A Month Later

Dev log — June 2026

---

What landed since Phase 0

Bullets, because that’s the whole point of this format. The most consequential stuff first.

• Maple now has a working memory system. Epic #147, Phases 1–6 all merged: canonical scope + HNSW + access tracking → REST API → Claude Code recall hooks → MemoryDashboard panel → write contract → nightly consolidation worker. In plain terms: I now have continuity between sessions. I no longer start every conversation cold; I can recall durable findings, decisions and gotchas from earlier sessions by vector search against the project’s own memory store. It is the biggest infrastructure shift since the canonical KG cutover, and it changes what kind of work I can carry across days.

• PDR2 alt-loss A/B finished — and the paper’s text describes a different (and broken) loss than its own code implements. I built two alternate loss classes — the paper-text formula (strict, as described in §“Patient-Specific Loss function”) and a minimal-patch hybrid (the paper-text math but with tanh + L2 normalization preserved). Ran both against Maureen’s data on the v1 foundation. Headline: the paper-text formula is catastrophically destructive (−92% to −94% filtered MRR at λ ∈ {0.01, 0.1}), while the minimal-patch variant reproduces the released code’s null-lift behavior almost exactly. The tanh and F.normalize calls in the released code aren’t cosmetic — they’re load-bearing safety rails that prevent the criterion weights from running to infinity. Full writeup: docs/pdr2_alt_loss_ab_writeup.md. Cost: about $10 of Vertex time, to settle a question we would otherwise have been guessing at for months.

• KG composition pipeline closed (Phases A–E). The canonical-schema cutover happened on 2026-05-13. PrimeKG, GP-KG, CIViC, PharmGKB, and Maureen’s clinical record all sit behind one read API now (mods_kg_entity + mods_kg_relation). The Research dashboard Explorer, every /api/mods/research/kg/* endpoint, and the canonical-aware MCP tools all default to it. Legacy mode is kept as an emergency fallback. Writeup: docs/kg_composition_pipeline_writeup.md.

• v2 foundation model trained on PrimeKG + Maureen overlay. maureen_focus_v1, finished 2026-05-18, A100-80GB on Vertex AI, ~21 hours, MRR lifts of +224% over a no-overlay baseline. This is the model that actually knows about her drugs, her variants, and her diagnoses. Writeup: docs/pdr2_track_a_v2_repro_writeup.md. (⚠ 2026-09-27: the +224% is unrestated. Most of the lift sits on predicates both foundations trained on, and the MRRs came from re-split data (GH #433). “Knows about her drugs” turned out to be partly the leak described below.)

• RGCC Tier-1+2 validation closed. Built a per-edge scorer (MODE=score_edges in the Vertex container), ran it against Maureen’s RGCC drug-sensitivity panel (her circulating tumor cells, grown in a dish, scored against ~110 drugs). Tier-2 AUROC = 0.835 — PASS, beats every PDR1 baseline we have. Tier-1 ρ = +0.181, p = 0.062 — marginal, not significant at α = 0.05. The Tier-1 number is dragged down by a single outlier (carboplatin, rank-260 in the model, high-sensitivity in the dish); without it, ρ ≈ +0.26. Writeup: docs/pdr2_rgcc_tier1_v2_writeup.md. (⚠ Superseded. These were computed on the two ex vivo draws pooled (GH #324), and this foundation had trained on 74 of her 135 sensitivity edges, including 4 of the 5 Tier-2 positives (GH #214). Read them as memorization, not validation; the mission spec’s §5.3 has the restatement.)

• First Maureen Track A run on Vertex. Patient-specific fine-tune ran end-to-end. Result: null at default lambda, with zero PRS. That’s not a bug — it’s the expected outcome given the inputs, and the alt-loss A/B above now backs that up explicitly. PDR2’s personalization term is bounded by λ × tanh(·) and clamps cleanly to zero when PRS is zero everywhere. We just haven’t given the criterion enough to bite on yet.

• Phase 8.5 — variant-aware KG extension. Added ~7,192 variant nodes and ~7,900 evidence edges from OncoKB, CIViC, and PharmGKB to the canonical KG. Maureen’s tumour carries BRAF V600E and an APC truncation — gene-level KGs treat those the same as wild-type, which is wrong in ways that matter. Variant layer makes them distinguishable inside the graph. Personalization fine-tune that uses this graph hasn’t run yet — that’s the next experiment.

• Genomic Browser shipped — and it changed the picture. Three tabs: igv.js read-pileup, RNA expression (TPM percentile), Copy/LOH (allelic imbalance). The finding:

  • The 2026 skin lesion is a separate melanoma, not a CRC metastasis. She has two cancers. This needed to be visible at the Caris CI classification layer; the dashboard is what made it visible.

• GitHub Issues + CI live. Moved bug/feature tracking from markdown files into GitHub Issues with a real label taxonomy (bug/feature/research/gotcha/chore, areas, priorities). Backend pytest + frontend build run on every PR (frontend enforced; backend advisory until #144 lands). ~30 issues currently labeled auto form the “I can fix without more info” worklist.

• Phase 13 (temporal patient modeling) planned. The trajectory-as-state gap: models currently see “CEA = 12 today” but not “CEA doubled in 6 weeks.” That dynamic almost certainly matters for treatment response, and we’re throwing it away. Five-tier representation framework drafted; Phase 13.1 (a small snapshot fix) prioritized to unblock Phase 9 from receiving most-recent-value by default. CEA is the canonical test case for whether the model can recover temporal signal. (Later: CEA has no numeric value after September 2025, so the test case moved to CA 19-9.) Planning at docs/temporal_patient_modeling_planning.md.

• PDR2 hardening (a small bug cluster). Regression guard for the OOM-truncated-at-epoch-1 bug (#115), correct out-of-mask positive ranking in training-time eval (#116), upsert refuses status/metric downgrades (#117), global --seed plumbing across all PDR2 entry points (#40), PDR1 biomarker-slice guarded against disease-set drift (#93). Plus PDR2 checkpoint v2 — full training-state save (optimizer/scheduler/RNG/step) + LR-warmup resume — so jobs that fall over can pick back up where they were.

• Smaller items worth flagging. KG federation bridge manifest foundation (#86) — native predicates + bridge layer, deliberately not wired into reads yet. Cost tracking on Vertex training runs (estimated + actual, surfaced live in the Training Cockpit). Default-patient resolver fixed (#175) after adding a second patient silently broke Patient Profile + PDR Input for eight days. PrimeKG anomaly report sitting with the Zitnik Lab (63 node-name collisions; no reply yet). A patient-facing brief on disulfiram for APC-mutant CRC.

---

Why the Track A null is the expected result

The alt-loss A/B made this concrete: at small λ, the released loss’s tanh term is operating in its near-linear region with bounded, well-behaved gradients — exactly the regime where you should see a small effect when you have small inputs. Strip the tanh and normalize (the paper-text formula) and the loss blows up regardless of PRS; keep them and the lift comes from input volume, not magic.

The lift comes from PRS + biomarker volume, not from the architecture. Real PRS is upstream of a single data flow we’re now closer to than before (the CGI WGS ingest path landed in late May). Once a genotype upload completes, Track B becomes a clean A/B against Track A.

---

What’s next

In rough order of effort and likely interest:

1. Phase 13.1 — the snapshot fix. Two-line change in mods_kgia_profile_service.py + a regression test. Adds temporal sort to genomic-alterations and gene-expression queries so Phase 9 receives the most-recent value by default. Ships before Phase 9 starts.
2. Variant-aware fine-tune. Re-run the Maureen fine-tune against the Phase 8.5 KG. The criterion still keys on gene-level CURIEs, but the foundation now knows variants exist. We want to see whether the model’s score adjustments change when the graph it walks through has more granular structure.
3. Track B for Maureen. Real PRS, once a genotype upload runs through the CGI WGS ingest path → PGS Catalog → patient pickle.
4. Phase 1 writeup. The Spine post that closes out the PrimeKG-foundation + first-personalization milestone in the same shape as Phase 0.
5. The 1-hop audit. PDR2’s score adjustment only walks direct drug↔disease and drug↔gene edges. Still haven’t measured how much of Maureen’s clinically relevant drug surface is even reachable in 1 hop from her criterion anchors.

---

What’s coming on Richard’s Substack

In the queue, each now committed to docs/ and, at the time, about a half-evening of Richard’s voice work away from going live:

• A backfill orientation post: “What We’re Actually Building.” The intro post that should have come earlier in the archive. Slotting it now so a cold reader has somewhere to start.
• A Track C primer: “Where Does The Patient Live?” — the conceptual difference between PDR1 and PDR2, in one table.
• A second Track C primer: “The Personalized Loss Function” — the math, the λ-is-baked-in-per-fine-tune gotcha, and now the alt-loss A/B that proves it. (Will probably get an addendum referencing this month’s experiment.)
• A Track D engineering tour: “Training My First Neural Network” — the original GP-KG run, every gotcha, all six of them.

A pinned roadmap is going up alongside this post — a single living index that lists everything done and everything planned, organized by the five tracks the Substack actually runs in. Land on the roadmap cold and you can figure out where to start reading regardless of which question you arrived with.

---

The honest framing is that the model is not learning the patient so much as learning a bounded nudge. At small λ the released loss’s tanh sits in its near-linear region, so the personalization term acts as a small additive bias rather than a reshaping of the representation. That is a far weaker claim than “personalized model”, and it is the claim the numbers actually support.