27, six arcs. Start here; the device-tier handover below covers only the last three days. The arcs: review-round closure (revisit outcomes removed the card before the write could fail); two device-feedback rounds (Collection and Share were dead on every sheet; storyline rows dead on tap); the Storyline/Theme rename, #1603, executed as A1–A4 + B1–B2 — the largest arc and the reason CodeQL went red, because renaming a file re-issues its dismissed alerts; played-state/queue rework plus two auth fixes (a token without an identity could block the login page); the Android tier stood up from nothing, #2139 — its only prior test was Capacitor scaffold asserting a wrong package name that nothing ran; and both tiers reaching green. Records the 6 issues closing on merge, the 4 held open with reasons, and what is NOT done — ~32 controls left on #2156, FCM unwired on #2157, iOS phase 1 not actually queueing, nothing resetting server-side accounts, and the e2e floor deliberately left at 37.0.	Active (handover)
DEVICE-TIERS-HANDOVER-2026-09-27.md	Both device tiers green on fix/ui-followups-2026-09-18 rebased onto main — TEST_IOS_EXIT=0 (25 suites) and TEST_ANDROID_EXIT=0 (14 suites), with ci-fast and ci-ui-fast at 0. Records the four PRODUCT bugs the tiers found that CI cannot (push enable killed the app on Android; the favourite heart had no accessible name; Profile never refreshed so stats froze until restart; five more unnamed icon-only controls), and the seven test defects that were passing for the wrong reason. The pattern worth carrying forward: five failures shared one shape — something persisted that nothing reset (stale lance_index in the fixture, server-side accounts, simulator app state, the server-side queue, account-dependent Home length). Both tiers now reset the DEVICE; nothing resets the server-side ACCOUNT, which is the biggest remaining source of false greens. Known-broken and deliberately not fixed: iOS phase 1's queueing step makes zero queue writes despite its name; the blank-WebView sign-in recovery has never fired so it is untested; the a11y audit walks 5 surfaces and iOS has none at all. Also carries the traps (scrollTo's stall detector vs network-loaded content, Android's on-screen-only tree, a renamed file re-issuing CodeQL dismissals, git hooks not inheriting unexport NODE_OPTIONS) and the CI question — neither tier runs in CI, so a green PR says nothing about them. Supersedes the 2026-09-25 and 2026-09-26 Android handovers (deleted).	Active (handover)
HANDOVER-2026-09-30-outage-fixed-corpus-mop-up.md	The 2026-09-30 reboot outage (public APIs down ~2 h: RAM secrets gone, keyless control plane, alloy not started) and its fixes on main — RemoveIPC=no (live + cloud-init), the recovery workflow's three bugs, a read-only prod recovery check for the homelab remediation fleet (agentic-ai-homelab RFC-0005). Corpus after the overnight queue: 4 bad KG graphs (2 emptied by relabel runs, now guarded by ad0d25f18; 2 blocked by the truncated-title transcript lookup, fixed in 4c9618820), 15 relabel violations left. Undeployed commits, the remaining order (deploy → rederive 4 → relabel 15 → full search rebuild → nightly), pending operator decisions, parked items, ops notes, NOT-verified list. Supersedes 2026-09-29 (renamed).	Active (handover + runbook)
DEVICE-TIER-PARKED-AND-MEASURED-2026-09-28.md	The PARKED half of the a11y-audit and test-independence work, so it is not lost. #2156's "37 unnamed controls, ~32 remaining" was an ESTIMATE, not a list: the measured aria-haspopup failure class is closed 6/6, and controls the old theory called broken are named on both engines (proven by app.buttons["Play"] driving four iOS suites, and "Skip forward 30 seconds" / "Change photo" tapped by name on both tiers). 19 speculative fixes were started against the old theory and fully reverted — the wider rule would have added absolutely-positioned sr-only spans to ~28 already-named controls, which is how testN4's "present but not tappable" branch gets triggered. Widening the audit 5 → 10 surfaces found 2 findings, both Settings toggles, and A11yProbe (new, reads AccessibilityNodeInfo directly) proves they are real: all five platform name fields empty, so not the UiObject2 blind spot that made 10 of the original 38 false positives. PARKED because two markup fixes both failed — aria-label, then explicit for/id + aria-label, each with the rebuilt bundle confirmed served; do not re-apply either. Also carries: the unresolved SavedColorControl-vs-OverflowMenu contradiction (settle it on the device, not with a regex), Profile ▸ Topics re-rendering 15 of 29 nodes mid-walk, lp-e2e-api measured up 19 hours still holding old accounts, sign-in flaking 2 of 9 runs before the callback path, and five advisor findings CONFIRMED by reading but not fixed — iOS openProfile's hardcoded identity list silently breaking startClean's offline normalisation, and the contact sheet photographing an account nobody seeds.	Active (handover)
TOPICS-AS-SENTENCES-2026-09-27.md	Tracked as #2164. 96 episodes (4%) render ZERO topic chips because GI wrote its insight sentences into the Topic nodes. Found from one screenshot showing "Topics & People · 1" on an episode with nine insights; the KG holds 10 Topic nodes and every one is a 27–35 word sentence that appears verbatim as an insight. is_filler_topic rejecting them is CORRECT — the defect is upstream, and the app should not change. Measured on prod across 2,283 episodes / 22,458 topic nodes: 97 have any sentence-shaped topic, 96 have every one (it is binary, the signature of a per-extraction failure), 830 nodes affected. Clusters in TIME, not by feed — 82 of 96 in 2026-08-11 → 08-19, peaking at 53 on 08-14, across 15 feeds and 24 runs, so it is a pipeline/model-state change in that window. NOT established: which change inside the window; whether the 9-episode Aug-27 → Sep-02 tail is the same cause; whether a re-derive fixes it (re-run ONE episode first — that is also the cheapest test of whether the defect is still shipping); and the blast radius across every other Topic consumer (discover signals, followable interests, feed_signals.top_topics, co-occurrence), all of which apply the same filter and are quietly missing these episodes.	Active (findings)
DEVICE-ONLY-DEFECTS-2026-09-27.md	Three defects the operator found on an iPhone against a build that passed 1,702 unit tests, vue-tsc and mkdocs --strict, none reproducible in jsdom or desktop Chromium. (1) Collection names render blank in the add-to-collection menu — UNSOLVED after two shipped fixes; carries the elimination table so the dead ends are not re-derived (prod data, wire contract, DOM across 1st/2nd/3rd open, class generation, :root variable scope, zero-width shrink, the real compiled CSS painted in Chromium, bundle freshness grepped out of the installed App.app, and a contrast-stretch of the screenshot showing NO glyphs at any intensity). Names the single observation that would split the remaining space — whether a long-press selects invisible text or nothing — and says not to ship a third speculative fix without it. (2) "Ask this episode" is a search wearing an Ask label (renders raw hit.text), plus a duplicated hit and missing timestamps that kill jump-to-moment; diagnosed, undecided, nothing changed. (3) Fixed chrome paints mid-list during momentum scroll; the usual translucency cause is ruled out (both bars opaque, backdrop-blur removed for WCAG), so undiagnosed. Closes with what the four device-only failures of the night — including the top-layer regression three of my own tests passed over — say about neither tier running in CI, and the cheapest smoke that would have caught two of them.	Active (handover)
REPAIR-10-CORRUPTED-EPISODES.md	Repair record for the 10 episodes the 2026-09-23 relabel_only batch corrupted by globbing {idx} - *.txt across the feed root (an on-disk idx is unique only inside one run_* dir, so 16 episodes were handed a different episode's transcript and 10 reached serving with gi.json/kg.json rebuilt from the wrong words). Cause fixed in 21bce7aa3+e2dedbd17; all 10 re-ingested and verified BY EPISODE ID on 2026-09-24 (FULLY VERIFIED: 10 of 10, corpus-wide mismatch count 0). Revision 1's delete-first route is recorded as REJECTED with the three reasons it was wrong — chiefly that metadata is both the skip marker and the selection record, so deleting it makes the run exit 0 having done nothing, and its own verification would have passed in that failure state. Open tail: LanceDB still holds the stale chunks (needs reindex-prod.yml mode=rebuild) and the 10 need re-enrichment after.	Done (cause + data); index rebuild pending
STORYLINE-INDEX-PROD-MIGRATION-2026-09-17.md	Post-deploy procedure for the prod operator now that storylines are real index rows (doc_type="storyline", one row per theme cluster). ONE additive cli index --output-dir /app/output via run --rm --no-deps api — not a rebuild, no schema migration, no api restart, app never started. Records why index-two-tier / --rebuild must NOT be used (drop_existing=True + divergent chunk params: 32/36 re-embedded, segments 131→115), the root-ownership trap that silently turns search dark (index_unreadable → no_index), the topic_theme_clusters.json precondition, verification curls, and an equal-weight NOT-verified section (nothing run against prod; prod runtime and storyline count unmeasured; ranking untuned).	Active (runbook)
ONBOARDING-RUNBOOK-200-700-EPISODES-2026-08.md	Operator runbook for onboarding 200–700 episodes across new shows (feat/onboarding-readiness arc). Preconditions in order (merge → deploy → backfill → snapshot → health drill), measured budget math ($0.238/ep → $60–210 with retry allowance), cost-chunked batch sizing (~$10–25/dispatch, batch one small), per-batch 7-point verification checklist, abort paths (#1785 STOP endpoint → stop-prod-pipeline.yml → SSH), and an explicit NOT-covered section (ML enricher re-enable #1817, retroactive fingerprints, feed selection, consumer-side capacity). Every prod dispatch operator-gated.	Active (runbook draft)
1192-1286-1285-guest-recall-arc-findings.md	Measured findings for the guest-recall tail. #1192: the transcript-intro guest lever names 0 new guest voices and causes 3 ASR name-flips on the prod-v2 90-ep relabel — the roster's active on-air-intro + self-intro paths already exhaust text-reachable recall → not shipped (validated negative, like the #1228 revert). #1286 (embeddings) is the only lever that reaches the ~113 un-introduced-panel tail; needs DGX. #1285 (ASR name canonicalization) confirmed real by the 3 flips (Cerisier/Serissier, Karnal/Carnell, Allardice/Allardyce).	Active (findings)
2026-07-26-production-security-ops-review.md	Fable 5 advisor holistic production security + ops review across the three surfaces (player / operator / tailnet-only privileged plane). P0 origin-lock (subsequently verified applied), P1s (operator viewer rate-limit gap, ?grant=creator authZ footgun, no operator-appdata/corpus backup schedule, public-repo bcrypt doorman), P2s, and an equal-weight not-covered/unknowns section. Analysis only.	Active (review)
2026-07-26-security-ops-mitigation-plan.md	Phased, owner-tagged mitigation plan for the review. Phase 0 verification done (origin-lock CONFIRMED applied via TF state — :443 locked to CF ranges; OPERATOR_SECRETS_VIA_FILES off; .com in admin emails), Phase 1 in-repo P1 quick wins, Phase 2 hardening (mix me/operator), Phase 3 maturity. Drives the security/prod-hardening-2026-07 PR.	Active (plan)
observability-app-surface-plan.md	Phased plan for full self-hosted o11y (metrics + logs) over the podcast operator surface (pipeline, LLM, ML, enrichers, player, search) plus orrery, grounded in a telemetry-emission recon. Phase 0 (infra collector, api /metrics, node + edge-security dashboards, Cloud-agent removal) done; Phase 1 = ship scoped structured app logs → VictoriaLogs + log-derived dashboards (LLM cost, pipeline, search); Phase 2 = app instrumentation for real-time ML/search/enricher metrics; Phase 3 = orrery; Phase 4 = alerts + retention.	Active (plan)
1000-EPISODES-REPROCESS-PLAN.md	THE canonical arc doc (reconciled 2026-08-02 — absorbed the former v2.5 handover). Reprocess v2→v3 with a fully-local (DGX) pipeline + expand to 500–1000 eps / 20–30 podcasts (10k horizon). Incremental single-variable arc v2.1→v2.5; v2.2/v2.3/v2.4 MERGED (#1335, #1355), v2.5 = current front = Gemini→DGX-local LLM swap (stages D bake-off → E swap → F freeze; gate = disjoint-vendor scalar judge-panel parity vs Gemini). Organizing principle = reprocess-once economics. v4 fixtures (#1189) as a growing harness. Open: go-live-vs-2.5 sequencing, parity-gate signal, expansion feeds (#630).	Active (plan)
DIARIZATION-SPLIT-HOST-CLUSTER-MERGE.md	Follow-up behind the #1330 naming fabrications. The v2.1.x re-cascade surfaced two "fabricated person" classes that are DIARIZATION artifacts the naming layer now CONTAINS (verified by deterministic replay): a cold-open montage cluster ("I'm Kevin Russo… I'm Casey Noon…" merged into one 13s SPEAKER_NN → refused a self-intro name when short, via distinct_self_introductions + MONTAGE_CLIP_MAX_TALK_S), and a detected-guest forced onto a bumper ("Robert Pape" onto a 30s "We'll be right back" → surname-aware spare, honorific-restricted). Both fail toward unnamed. The real cause (merge/split of clusters) is a diarization-layer normalisation deferred to v2.2 (community-1); blast radius rewrites voice ids so it's its own branch.	Active (follow-up)
SPEAKER-RESOLUTION-ROADMAP.md	Roadmap to reduce unknown diarized speakers, with before/after measurement per step on the real prod-v2 corpus (90 diarized eps / 579 voices). Shipped: #1a (+41 voices), #1b (episode-scope), #2 (publisher denylist), #3 (host/guest role), voice_type classification, Step B known_hosts (+19 voices, cumulative 36.6%→47.0%), Step C (unnamed host→"Host" label), Step D (intro NER guest detection, +7 voices). 466/579 (80%) handled; 113 truly unknown remain (un-introduced panel guests). Open decisions: promote measure scripts to make target, chase talk-time threshold.	Active (plan + measured)
GOAL1-ORRERY-CRITICAL-PATH.md	Trimmed scope for the operator's real goal order (1 orrery on minimal common infra, 2 player, 3 operator surface). Pins what Goal-1 needs (live rollout + T-11 + orrery vhost coordination) vs what is PARKED (all player work) vs deferred (goal-3 privilege-split).	Active (scope)
GOAL1-GO-LIVE-PLAN.md	Owner-tagged (🧑 You / 🤖 Me / 🤝 Both) phased plan to take orrery live on the shared edge. Config is built but not applied to the running box (cloud-init ran pre-config). Recommends Option A (imperative-once, no rebuild → no data-loss); phases: safety net → converge box (firewall closed) → o11y live → pre-public gate → open firewall → orrery onboard → Cloudflare → post-live. Open decisions: apply path A/B/C, corpus location (volume vs boot disk), re-rebase timing.	Active (plan)
GOAL1-PHASE3-PREP.md	Phase 3 hard-gate prep: (a) pre-public-gate pre-walk for orrery — 8 items mapped to done (substrate: catch-all Caddy, metadata-egress, T-11 alerting) vs pending (orrery cap_drop / digest-pin / rollback rehearsal); (b) ADR-115 secrets cutover exact cutover-gate-ordered steps (age key → prod.enc.yaml → decrypt on box → add secrets overlay → drop plaintext keys).	Active (prep)
GOAL1-AUDIO-ARCHIVE-ROLLOUT.md	Audio-archive Storage Box (#1199) rollout/transition plan: current state (storage_box.tf gated off, audio_storage_backend=local/remote,archive pull CLI, reprocess-prod), data-at-risk analysis (corpus backed up daily; audio archive NOT covered + best-effort-irreplaceable), 7-step rollout after DR-green (provision → verify → backfill+verify → flip backend → e2e → backup policy → prune), new-command runbook. Folds in as go-live Phase 7.	Active (rollout)
GOAL1-MERGE-TO-LIVE-ROADMAP.md	Living from-here roadmap: the gated sequence from "infra PR merges to main" → orrery live (harden → push → PR → merge → DR-drill-green → go-live phases). Holds the invariant (merge applies nothing live) + the open-decisions tracker (apply path, corpus location, DR-drill gate, branch hygiene, merge style). Links GOAL1-GO-LIVE-PLAN.md for phase detail.	Active (roadmap)
INFRA-HARDENING-PLAN.md	Sequenced execution plan for the infra security hardening effort (T-01…T-12 from docs/security/THREAT_MODEL.md). Infra-first: Docker/orchestration → Host → Edge → App. Precedes the shared public edge (ADR-114). Tracked in one hardening issue.	Active (plan)
1161-API-SEPARATION-ROUTE-INVENTORY.md	#1161 reassessment v2 (corrected): three surfaces — consumer player (/api/app/*, OAuth, settled), kg/gi web app (gi-kg-viewer = today the operator console driving the FULL privileged /api/*), operator control plane. Two gaps: (1) /api/* has NO per-request authz (role system listener<creator<admin wired only into /api/app/*); (2) docker.sock behind a public RBAC gate reintroduces T-01. Recommends splitting the backend by privilege (public-api no-sock vs control-api tailnet-sock, enqueue→drain). Key fork: does admin trigger privileged actions from the public web app?	Reassessment v2
EVAL_1016_metrics/vllm_metrics_*_phase2c.log	Raw vLLM /metrics polls per candidate (input data for the canonical per-model param compendium, which lives at autoresearch/PER_MODEL_OPTIMAL_PARAMS.md)	Reference
PUBLIC-EXPOSURE-AND-PRIVATE-SPLIT.md	What the public repo actually exposes, and the private-split options. Separates two boundaries that get conflated: prod content (transcripts, segments, GI artifacts — never public, boundary holds; the 30 "committed episodes" in prod_validation_v1/ are symlinks, 4.8 KB of paths) versus method (456 eval configs, the 2,175-line autoresearch playbook, 49 scorecards, 148 WIP notes — fully public, operator accepted 2026-07-14, will revisit). Records the enforcement gap: the content boundary rests on .gitignore alone, with no pre-commit check — a measured shape+allowlist guard (0 false positives across 203 synthetic-fixture hits) is proposed but NOT built. Notes that history is permanent, so forward-only and retroactive purge are different decisions with wildly different costs.	Active (analysis, no action taken)
1046-WHISPER-DUAL-MODEL-FUTURE-USES.md	Parks 5 alternative uses of the dual-model machinery (dual-pass reconciliation, confidence-weighted NER, sniff-driven NER pre-pass, speculative pipeline, cross-model dispatch) — all align with the intelligence-extraction goal that the skip-deep gate violated. None queued; planning material for next session pickup. Includes the offline-prototype-able subset using saved transcripts under data/eval/runs/1046-measurement-pass-2/.	Backlog
DIGEST-TOPICBAND-THREAD-UNSAFETY-ARM64.md	Follow-up to the parallelisation ranking above. Concrete SIGSEGV (api container exit 139) inside ThreadPoolExecutor.map around run_corpus_search on macOS arm64 stack-test, faulthandler pointing at sentencepiece / torch C extensions. Documents the workaround (unconditional sequential map() in corpus_digest.py — same code path in prod and stack-test) and four options for a real fix (warm the tokenizer on api startup — recommended first; module lock; process pool; upstream wheel patch).	Followup
super-theme-signal-comparison.md	Queued: compare 3 signals for the super-theme rollup on top of topic_theme_clusters — cross-cluster topic-lift (shipped default in v1.1.0), centroid cosine, member Jaccard on 1-hop lift neighbourhood. Pick winner by editorial-read of super-theme labels on prod-v2. graph-v3 tier 7-1a follow-up.	Queued
DGX_NEXT_STEPS.md	Living strategic doc on what runs on DGX vs local, vLLM vs Ollama, offload decision frame	Living
LORA_HYBRID_PIPELINE_PLAN.md	LoRA + hybrid pipeline exploration	Idea
manual-test-plan-gi-kg.md	Manual GI/KG smoke checklist	Reference
wip-concurrent-pipeline-http-retry-metrics.md	Open documentation gap for http_urllib3_retry_events	Open
wip-topic-clusters-validation-reference.yaml	Reference topic-cluster validation config	Reference
player/mockups/	Phone mockups (HTML + PNG) of the three explored Player aesthetics; Direction B (Editorial Bold) adopted → UXS-011. Design aids, not shipped assets.	Reference
NER_FP_SAMPLE_LABELLED_2026-06-24.json	Operator-labellable 50-row sample produced by scripts/dev/measure_ner_mentions_diff.py against the prod-v2 corpus; backs the TP/FP claim in the determinism investigation doc	Reference
enrichment-visual-inspection-plan.md	4-stage plan for restarting the viewer against a freshly-enriched small corpus and inspecting where each RFC-088 enrichment signal surfaces in the player/viewer UI (uncovering UX gaps).	Active
ONBOARDING-SHOWS-FOR-ENRICHER-VALUE.md	THE canonical corpus-expansion doc (current state refreshed 2026-08-29). How to grow the eval corpus with more real shows so the enrichers produce visible value. Value model per enricher; highest lever = topic OVERLAP across shows. Read order: §5f (final 15+10 feed list, all RSS verified) → §5g (smoke→assess→deepen protocol + buckets) → §5i (evidence-based thresholds; bridge_partition.both is the primary signal) → §5j (live corpus = 14 feeds / 765 eps; probe group 1 all DEEPEN; 10 Batch A feeds still to go). §1 and §6 are STALE and marked as such — §6 is the pre-verification sketch and sits after §5f purely by append order. Expansion vehicle = #630.	Active
2026-09-02-enricher-reenable-prep.md	Prepared runbook (#1921) for turning the three DETERMINISTIC corpus enrichers back on (grounding_rate, topic_cooccurrence_corpus, topic_theme_clusters) after the Batch A ingestion lands — the two ML ones (topic_similarity, topic_consensus) stay off by operator decision. Carries the full enabled/disabled inventory, the PUT /api/enrichment/config whole-block-replace hazard, a schema-validated payload, and a testable hypothesis for WHY they were disabled (30s manifest cap vs the 300s override the two enabled corpus-scope enrichers carry). Also lists what to measure, including whether populated co-occurrence reframes #1918.	Active
HANDOVER-2026-08-29-batch-a-remainder.md	Executable runbook for the next batch: ingest 10 episodes each for the 10 remaining §5f Batch A feeds. Unit of work = 100 episodes (10 eps x 10 feeds), a REPEATABLE pass — not "onboard ten feeds". Run three times: 08-30 08:33, 08-31 09:05, 09-01 22:41 (in flight). episode_selection=unprocessed means each pass takes 10 NEW episodes, so the ten feeds sit at 20 eps and go to 30 — corpus 24 feeds / 966 eps, fully local at $0. Steps 1–2 are marked superseded: the PUT /api/feeds merge was never a prerequisite, and the selection control is episode_selection (per-request, 998d5312), not episode_order — see docs/guides/INGESTION_RUNBOOK.md, now canonical. Steps 3–4 are the live work: no Batch A feed has a §5i grade or a §5g bucket yet. Open + unexplained: KG node_count pinned at exactly 29 across 7 episodes; insight_salvage discarding insights 26–30 by arrival order.	Active
HANDOVER-2026-09-14-enrichment-o11y.md	Enrichment o11y (#2071) + the unwatched pipeline. Four podcast_scraper commits pushed, none deployed (prod runs sha-080403f). Enrichment emitted to run.jsonl only — 0 lines in VictoriaLogs, 0 series in VictoriaMetrics; fixed at the _safe_append_event chokepoint plus a metrics exporter and per-enricher spans. org_web had never worked (no profile set; read GI for a KG-only node type; P31 allowlist rejected NIH/Newsweek/Nature and resolved Stanford the town) — 0/5 to 4/5. Wider finding: no dashboard in Grafana referenced enrichment, pipeline, ingest, asr or corpus at all, while ~2600 pipeline_stage events/week sat unread; panels + dead-man alerts added in homelab ffb0cae. Blocked on one operator command (§4, mini repo is 0700). MUST revert prod experiment caps max_persons: 3 / max_orgs: 3 (§6).	Active
EPIC-HOST-IDENTIFICATION.md	Epic spec: reliably know every show's host(s), persist as metadata + a person→HOSTS→podcast graph edge, and render host-aware behaviours (graph/digest/library/person card). TDD spine = a host scorecard (Slice 0: gold set + coverage/precision/recall/sample-bias metrics) that every path must move toward target. Paths A(per-show edge)→C(show-notes parsing)→B(voice embeddings)→D(multi-guest)→E(role taxonomy). Supersedes the "precise per-show host/guest" follow-up.	Specced
PLAYER-GOLDEN-WALKTHROUGH-v3.md	v3 — supersedes v2. Same exhaustive player walk, folding in the new consumer "Where they agree" consensus row (topic_consensus is no longer operator-only): §2.3 person card + §3/§4/§5/appendix updated, with a fresh person-card capture (Nic Harrigan, live consensus row) in assets/player-walkthrough-v3/ (8 unchanged surfaces reused). Cross-refs the operator show-landing Signals band (RFC-104/UXS-015 Phase 2).	Active
ENRICHER-HARDENING-ROADMAP.md	Fixtures→coverage→surfaces roadmap grading every enricher surface/cross-link against a weak/good/excellent rubric. PR-A (v3 topic_similarity fixture + invariants) + PR-B (per-enricher emission non-degeneracy + consensus executor smoke) landed on fix/adr108-enricher-surfaces; PR-C = #1168 (operator-viewer served-corpus e2e harness); PR-D reconciled 2026-07-09 — the "thin" surfaces were already built + tested (data-absence, not code-absence), so PR-D is capture+verify, not build.	Active
graph-v3/HARDEN-FOLLOWUPS-2026-07-17.md	2026-07-17 harden pass on feat/graph-v3 surfaced 3 items self-noted as "untracked" inside other WIP docs (USERPREFS-1 three deferred items, aggregatedEdges V1 enricher-gate, Speaker + Quote shape live-verify). Consolidates them so they're findable by title; no GH issues opened per operator's standing rule. HD1–HD9 harden fixes landed on the branch and are documented in graph-v3/SUMMARY.md § "Post-tier-8 harden follow-ups".	Active
graph-tech-debt.md	Running log of graph-viewer improvements surfaced during PR work but deferred as out-of-scope. Template + convention for adding new items; two entries from PR #1207 (residual +408 ms wave-1 fcose cost, quality: 'draft' as documented off-lever). Launching pad for future perf sessions.	Active

Guidelines

• Documents here are not part of the official documentation site.
• Documents may be incomplete, outdated, or experimental.
• Periodically review and either:
• Promote to the appropriate doc category (PRD / RFC / ADR / guide / release note).
• Delete if obsolete or superseded.
• Keep as backlog / reference if it still has signal.

Recent cleanups

• 2026-06-24 second pass — Removed 7 more notes for shipped/superseded work (47 → 40 files):
• Autoresearch programme planning docs (2): AUTORESEARCH_NEXT_PHASE_AGENT_PLAN.md, AUTORESEARCH_NEXT_PHASE_DEPENDENCIES.md — #907 + #927 epics + children all closed per [[project_autoresearch_programme]]; plans no longer live-bearing.
• Plans for shipped features (3): RESEARCH_POWERED_REGISTRY_PLAN.md (registry promotion shipped in #1060), VLLM_RELOCATION_TO_HOMELAB_REPO.md (homelab repo already owns /opt/vllm-autoresearch/docker-compose.yml), EVAL_1016_LANDSCAPE_2026_06.md (#1016 children all closed; raw metrics dir retained as historical input under EVAL_1016_metrics/).
• Superseded measurement notes (2): VLLM_GB10_TUNING_VALIDATION_2026-06-18.md (canonical compendium is autoresearch/PER_MODEL_OPTIMAL_PARAMS.md), FIXTURE_AUDIO_TOOLING_COMPARISON_2026_06_13.md (tooling choice landed in shipped fixture pipeline).
• Ghost-row cleanup: 2 entries (967-interaction-cost-trace.md, POST-MIGRATION-GRAPH-VALIDATION-967-974-876.md) were in the README table but their files were deleted on 2026-06-16. Table now matches disk.
• 2026-06-24 — Pruned 25 stale notes (71 → 47 files):
• Done/closed work (4): V27-OBSERVABILITY-SCOPE-803-805-426_2026-06-21.md (#803/#805 in #1047, #426 → #1052), CLOUD-PROVIDER-RESILIENCE-E2E-GAP-1003.md (#1003 closed), 1046-WHISPER-MULTI-MODEL-DESIGN.md (gate rejected, parked ideas live in sibling future-uses doc), RFC097_CHUNK9_PLAN.md (closed by #1073, see ADR-101).
• Session/handoff notes ≥3 days old (5): SESSION_BRIEFING_2026-06-19.md, SESSION_2026-06-20_1033_FOLLOWUPS_STATUS.md, SESSION_2026-06-20_FINAL_STATUS.md, OPERATOR_HANDOFF_NOTES_2026-06-21.md, NEXT_SESSION_PLAN.md (2026-06-13).
• Planning docs for shipped work (5): NEXT_BATCH_REGISTRY_RUNTIME.md (registry promotion in #1060), SPEC_1035_NER_PREPASS_DESIGN.md + EVAL_1035_NER_PREPASS_VERDICT.md (#1035 shipped), KILL_CODESPACE_COLLAPSE_TO_DEV_PROD.md (subsumed by docs/guides/DEV_PROD_ENV_DETECT_REMOVAL.md), WAVE-3-PLAN.md (audio waves 1–3 covered by docs/guides/AUDIO_PIPELINE_GUIDE.md).
• Eval verdicts for closed/superseded cohorts (8): EVAL_1016_FINAL_REPORT_2026_06_17.md (superseded by §11 in itself), EVAL_1016_OVERNIGHT_REPORT.md, EVAL_1016_ROUND3_REVIEW.md, EVAL_1033_COHORT_RERUN_2026-06-19.md, EVAL_112_ENTITY_FOCUSED_KG_2026-06-19.md, EVAL_113_SMALL_MODEL_STANDOFF.md, EVAL_116_CELL_C_REBASELINE_2026-06-20.md, EVAL-hybrid-search-validation.md (#1010 LanceDB shipped).
• Pre-impl reviews for shipped features (3): SPEAKER-ATTRIBUTION-PIPELINE-REVIEW.md (#876 shipped), TEST-SUITE-REVIEW.md (post-migration scan; action items addressed), CHUNK7_SILVER_REBUILD_RUNBOOK.md (chunk 7 audit landed in docs/guides/eval-reports/).
• 2026-06-16 — Removed 11 notes for shipped/closed work: APPROACH-913-909-964-993.md, AUDIO-WAVES-HARDENING-AUDIT.md (#964 done; #913/#400 closed), 967-interaction-cost-trace.md, 974-adfree-validation.md, POST-MIGRATION-GRAPH-VALIDATION-967-974-876.md (#967/#974 in #1010), DEP-EXTRAS-SEPARATION-1019-SCOPE.md (#1019), BATCH-PLAN-diarization-followups.md, SPOKEN_BY-REPROCESS-876.md, REHEARSAL-876-findings-20260609.md, NEXT_SESSION_HANDOFF-feat-946.md (#876/#946 done), COVERAGE-DEBT-deepgram-diarization-pr908.md (PR #908). Working notes now live in the git-ignored .journal/ (see AGENTS.md → Document location).
• 2026-05-22 — Removed GRAPH_NAVIGATION_HANDOFF_ANALYSIS.md; superseded by shipped graph handoff orchestrator (ADR-094, RFC-085).
• BAKEOFF-18EP-RESULTS.md (moved to the private eval repo) (moved to the private eval repo) — final medium-tier multi-provider bake-off on the 18-episode prod-v3 corpus, 8 clean arms scored with grok-4.3 as the single vendor-disjoint judge (gemini-2.5-flash-lite leads on surf/ep; sonnet-5 and gpt-5.5 dropped — empty-content failures on bundled evidence calls, the latter also triggering the entailment-fallback money-guardrail fix).
• BAKEOFF-OVERNIGHT-2026-08-05.md — v2.5 10-model bake-off frontier (9-ep control): qwen3.7-flash (value) + deepseek-v4-flash (quality) beat every Gemini tier/GPT/pro-reasoning model; gemini-2.5-pro GI-bug fixed + fairly re-judged. Active front.
• BAKEOFF-FINALE-INVENTORY.md — living work-inventory for the 100-ep finale (fix/improve/auto-research, tagged provider-agnostic vs deepseek-specific) + the iterate-deepseek-vs-proceed-to-qwen decision framework. Active.
• CORPUS_COMPARE_V2_V3_PILOT.md — deterministic-only metric comparison of v2-cloud vs v3-dgx over 9 shared episodes (joined by GUID): v3 names far more voices (voices_named +3.44) and nearly eliminates timeline drift (timeline_error_pct 1.92% → 0.19%), but surfaces fewer quotes and insights. Text-quality judging is out of scope here — tracked separately via the cross-vendor judge panel.
• OPEN — airgapped/local ML grounder at 8% coverage (EVAL_GROUNDING_WHO_FINDS_THE_QUOTE_2026_07) — the local extractive-QA + NLI grounder finds evidence for only 8% of insights vs 82% for the LLM (qwen) grounder. Two structural faults: no retrieval step, and the NLI verifier demands strict entailment that insights don't satisfy. Fix = add embedding retrieval and replace the verifier. Only affects the local/offline (airgapped) profiles — cloud/DGX profiles ground with the summarising LLM. This is the "tracked separately" item from the report; parked here (no GH issue opened) until scheduled.
• DGX-SERVICE-PERF-BASELINE-PLAN.md (moved to the private eval repo) (moved to the private eval repo) — isolated load/perf harness + baseline plan for each DGX inference service (diarization/Whisper/vLLM) after the 2026-08-04 OOM (#1397); scripts/perf/dgx_service_loadtest.py ramps concurrency + correlates homelab memory/Shmem. Per-service issues #1398/#1399/#1400.
• ISSUE-1540-ml-optional-assessment.md — blast-radius assessment for making local-ML opt-in (#1540). Split: CI wins (lever F, already implemented) stay in #1540; the torch-free runtime image (A), CI path-narrowing (E, re-opens #1527), and cloud NER prepass (D, latent-quality question) move to a discussion issue. [search]→torch is the whole cost; Search is hard-coupled to a locally-built LanceDB index (no keyword fallback). Gateway-embeddings path already exists — gated on whether LiteLLM serves /embeddings.
• ALLOY-DOUBLE-SHIP-FIX-2026-08-25.md — pipeline container logs reach VictoriaLogs twice (operator.alloy + legacy base.alloy docker sources; measured ×2.0). Podcast-repo half done (dashboards repointed); homelab half (remove the legacy block, root-owned on-box copy) documented with apply + verify steps. Open until the homelab half lands.
• SPEND-ESTIMATOR-OVERSTATEMENT-2026-08-28.md — run-reported LLM cost overstates gateway truth ~3.5× ($2.2685 estimated vs $0.6518 in SpendLogs over one sweep). Cause: LiteLLM returns cost in the x-litellm-response-cost HEADER, which the plain openai SDK never surfaces, so _openai_response_cost_usd always falls back to the direct-DeepSeek pricing table. Consequence: the $10 run cap halts runs ~3.5× earlier than intended. Open — fix not chosen (prefers reading the header).
• DEVICE-FEEDBACK-2026-09-19.md — the 2026-09-19 device round: every item the operator reported from the installed app, what was done, and what was NOT. Two real bugs behind the layout asks (storyline trend rows dead on tap — the client joined two lists that cannot cover the same set; Discover's "all" linked to the page you were already on), plus the Storyline/Theme vocabulary settled for interests (thc: = Storyline, tc: = Theme — the wire names invert against the reader-facing ones; see UXS-013 §Vocabulary). Screenshots per item in feedback-2026-09-19/. Open items listed in its NOT-done section — the in-app PDF gate and the Markdown filename are untouched.