choiyounggi · choiyounggi · Aug 7, 2026
@@ -1,53 +1,39 @@
-# Knowledge consolidation — 15 open PRs (#17–#40) → one reconciled state
+# Knowledge flush — 1 insight ingested (2 dropped as pending-duplicates of open PR #51)
 
-The 15 open `knowledge/*` PRs (created 2026-08-04 → 2026-08-05, before the
-harvest processed-store dedupe fix in #41) contained 123 file-versions of ~75
-unique pages, with the same insight landing at up to 3 different paths across
-up to 8 PRs. Per-PR review would re-import those duplicates, so — as with the
-#6–#13 consolidation — this branch carries the reconciled end-state and the 15
-PRs are closed in its favor.
+Queue drained: 3 pending candidates across 2 session files.
 
 ## Verified best-practice
 
-Every adopted page's sources were carried from its originating PR's flush, where
-they were live-verified at flush time; no new URLs were introduced during
-consolidation (checked mechanically: every `http(s)` URL in every merged page
-appears in a source PR's diff; every added body line in amended pages traces to
-a source PR hunk — orphan-line verification). Confidence fields were kept as the
-originating flushes set them, except client-side-rate-limiting where the union
-of provider-doc citations (Okta, Auth0, GitHub, OpenAI, RFC 6585) supports
-`verified` for the load-bearing claims. One subagent's fabricated content (12
-files matching neither main nor any PR, with invented source URLs) was detected
-by the same verification and replaced with true PR content.
+**Ingested — MAPE-aligned point prediction** (from `avm-hackathon-jycho` session):
+
+- **Claim:** A regression model trained to predict the conditional median (log target + L1 loss) and scored by MAPE is a structural overpredictor; the MAPE-optimal point prediction is the median of the 1/y-reweighted predictive distribution, which equals `median × exp(−σ²)` for a lognormal conditional. Correct per-row with `pred × exp(−λσ²)`, σ estimated from q16/q84 quantile spread, λ selected from 0.5 by multi-period holdout consistency rather than the theoretical 1.0.
+- **Sources checked:**
+  - https://arxiv.org/abs/0912.0902 — Gneiting, "Making and Evaluating Point Forecasts" (JASA 106:746–762, 2011). Table 5: the Bayes rule under absolute percentage error is the β-median with β = −1 (median of the y⁻¹-reweighted predictive distribution). Confirmed via WebSearch against arxiv.org and the tandfonline JASA record.
+  - https://arxiv.org/abs/1605.02541 — de Myttenaere, Golden, Le Grand, Rossi, "Mean Absolute Percentage Error for regression models" (Neurocomputing 2016). Confirms MAPE-optimal regression ≡ 1/y-weighted MAE regression. The citation given in the raw candidate was checked and is real (also mirrored at hal.science/hal-01312590).
+  - Lognormal algebra re-derived independently: for Y ~ LN(μ, σ²), y⁻¹·f(y) normalizes to LN(μ − σ², σ²), whose median is exp(μ − σ²) = median(Y)·exp(−σ²).
+- **Verification result:** the shrinkage mechanism and its direction are **verified** against the two papers plus the closed-form derivation. The practical λ = 0.5 starting point and the "quantile spread overestimates σ" claim rest on the session's holdout evidence only (Seoul commercial-building AVM, LightGBM + CatBoost: λ = 0.5 improved MAPE on all 4 holdout years, 19.18% → 18.86% excluding 2025 outliers; λ = 1.0 degraded 2 of 4 years) — that part is **field-tested** and the page's Sources section says so explicitly. Page frontmatter is `confidence: verified` for the sourced mechanism, with the field-tested scope of the λ practice stated inline.
+
+**Dropped candidates** (both from the linkly r3 orchestrate session) — see Open-PR check; neither was ingested, so no verification pass was spent upgrading them.
 
 ## Existing-layer check
 
-- Merged-main near-dup scan before consolidation: pairwise Jaccard over
-  title + "When this applies" across all 141 merged pages → **0 flagged pairs**;
-  previously merged content carries no duplication.
-- Cross-PR dedup during consolidation: 10 duplicate clusters collapsed to one
-  canonical page each (rate limiting 8→1, call-site enumeration 7→folded into
-  the canonical merged in #20, stderr/exit-0 diagnostics 4→1, sysroot 2→1,
-  env-off-switch 2→1, completion predicates 2→1, robots.txt 2→1,
-  harness-mediated results 2→1, leaked artifacts 2→1, orchestration category
-  naming unified). Three near-pairs kept distinct after trigger comparison,
-  with mutual `related:` links (differential setup vs interpretation; expansion
-  semantics vs off-switch design; import-time tactics vs level choice).
-- 24 existing pages received union-merged amendments; additions already present
-  in main (from #16/#20) were skipped, and all non-canonical `related:` ids
-  were remapped to canonical page ids (post-merge broken-link scan: 0).
+Routing went INDEX.md → backend → domain index. A full-text sweep of the checkout's `wiki/` for `mape`, `percentage error`, `lognormal`, `quantile`, `lightgbm`, `calibration` returned **zero hits** — nothing in any layer covers regression-model point-prediction calibration, so this is a create, not a merge. The closest existing category is `backend/common/llm`, whose pages cover *consuming* LLM completion APIs, not training/evaluating predictive models; I opened its nearest page to confirm the trigger space does not overlap (it gates on HTTP completion responses, not on metrics or point forecasts).
+
+Pages read: backend-common-llm-completion-response-validation
+
+No conflicts flagged. `related:` left empty — no genuinely adjacent page exists yet (the new category is a seed; future ML ingests should link here). Plumbing updated: `wiki/backend/index.md` gained a `### ml` section with a load-when line, the common-subtree route line and root `INDEX.md` backend line both mention MAPE-aligned point-prediction calibration, and `log.md` has the ingest entry.
+
+## Open-PR check
+
+Open `knowledge/*` heads listed via `gh pr list --search "head:knowledge/"`: #47 (`knowledge/dch0202-20260806-130040`), #49 (`knowledge/dch0202-rsquare-20260806-142309`), #50 (`knowledge/dch0202-20260806-172420`), #51 (`knowledge/dch0202-20260806-183029`), #52 (`knowledge/dch0202-rsquare-20260807-100149`).
+
+Per-candidate verdicts:
+
+- **Candidate 1 — worktree_escape guardrail escalation on cross-worktree reads** (linkly): **drop**. Fetched and diffed `origin/knowledge/dch0202-20260806-183029` (#51) against main: its `wiki/infrastructure/agent-orchestration/worktree-isolated-workers.md` additions already carry this insight in strictly better form — including the exact "budget the escalation round trip (read → approve → clear escalations/ → restart the watcher) and state in the worker's first brief that reads are approved" row. #51 additionally *corrects* the raw candidate: its local reproduction (guardrails 1.2.0) shows pure reads pass and the rule fires only when a main-root mention co-occurs with a write verb or an absolute-path redirect — the candidate's blanket "fires on read-only access" phrasing is the pre-correction version. Nothing unique to fold; retired as pending-duplicate.
+- **Candidate 2 — Orca dispatch binding taxonomy (idle-prompt check; runtime_unavailable vs agent_unconfigured vs terminal_worktree_mismatch)** (linkly): **drop**. The same #51 diff's `wiki/infrastructure/agent-orchestration/pane-delivery-confirmation.md` additions carry all four rows (bind only on idle prompt because "done" is a report not the turn's end; wait-and-rebind for occupied runtime; close pane + new worker-mode agent for a dead one; always pass worktree with pane) citing the same three 2026-08-06 field incidents the candidate cites. Nothing unique to fold; retired as pending-duplicate.
+- **Candidate 3 — MAPE-aligned point prediction** (avm-hackathon): **new**. No overlap with any open head — #47/#49/#50/#52 are testing/shell/encoding-themed and #51's wiki diff (11 files) touches no ML content. Ingested here.
 
 ## Routing decision
 
-- New categories: `infrastructure/agent-orchestration` (5 pages; unified the
-  competing `orchestration`/`agent-orchestration` names), `databases/data-survey`
-  (1), `qa/deliverables` (1). All other pages route into existing categories.
-- Canonical-path decisions: rate limiting → `backend/common/reliability/`
-  (sits beside timeouts-and-retries; 6 of 8 variants chose it); stderr
-  diagnostics → `platforms/processes/` (concern spans beyond shells); leaked
-  artifacts → `testing/data/artifact-leakage-from-a-suite`; call-site
-  enumeration → the existing `backend/common/change-impact/` page.
-- All 38 new pages listed in their domain indexes (nearest-index rule; backend
-  routes via its python sub-index for bytecode-cache-staleness); INDEX.md domain
-  summaries updated for infrastructure/qa/databases. Full-wiki lint: frontmatter,
-  ids, related-links, index coverage, size, qualifiers, staleness → 0 findings.
+- **Candidate 3 → `backend/common/ml/mape-aligned-point-prediction.md`** (id `backend-common-ml-mape-aligned-point-prediction`), **new category `ml`** under backend/common. Justification for the new category: the harvested domain hint was `backend`, and the backend domain's 12 existing categories (api-design, auth, caching, change-impact, concurrency, errors, integrations, jobs, llm, orm, reliability, storage) all cover server-side application code concerns; none covers training or evaluating a predictive model. `llm` is the nearest name but its scope is consuming LLM completion APIs from server code — putting metric-aligned regression calibration there would corrupt its load-when gate. No other domain fits better (databases owns SQL/schema; qa owns release process). The category seeds with one page.
+- Candidates 1–2: no routing — dropped as pending-duplicates of open PR #51 (see above); their queue rows are retired to `.processed.jsonl` so the auto-flush cannot re-surface them.
@@ -10,7 +10,7 @@ follow the cross-pointers in their index or take the next matching seeded domain
 | Domain | Status | Route here when |
 |--------|--------|-----------------|
 | [databases](wiki/databases/index.md) | **seeded** | Designing schemas/tables/keys, choosing or evaluating indexes, writing or optimizing queries, choosing transaction/isolation behavior, surveying live data to derive a rule, verifying additive migrations |
-| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, call-site enumeration before a contract change, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors, LLM completion validation & context budgeting, consuming external-API responses, externally-owned defaults, object-storage references) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps) |
+| [backend](wiki/backend/index.md) | **seeded** | Server-side application code — language-agnostic (`common/`: API contracts, call-site enumeration before a contract change, idempotency, JWT, timeouts/retries, caching, jobs, transactions in app code, shared state/pools, errors, LLM completion validation & context budgeting, MAPE-aligned point-prediction calibration, consuming external-API responses, externally-owned defaults, object-storage references) plus stack subtrees: `java/` (JPA, Spring proxies, JVM threads/memory), `node/` (event loop, promises, runtime validation, shutdown), `python/` (GIL/asyncio, pydantic, WSGI/ASGI workers, language traps) |
 | [frontend](wiki/frontend/index.md) | **seeded** | Web UI code: state placement, rendering performance, in-UI data fetching (races, infinite scroll), auth token handling, forms, XSS-safe output, accessibility |
 | [infrastructure](wiki/infrastructure/index.md) | **seeded** | CI/CD pipelines, secrets in build/deploy, container image builds, rollout/rollback strategy, observability (logs/metrics/alerting), per-environment/path-valued config, multi-agent orchestration (worker liveness signals, shared run state, tmux pane delivery, completion gates, worktree-isolated workers) |
 | [testing](wiki/testing/index.md) | **seeded** | Writing or structuring automated tests: level choice, cases/assertions, test data, mock decisions, flaky tests (release-process quality → qa) |

@@ -43,3 +43,4 @@ Append-only. Format: `## [YYYY-MM-DD] <ingest|revise|lint|gap|contradiction|drif
 ## [2026-08-06] revise | Union-merged amendments from the same 15 PRs into 24 existing pages (largest: tests-that-cannot-fail — 7 PRs: per-assertion mutation granularity, restore-mechanism-by-commit-state, suite-total-drop detection; portable-shell-scripts — 8 PRs: POSIX inline set-- reordering, ${VAR:-} vs ${VAR-} colon semantics, quote-by-what-the-text-is; test-data-and-isolation — 6 PRs: env-derived write paths, absent-variable cases, exec-bit fixtures, leak attribution; change-impact/call-site-enumeration — 8 unique additions from 7 path-variants: rename-in-place, same-type reorder, *args forwarding, two-searches-agree fallacy, codemod, coverage completion check). All non-canonical cross-page ids remapped to canonical pages; every added line traced to its source PR hunk (fabricated worker output was detected by orphan-line verification and replaced with true PR content).
 ## [2026-08-06] dedup | Cross-PR duplicate clusters collapsed: client-side rate limiting (8 PR versions at 3 paths → 1), call-site enumeration (7 versions at 7 paths → additions folded into the already-merged canonical page), stderr/exit-0 diagnostics (4 versions at 4 paths → 1), macOS sysroot (2→1), env-var off switch (2→1), completion predicates (2→1), robots.txt (2→1), harness-mediated results (2→1), leaked test artifacts (2→1), orchestration category naming (orchestration vs agent-orchestration → agent-orchestration). Kept distinct after trigger comparison: differential-testing vs differential-run-agreement (setup vs interpretation), unset-versus-empty-parameters vs env-var-off-switches (semantics vs design, cross-linked), import-time-side-effects vs test-level-choice edge (tactics vs level choice, cross-linked). Merged-main near-dup scan (Jaccard over title+trigger, 141 pages): 0 pairs — no duplication among previously merged content.
 ## [2026-08-06] ingest | Reconciled post-consolidation flush PRs #42–#43 (created against pre-#44 main). Folded unique content: platforms/filesystems/permissions-and-exec-bits (+`sh "$SCRIPT"` interpreter-invocation rows — stubs need no exec bit, avoids EDR chmod+x heuristics), platforms/shells/command-text-inspected-before-execution (+gate-AUTHOR step 8: parse all three POSIX quoting forms, expand only ~/$HOME/${HOME}; red-then-green bats evidence), testing/quality/checks-that-cannot-pass (+vacuously-green pre-implementation test row), platforms/processes/tool-diagnostics-without-a-failing-exit-code (+repeat-suppression edge, +clang reproduction source), infrastructure/ci-cd/secrets-handling (+independent credential channels row: gh token vs git SSH vs API token, git push --dry-run / author:@me probes), testing/mocking/what-to-mock (+ESM read-only-bindings DI rows, +PATH-emptied tripwire negative-proof). Dropped as duplicates of the #44 state: #42's rate-limit row (already in timeouts-and-retries + client-side-rate-limiting), #42's warning-only-diagnostics page (same trigger as tool-diagnostics-without-a-failing-exit-code), #42's leak-attribution row (already in test-data-and-isolation).
+## [2026-08-07] ingest | backend-common-ml-mape-aligned-point-prediction — NEW category backend/common/ml (no existing category covers training/evaluating predictive models; llm covers only consuming LLM APIs). A median-predicting regression model (log target + L1) scored by MAPE is a structural overpredictor: the MAPE Bayes rule is the 1/y-reweighted median (Gneiting arXiv:0912.0902 Table 5; de Myttenaere arXiv:1605.02541), = median × exp(−σ²) under lognormality. Directive: per-row `pred × exp(−λσ²)` with σ from q16/q84 quantile spread, λ selected from 0.5 by all-periods holdout improvement (theory λ=1.0 overcorrects because spread-estimated σ is inflated). Theory verified vs sources; λ practice field-tested (Seoul AVM, 4/4 holdout years improved at λ=0.5, 2/4 degraded at λ=1.0). Two queued dev-loop orchestration candidates dropped as pending-duplicates of open PR #51.