Files
waggle-os/docs/research/03-paper-skeleton-v2-2026-04-30.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

21 KiB

Arxiv Paper Skeleton v2 — Pre-Launch Submission Ready

Date: 2026-04-30 Author: PM Status: SKELETON DRAFT — awaiting Marko 7 decision points ratifikacija → drafting full content Predecessor: 00-paper-outline.md (2026-04-26) + 02-section-5-refresh-2026-04-30.md Target venue: arxiv preprint (cs.AI primary, cs.CL secondary), Day 0 launch submission Authority: decisions/2026-04-30-pre-launch-sprint-consolidation-LOCKED.md Wall-clock: 7-9 dana drafting (PM) + 1-2 dana review (Marko) = ~10 dana to submission


§1 — Marko 7 decision points — RATIFIED 2026-05-02

1. Title — RATIFIED: "Apples-to-Apples on LoCoMo: A Bitemporal Local-First Memory Substrate and a +27.35-Point Methodology Gap."

Marko rationale: methodology-led, dvodelno, konkretan broj kao hook. Najjači za arxiv discoverability jer "methodology gap" privlači citate iz svake naredne LoCoMo evaluacije. Marko REJECTED PM rec (d) jer cross-family generalization je sekundarna narativna linija, ne primary anchor; primary contribution je substrate vs Mem0 + +27.35pp self-judge bias methodology gap.

Abstract refinement (Marko ratified): prva rečenica metodološka, druga arhitekturna, treća broj. Reorder iz prethodne 4-claim strukture da odražava ovaj sequence.

2. Co-author roster — RATIFIED 2026-05-02 final:

  • Lead/corresponding: Marko Marković, CEO Egzakta Group
  • Co-author 1 (Egzakta technical lead): Michail (Mikhail) Pavlukhin
    • LinkedIn (Marko-confirmed 2026-05-02): https://www.linkedin.com/in/mpavlukhin/ (rs.linkedin = Serbia)
    • Profile verified via WebSearch 2026-05-02: Public output = DSPy + GEPA practitioner tutorials ("Improving performance of small LLMs on creative tasks with DSPy and GEPA"), Timeboat Adventures experimental narrative game, AI engineering content creation.
    • EVOLVESCHEMA author — VERIFIED per memory/project_benchmark_strategy.md LOCKED 2026-04-20: Pavlukhin je first-author EVOLVESCHEMA paper "LLM-Guided Evolutionary Optimization of DSPY Signature Schemas" (Marko uploadovao PDF 2026-04-20). Paper demonstrates composition pipeline EVOLVESCHEMA→GEPA, najbolji javno objavljeni brojevi 0.925 FIRE NER, 0.903 SGD Hotels (per Pavlukhin Table 2). Teza: "schema structure matters more than instructions". 74% HotPotQA gain od jedne strukturne mutacije (replace_output_fields).
    • Strategic alignment: Pavlukhin radi tačno u methodology orbiti našeg §5.4 GEPA finding + §2.4 Related Work EVOLVESCHEMA reference. Coauthorship strengthens cross-citation network and signals Egzakta team deep methodological investment u trojnoj kompoziciji optimization layer-a.
    • Arxiv presence: EVOLVESCHEMA paper trenutno NIJE NA ARXIV-U (PDF was internal Egzakta material per Marko 2026-04-20 upload). WebSearch arxiv author profile 404 + zero hits potvrđuje no prior arxiv first-author submission yet. → cannot currently endorse, ALI strategic option below.
    • STRATEGIC OPTION (Pavlukhin EVOLVESCHEMA arxiv submission): Pavlukhin submit EVOLVESCHEMA kao standalone short preprint (4-6 strana, cs.AI primary) u sledećih 3-5 dana. Three benefits: (1) Pavlukhin postaje arxiv author → Scenario A re-activated (auto-endorsement kroz coauthorship); (2) Cross-citation network: EVOLVESCHEMA u §2.4 + §5.4 composition pipeline reference; (3) Egzakta team javan akademski signal pre Day 0 launch. Marko 2026-05-02 to ask Pavlukhin if feasible.
  • Co-author 2 (methodology consult): DROPPED per Marko ratification 2026-05-02 (Opcija C). Paper ide kao 3-author. Barać preuzima methodology cross-check rolu pored academic advisor.
  • Co-author 3 (academic advisor): Dušan Barać
    • Scholar profile (primary, Marko-confirmed 2026-05-02): https://scholar.google.com/citations?user=WvzsdGkAAAAJ&hl=sr&oi=ao
    • Scholar profile (alternative ID seen earlier, possibly stale): https://scholar.google.com/citations?user=wn7N_HsAAAAJ&hl=sr&oi=sra — confirm sa Markom da li je isti čovek dva profila ili greška
    • Profile verified via web search 2026-05-02: Full Professor, Department of E-business, Faculty of Organizational Sciences (FON), University of Belgrade. 56 publikacija ukupno; 25+ u peer-reviewed international journals sa impact factor; 8 papers u SCI/SCIe journals. 18+ godina IT consulting experience. Primary research domains: digital transformation, e-commerce, IT project management, e-learning technologies, AI-based apps, blockchain.
    • Academic credibility: STRONG — Full Professor + University of Belgrade (recognized institution) + SCI track record. Strong co-authorship credibility signal for peer reviewers.
    • Methodology cross-check capacity: ADEQUATE — IT/e-business background covers general empirical research methodology rigor; možda not specifično cs.AI/judge ensemble expertise, ali general academic peer-review standards covered.
    • Arxiv endorsement potential: WEAK — primary domain je e-business / IT ecosystems, NE cs.AI/cs.CL/cs.LG; plus Marko confirmed "nema na arxiv" → no prior arxiv first-author submission → cannot endorse cs.AI primary submission. Domain mismatch + zero arxiv presence = double pre-empt.

3. Endorsement path — RESOLVED 2026-05-02 final via Path A + Path D combination.

Verified results:

  • Pavlukhin: NEMA arxiv first-author submission yet — but EVOLVESCHEMA standalone arxiv submission planned (Marko-side action sutra 2026-05-03 ponedeljak) → Path A reactivated
  • Barać: NEMA arxiv first-author submission + e-business domain mismatch → cannot endorse
  • LinkedIn 1st-degree network sweep (Bojan Djuric verified 2026-05-02): no arxiv-eligible candidate match without cold outreach
  • Marko 2026-05-02 verdict: cold outreach is out of scope for current sprint

Path A — PRIMARY (Marko-side action sutra): Pavlukhin podnosi EVOLVESCHEMA na arxiv kao standalone short preprint (4-6 strana, cs.AI primary). Posle arxiv approval (~24h), Pavlukhin postaje cs.AI/cs.LG eligible endorser. Critical insight: Pavlukhin NE MORA biti coauthor našeg paper-a — endorses kao nezavisni researcher kroz EVOLVESCHEMA author credentials. Single email request to direct contact, NE cold outreach. Timeline: 3-5 dana Pavlukhin draft + 24h approval = ~5-6 dana to endorsement-ready. Compatible sa Day 0 ETA 6-10 dana.

Path D — FALLBACK + DECOUPLING: Day 0 launch ne treba biti gated na arxiv preprint timing. Landing trust signal "Published methodology — arxiv preprint" (Trust Band Card 4 placeholder + Sources reference) zamenjuje se sa "Open methodology — github.com/marolinik/waggle/docs/methodology" (markdown doc u OSS repo). Credibility signal ostaje (open methodology = open source rigor), arxiv timing više nije launch dependency. Posle arxiv preprint linkujemo retroactively kao news cycle update.

Landing copy implications (apply paralelno sa Track D apps/www port):

  • Trust Band Card 4 swap: "Published methodology — arxiv preprint" → "Open methodology — github docs"
  • Footer Research column: "arxiv preprint" link → defer to placeholder ili remove until arxiv preprint dođe
  • Hero Variant D Sasha (developer): no copy change needed (no arxiv mention)
  • Proof Card 1 GEPA: keep "Methodology in arxiv preprint" — to je forward reference, cleanly fulfilled when Path A succeeds OR replace sa "Methodology in companion docs" if delayed

PM action queue (sutra 2026-05-03):

  • Marko 1-line Pavlukhin message za EVOLVESCHEMA arxiv submission timing
  • PM update Track D apps/www CC sesija D §3 acceptance review sa Trust Band Card 4 copy swap
  • PM draft methodology markdown doc skeleton za github (post Day 0 link target)

3. Endorsement path — DEFERRED: Marko asked "kako da proverim ko od mojih kontakata ima prior arxiv publications" — PM follow-up sa 3 actionable proverava metoda (vidi §1bis below).

4. Multiplier section disposition — RATIFIED with sharper framing:

NOT "Branch B prerequisites for re-test (deferred)" — that reads evasive after Zep/Mem0 dispute reviewer-instinct.

INSTEAD: Eksplicitna "Negative Result" subsekcija u §6 Discussion sa 5 elementa:

  1. Hypothesis statement
  2. N=12 protocol
  3. Numbers verbatim: h2=1/3, h3=0/3, h4=0/3
  4. Preconditions for re-test imenovane: B-frame compaction stability, harvest timestamp fix, minimum N
  5. Qualification: "ne aplicira na primary contribution (substrate-vs-Mem0 + methodology gap)"

"Deferred" ide u footnote, NE u headline. Negative result se hvata u CV publikacije jer je honest, ne uprkos tome.

5. GEPA scope u §5.4 — RATIFIED (a) sa companion paper signal:

Zadržava ~1.5 page §5.4 sub-section. PLUS footnote: "Extended cross-family treatment in companion paper, in preparation."

Marko rationale: standalone §5 GEPA bi te koštao desk-reject rizika ("this is two papers stitched together") što ne želiš na prvoj submisiji. Cross-family generalization sa GEPA Faza 1 je dovoljno jako da nosi sopstveni arxiv preprint za 6-8 nedelja sa ERL methodology framing. Footnote daje opciju za drugi launch ciklus bez current commitment. Ako kapaciteta za companion paper nema, footnote se izostavi pre submission.

6. §5.5 framing — RATIFIED with active-voice rewrite:

NOT "methodology maturity demonstration" — chita se defanzivno, apologetičko.

INSTEAD: "Bias-detection guardrails functioning as designed" — aktivan claim sa proverim outcome:

  • "GPT selection bias detected and filtered by held-out validation at N=..."
  • "Qwen-non-thinking decoupling probe revealed effect Z magnitude"
  • "Calibration evolved across Amendments 7-11 with adjustments documented in Appendix"

Aktivni glagoli protiv pasivnih: "guard-rail caught and filtered" vs "bilo je pažljivo." Druga formulacija nosi metodološku težinu, prva zvuči apologetski.

7. Phase 5 forward reference — RATIFIED with two terminological refinements:

REVISED forward reference statement: "Production traffic Pass II rates, p95 latency, and recall@K on production traffic distribution from Phase 5 deployment will be reported in v2 of this preprint, scheduled within 60 days post-publication."

Refinement 1: "expected ~6 weeks" → "scheduled within 60 days" (operational discipline signaling + 14-day margin if Phase 5 slips two weeks; missed v2 datum bi bio credibility liability za naredni paper).

Refinement 2: imenovati konkretne metrike (Pass II rates + p95 latency + recall@K) — vague forward reference izgleda kao vaporware; specifična obećana metrika izgleda kao discipline.


Marko net evaluation note (2026-05-02): PM intencije su sve ispravne; tri od četiri framing-a (4, 6, 7) treba da se pomere stepenicu ka aktivnijem, ne dodatak na sadržaj. Cilj nije da paper deluje skromno ili pažljivo — cilj je da deluje tačno.

Posle ovih ratifikacija, drafting kick-off može da krene paralelno sa rešavanjem Co-author 2 + Endorsement path.


§1bis — Endorsement path actionables (PM follow-up za Decision 3)

Marko question: "kako da proverim ko od mojih kontakata ima prior arxiv publications". 3 metoda po brzini:

Method A — arxiv search (5 min per name):

  1. Otvori https://arxiv.org/search
  2. Unesi "FirstName LastName" u Author search
  3. Filter: cs.AI, cs.CL, cs.LG (relevant categories)
  4. Ako se pojavi makar 1 paper kao first author → endorsement candidate. Ako appears samo kao co-author → weaker, endorsement may still work ako paper je u cs.AI specifically (arxiv requires first-author publication u target category).

Method B — Google Scholar (3 min per name):

  1. Otvori https://scholar.google.com
  2. Search "FirstName LastName" + relevant keyword (e.g., "memory" or "agent")
  3. Ako ima Scholar profile → click → vidi publication list. cs.AI/cs.CL papers = endorsement potential.
  4. Bonus: h-index ≥3 u relevant category = strong endorser.

Method C — Brain dump → PM verifies: Marko mi da listu 10-20 imena (LinkedIn 1st-degree connections u academic + AI research). PM uradi Method A + B za svaki kroz web search. Output je rang-lista candidates sortirana po endorsement probability + first-message draft za top 3.

PM preporuka: Method C najbrži za Marka (sve što treba je brain-dump 20 imena), ja u 30-45 min vratim ranked list. Pavlukhin kao confirmed Co-author 1 — proveriti da li ima cs.AI prior publications (ako da, automatic auto-endorsement, NEMA need za external).

Plus za Co-author 2 methodology consult — kandidati koji dolaze iz EVOLVESCHEMA, GEPA (Agrawal), ACE (Zhang) author networks su strong (već su empirical AI rigor). Pavlukhin može introduce.


§2 — Section structure (10 sekcija, ~10-13 strana)

Abstract (≤ 250 words)

Three-claim structure (preserved iz v1):

  1. Architectural separation of memory substrate from retrieval is novel and methodologically important.
  2. Substrate beats peer-reviewed Mem0 on LoCoMo oracle ceiling: 74% vs 66.9%.
  3. We document +27.35pp self-judge bias; we propose trio-strict ensemble as remedy.

NEW fourth claim (per §5.4 refresh): 4. Genetic-evolutionary prompt adaptation (GEPA) on Waggle harness produces +12.5pp Pass II uplift on held-out validation; method generalizes cross-family (Claude flagship + Qwen 35B open-source); on-prem Qwen achieves Opus-class quality on validated samples.

Word budget: 230-250 words.

1. Introduction (~1 page)

Subsections:

  • 1.1 Problem framing (LLM agents lack persistent memory; products bundle layers; independent measurement impossible)
  • 1.2 Why substrate-retrieval separation matters (no one knows which layer to improve)
  • 1.3 Contributions (4 — architectural + empirical + methodological + GEPA cross-family + open)
  • 1.4 What this paper is not (no system-level competitive claim, V1 retrieval acknowledged)

Subsections (preserved iz v1):

  • 2.1 Memory systems for LLMs (Mem0, MemGPT, Letta, LongMem, GraphRAG)
  • 2.2 Long-term conversational memory benchmarks (LoCoMo, LongMemEval; Gaia2 + τ³ za future work mention)
  • 2.3 Methodology critiques in LLM evaluation (self-judging bias, judge ensembles, reproducibility)
  • 2.4 NEW: Genetic-evolutionary prompt adaptation (Agrawal et al. 2025 GEPA, Pavlukhin et al. 2025 EvolveSchema)
  • 2.5 NEW: Experiential reflective learning (ERL — arxiv:2603.24639; Waggle bitemporal-KG-conditioned extension framing)
  • 2.6 NEW: Closed-learning-loop agents (Hermes Agent positioning differentiation)

3. Architecture (~2 pages)

Subsections (preserved iz v1):

  • 3.1 Substrate: bitemporal knowledge graph
  • 3.2 Memory compression: I/P/B frame analogy (MPEG-4 inspired)
  • 3.3 Retrieval interface (intentionally pluggable)
  • 3.4 Audit + governance layer (EU AI Act compliance, sovereign deployment)

4. Methodology (~1 page)

Subsections (preserved iz v1):

  • 4.1 LoCoMo-1540 evaluation protocol (5 cells, 5 question types)
  • 4.2 Trio-strict judge ensemble (Opus + GPT + MiniMax, κ_trio = 0.79)
  • 4.3 Apples-to-apples re-judging (GPT-4o-mini self-judge methodology re-eval)
  • 4.4 Pre-registration & reproducibility (manifest v6 frozen, git SHA, seeds locked)
  • 4.5 NEW: GEPA evolution methodology (manifest v7 + 11 amendments, held-out validation, §F gates)

5. Experiments & Results (~4-5 pages)

5.1 Substrate quality (oracle ceiling)

  • Claim: 74% trio-strict (N=400)
  • Compared: Mem0 peer-reviewed 66.9% basic / 68.4% graph
  • Headline: +7.1pp over peer-reviewed baseline
  • Statistical test: Fisher one-sided p < 8.07e-18 vs no-context (H1 PASS)

5.2 Methodology bias quantification

  • Self-judge: aggregate +X.X% inflation
  • Trio-strict: true performance band
  • Measured gap: +27.35pp
  • Implication: published "91.6% LoCoMo SOTA" claims likely 18-30pp inflated

5.3 V1 retrieval performance

  • Retrieval cell: 48% trio-strict
  • Honest: 26pp below substrate ceiling
  • Direction: 5 specific improvements identified (community contribution invitation)

5.4 GEPA Evolution + Cross-Family Generalization (NEW, ~1.5 page)

Subsections:

  • 5.4.1 GEPA evolution methodology summary (~0.3 page)
  • 5.4.2 Cross-family generalization results (~0.5 page) — table:
    • claude::gen1-v1: +12.5pp in-sample / +12.5pp held-out / 0pp gap
    • qwen-thinking::gen1-v1: +12.5pp / +12.5pp / 0pp
    • gpt::gen1-v2: +25.0pp / +5.0pp / 20pp gap (FAIL §F.5 cond_2)
    • §F.1 PASS 3/3, §F.5 cond_2 PASS 2/3
  • 5.4.3 Qwen 35B = Opus-class on out-of-distribution (~0.4 page)
    • Retrieval engagement 2.231 = 96% Opus parity
    • Pass II 100% on N=13 (8 in-sample + 5 held-out)
    • Sovereign deployment without flagship-tier compromise
  • 5.4.4 Phase 5 deployment status (~0.2 page) — both AUTHORIZED variants
  • 5.4.5 Pilot 2026-04-26 disposition (~0.1 page) — multiplier conditional finding

5.5 Scoping Discussion (NEW, ~1.2 page)

4 methodological findings:

  • 5.5.1 gpt selection bias exposed by held-out methodology (~0.3 page)
  • 5.5.2 qwen-non-thinking retrieval-quality decoupling (~0.3 page)
  • 5.5.3 generic-simple necessary-but-not-sufficient retrieval (~0.2 page)
  • 5.5.4 Calibration evolution Amendments 7-11 (~0.4 page)

5.6 Phase 5 Production Validation Hook (NEW, ~0.2 page)

Forward reference za v2 preprint sa production traffic results.

6. Discussion (~1-2 pages)

Subsections:

  • 6.1 Substrate-retrieval separation as paper contribution
  • 6.2 Sovereignty axis (local-first + Apache-2.0; KVARK secondary downstream)
  • 6.3 Cross-family generalization implications (NEW per §5.4)
  • 6.4 Limitations (single benchmark + retrieval gap + multiplier conditional)

7. Future Work (~0.5 page)

  • Retrieval V2 (entity-aware reranking, hybrid scoring)
  • Cross-dataset replication (LongMemEval + Gaia2 + τ³-bench banking_knowledge — per benchmark portfolio Phase 3 + Phase 4 sprints)
  • Faza 2 sprint (gpt N=16 re-validation + scoping investigations)
  • Multiplier thesis at scale (Branch B prerequisites)
  • ERL methodology extension (per arxiv:2603.24639) submission MemAgents Workshop

8. Conclusion (~0.25 page)

Restate 4 contributions. Restate Apache-2.0 community invitation.

9. Acknowledgments

  • Anthropic (Claude AI assistance, explicit per ICML/NeurIPS/arxiv standards — Claude does NOT appear as author)
  • LiteLLM gateway za multi-vendor judge orchestration
  • LoCoMo dataset authors

10. References (~30-50 BibTeX entries)

Categories:

  • Memory systems (Mem0, MemGPT, Letta, LongMem, GraphRAG): 8-10
  • Conversational memory benchmarks (LoCoMo, LongMemEval): 5-7
  • LLM evaluation methodology (judge bias, ensembles): 6-10
  • Genetic-evolutionary prompt adaptation (GEPA, EvolveSchema): 3-4
  • Experiential reflective learning (ERL): 2-3
  • Bitemporal databases + KGs: 4-6
  • MPEG-4 video compression analogy: 2-3
  • Reproducibility & pre-registration: 3-5

§3 — Claim trace (where each numerical claim comes from)

Preserved iz v1 sa NEW additions:

Claim Source artifact Verification path
74% oracle ceiling benchmarks/results/v6-trio-strict/oracle-context-*.summary.json Stage 3 v6 N=400
66.9% Mem0 baseline Mem0 paper Table 2 arxiv:2504.19413
+7.1pp gap direct subtraction verified
Fisher p < 8.07e-18 Stage 3 v6 H1 benchmarks/results/v6-trio-strict/h1-fisher-result.json
+27.35pp self-judge bias Apples-to-apples re-eval benchmarks/results/v6-self-judge-rebench/
κ_trio = 0.79 Stage 3 v6 calibration benchmarks/calibration/2026-04-24-trio-strict-recal.json
+12.5pp held-out (claude::gen1-v1) Faza 1 §F.1 PASS decisions/2026-04-29-gepa-faza1-results.md
+12.5pp held-out (qwen-thinking::gen1-v1) Faza 1 §F.1 PASS same
0pp held-out gap (claude + qwen-thinking) §F.5 cond_2 PASS same
20pp gap (gpt::gen1-v2 FAIL §F.5 cond_2) held-out vs in-sample same
96% retrieval parity (qwen-thinking) retrieval 2.231 / Opus 2.33 Faza 1 evidence
100% Pass II combined N=13 Faza 1 Checkpoint C same
11 amendments manifest v7 Faza 1 manifest chain decisions/2026-04-29-gepa-faza1-results.md §11-amendment chain

Every numerical claim mora trace to JSON artifact + git SHA + manifest anchor. No hand-computed numbers u paper text.


§4 — Authoring sequence

Day 0 (today, 2026-04-30): This skeleton authored. Pending Marko 7 decision points ratifikacija.

Day 1 (post-ratifikacija): PM autoring abstract + introduction (~1500 words).

Day 2: PM autoring related work + architecture (~3000 words).

Day 3: PM autoring methodology (~1500 words).

Day 4-5: PM autoring experiments & results — heavy lifting (~5000 words sa §5.4 + §5.5 + §5.6 detail).

Day 6: PM autoring discussion + future work + conclusion (~2000 words).

Day 7: PM autoring acknowledgments + references (BibTeX entries).

Day 8-9: Internal review (Marko + co-authors) + revisions.

Day 10: LaTeX formatting + final polish + arxiv submission.

ETA submit: Day 10 wall-clock from ratifikacija. Posle submission, preprint live within 1-2 days.


§5 — Cross-references

  • v1 outline: 00-paper-outline.md
  • §5 refresh: 02-section-5-refresh-2026-04-30.md
  • Faza 1 closure: decisions/2026-04-29-gepa-faza1-results.md
  • Pre-launch sprint consolidation: decisions/2026-04-30-pre-launch-sprint-consolidation-LOCKED.md
  • Benchmark portfolio refresh: briefs/2026-04-29-benchmark-portfolio-refresh-2026-venues.md
  • This skeleton: research/2026-04-26-arxiv-paper/03-paper-skeleton-v2-2026-04-30.md

End of skeleton. Awaiting Marko 7 decision points ratifikacija → drafting kick-off.