Health KB LLM Consistency Audit
Generated 2026-04-29 from topics/*.md after refactoring thirty-day-experiment.md. This report uses LLM judgment over the topic corpus to compare role fit, level of detail, evidence/action balance, writing style, cross-topic overlap, and what should be compressed, expanded, split, or left alone.
37
topic files audited
1
high-priority item fixed
4
remaining high-priority fixes
5
medium structural fixes
4
topics too shallow for their role
Main judgment: the 30-day experiment article is now much closer to the intended operational-protocol role. The remaining highest-value consistency work is not global shortening; it is assigning clear ownership to router, anchor, satellite, protocol, and ledger files.
Recently Fixed
| Status | Topic | What changed | Next state |
|---|---|---|---|
| Fixed / monitor | thirty-day-experiment.md | Refactored from mixed protocol/app-spec/lab-plan into a cleaner clinical protocol. Removed app implementation details, compressed stool-photo and abdomen-measurement guidance, made decision questions explicit, and kept the live counter because it is useful in the report view. | Leave it stable unless the experiment status changes. After day 30, convert from active protocol into results/interpretation summary or archive the protocol detail. |
Executive Findings
- Best current pattern: current signal → decision pathway → what changes management → action thresholds → compact references.
- Worst remaining pattern: exploratory review prose inside an action KB article, especially where mechanism depth exceeds decision value.
- Most important structural fix: define canonical owner files per decision. Repeated claims should live once, with satellites linking to that owner.
- Style consistency issue: the KB mixes clinical notes, live ledgers, literature reviews, and safety action plans. That is fine only when the article role is obvious.
- Research policy: deep/wide research should add only practice-changing deltas: new thresholds, diagnostic pathways, local access changes, safety boundaries, or therapies.
Target House Style by Article Type
| Article type | Good shape | What to avoid |
|---|---|---|
| Clinical anchor | 900–1,300 words. Summary, current signal, decision map, action thresholds, brief evidence/context, references. | Pipeline tracking, local logistics lists, repeated background physiology, every related side question. |
| Satellite workup/treatment | 500–1,100 words. Narrow question, exact triggers, test/treatment sequence, what result changes action. | Re-explaining the parent condition or duplicating the global profile narrative. |
| Live ledger/protocol | Tables/checklists are fine. Freshness, order status, trigger conditions, and execution clarity matter more than PMIDs. | Becoming a permanent clinical essay; app implementation details in the clinical body. |
| Evidence exploration | Evidence tiers and cautious implications. Keep only if it materially affects monitoring, testing, or interpretation. | Long mechanism review with low action yield. |
| Safety/action page | Short, calm, explicit: red flags, what to avoid, who to contact, what to test next. | Theatrical language, overconfident instructions, or burying urgent steps under rationale. |
Priority Topic Findings
| Priority | Topic | LLM consistency finding | Target state |
|---|---|---|---|
| High | bloating-vs-bleeding-risk.md | Best remaining example of role-depth mismatch: the clinical distinction is important, but the article over-proves it with repeated mechanisms and neighboring-topic evidence. It should be a router, not a GI review. | Compress into symptom branch vs bleeding branch. Own the boundary rule: bloating/SIBO does not explain positive occult blood/RBC. Link out for SIBO, diverticular disease, calprotectin, stool-blood escalation. |
| High | eczema-diverticular-connection.md | Evidence density is much higher than action yield. It reads like exploratory literature synthesis, while the practical conclusion is cautious and modest. | Convert to evidence tiers: established, plausible, weak/speculative, not supported. Keep strongest sources and the action implications; remove broad immunology review. |
| High | lp-a.md | Anchor article is carrying satellite jobs: baseline Lp(a) interpretation, imaging, aspirin, PCSK9, emerging therapies, smoking, family screening, prevention strategy. This creates duplication and drift risk. | Make it the stable anchor. Delegate pipeline to therapy watchlist, aspirin to antithrombotic strategy, imaging to CVD/imaging pages, and prevention sequence to CVD burden router. |
| High | thrombocytosis-lpa-thrombosis.md | Strong workup material, but the title/frame implies combined thrombotic-risk modeling while the useful core is platelet workup. Style is more synthesis-heavy than decision-heavy. | Reframe around platelet workup pathway first; keep Lp(a)/smoking/inflammation as risk modifiers. Preserve JAK2/CALR/MPL/smear triggers. |
| Medium | prevention-status-cvd-burden.md | Useful router, but overlaps with Lp(a), CCTA/LDNCP, CT screening and aspirin pages. Breadth is justified; repetition is not. | Make this the CVD decision map: primary vs secondary prevention, LDL/ApoB targets, plaque-detection sequence, aspirin boundary. Link rather than restate modality detail. |
| Medium | ct-scan-screening.md | Good practical imaging access article, but tends to become an imaging encyclopedia. | Keep local/logistical sequence and modality choice. Push advanced plaque interpretation to LDNCP article. |
| Medium | diverticular-disease.md | Good GI anchor, but it can absorb every gut update unless constrained. | Use as stable hub: disease baseline, recurrence prevention, core medication/diet rules. Push bleeding escalation, SIBO symptoms, alcohol recurrence, and calprotectin details to satellites. |
| Medium | planned-blood-tests.md | Highly actionable, but not a reusable clinical article. It is a live ledger and should be judged by freshness and execution clarity, not evidence density. | Label/shape as visit ledger: ordered, pending, trigger-based round 2, interpretation owner links. Avoid duplicating rationale from each clinical topic. |
| Medium | phnom-penh-medical-access.md | High practical value, but under-structured for the role. It should answer service queries faster. | Reformat as service matrix: service, confirmed local option, uncertainty, exact question to ask, fallback. |
Language and Style Consistency
| Style mode | Where it appears | Consistency action |
|---|---|---|
| House style | Decision-first, neutral clinical language, compact evidence, explicit action thresholds. | Use Summary / Current signal / Decision pathway / Action / References. Personalize via data, not chatty direct address. |
| Legacy review style | Long mechanism sections, many PMIDs, broad background, “evidence map”/“trace” labels. | Use only when the topic is genuinely an evidence review; otherwise collapse into evidence tiers. |
| Ledger/protocol style | Tables, checklists, dates, round logic, app/logging detail. Low PMIDs can be fine. | Keep if the file’s role is operational. Make that role explicit so it is not compared to clinical articles. |
| Safety/action style | Short, imperative, red flags, escalation thresholds. | Good for recurrence plans and medication avoidance. Avoid theatrical emphasis; use calm urgency. |
Specific language cleanup: standardize headings toward
Summary, Current signal, Decision pathway, Action, Evidence/context, and References. Avoid reader-facing process labels such as SearchPlan, Compact Evidence Map, Practical meaning, Verdict, and emphatic all-caps instructions unless they are true safety warnings.Cross-Topic Architecture Map
| Cluster | Canonical ownership | Drift risk |
|---|---|---|
| Cardiovascular / Lp(a) | lp-a.md should own fixed genetic risk; prevention-status-cvd-burden.md should own management routing; ldncp-advanced-imaging.md and ct-scan-screening.md should own imaging details; antithrombotic-strategy.md should own aspirin/bleeding boundary. | Current risk: PCSK9/aspirin/imaging rationale repeated with slightly different emphasis. Harmonize “best bridge” vs “not mandatory until plaque/overall risk justifies it.” |
| GI / bleeding / SIBO | diverticular-disease.md as hub; occult-stool-blood-workup.md as escalation tree; bloating-vs-bleeding-risk.md as boundary router; sibo-mmc.md as symptom differential; rebleeding-risk-alcohol.md as current recurrence-risk context. | Current risk: multiple pages prove that bloating does not explain stool blood. Keep that rule in the router and refer out. |
| Iron / inflammation / platelets | ferritin-iron-workup.md owns diagnosis; oral-iron-repletion.md owns treatment; thrombocytosis-lpa-thrombosis.md owns platelet workup; inflammatory-thrombotic-axis.md interprets the combined pattern. | Current risk: April lab story repeated across files. Use one compact current-signal block per topic. |
| Supplements / probiotics | recommended-supplement-adjustments.md can become the stack hub; omega-3/vitamin-d-k2/probiotics as compact notes; clostridium-butyricum.md as targeted evidence page. | Current risk: fragmented short pages are readable individually but hard to review as one stack. |
| Local logistics | phnom-penh-medical-access.md should own stable service availability; clinical pages should include logistics only when it changes the decision. | Current risk: local access facts drift across SIBO, imaging, Repatha, Biomed, colonoscopy content. |
Topics That Are Too Shallow for Their Role
| Topic | Why it is shallow | Target expansion |
|---|---|---|
antithrombotic-strategy.md | Too short for its importance: it should be the canonical aspirin/antithrombotic boundary in the presence of high Lp(a) plus GI bleeding uncertainty. | Expand modestly to 700–900 words; include what evidence would reopen aspirin, who decides, and what GI facts must be clarified. |
phnom-penh-medical-access.md | Practical value is high but query structure is weak. | Turn into service/access matrix rather than prose facility list. |
probiotics.md | Readable but decision role is unclear beside CBM588 and SIBO pages. | Add current-regimen table or merge into supplement hub, keeping CBM588 as focused evidence page. |
vitamin-d-k2.md | Short is fine, but if it remains “important” it needs a clearer monitoring link to calcium/PTH. | Either downgrade to compact supplement note or add explicit calcium/PTH boundary. |
Model Topics to Imitate
| Topic | Why it works |
|---|---|
elevated-iga-workup.md | Excellent abnormal result → differential buckets → trigger tests structure. |
occult-stool-blood-workup.md | Canonical escalation logic; clear ownership of stool-blood decision tree. |
b12-functional-deficiency.md | Good balance of mechanism, lab interpretation, supplement interaction, and next step. |
oral-iron-repletion.md | Treatment satellite with clear dose/timing/monitoring purpose. |
blood-pressure-profile.md | Protocol style is justified and readable; measurement acquisition is the point. |
recurrence-action-plan.md | Short enough to use during a problem; clear red flags and next actions. |
ldncp-advanced-imaging.md | Recently compacted specialty satellite; should stay linked, not broadened. |
thirty-day-experiment.md | After refactor, a good example of an operational protocol kept separate from lab logistics and app implementation. |
Full Topic Triage Matrix
| File | KB / lines | Evidence signals | Role fit | Detail level | Style | Recommendation |
|---|---|---|---|---|---|---|
recommended-supplement-adjustments.md | merged hub | PMIDs: mixed; action terms: router | Stack hub after merge | Right-sized after psyllium expansion | Practical | Keep as single supplement hub; standalone add-on page retired |
antithrombotic-strategy.md | 3.8 / 57 | PMIDs: 0; action terms: 4; meta/process hits: 2 | Too shallow for canonical boundary | Under-detailed | Neutral but too thin | Expand modestly |
apple-health-signals.md | 12.2 / 219 | PMIDs: 11; action terms: 14; meta/process hits: 1 | Appropriate data-signal article | Long but justified | Data-heavy; acceptable | Keep; prune passive speculation over time |
b12-functional-deficiency.md | 6.5 / 118 | PMIDs: 0; action terms: 29; meta/process hits: 0 | Good workup satellite | Right-sized | Decision-first | Keep as model |
bloating-vs-bleeding-risk.md | 15.6 / 215 | PMIDs: 37; action terms: 20; meta/process hits: 1 | Too deep for boundary role | Over-detailed | Useful but repetitive | Compress as router |
blood-pressure-profile.md | 8.0 / 127 | PMIDs: 1; action terms: 29; meta/process hits: 2 | Good acquisition protocol | Right-sized | Operational | Keep |
calcium-parathyroid-vitamin-d.md | 8.2 / 113 | PMIDs: 0; action terms: 14; meta/process hits: 0 | Good metabolic anchor | Mostly right-sized | Plain clinical | Keep; minor links if revisited |
clostridium-butyricum.md | 6.9 / 124 | PMIDs: 11; action terms: 4; meta/process hits: 1 | Good probiotic satellite | Right-sized/high evidence | Evidence-heavy but focused | Keep as targeted evidence page |
ct-scan-screening.md | 11.2 / 172 | PMIDs: 0; action terms: 68; meta/process hits: 1 | Broad imaging menu | Slightly over-detailed | Practical but encyclopedic risk | Compress around sequence |
diverticular-disease.md | 10.5 / 140 | PMIDs: 10; action terms: 30; meta/process hits: 0 | Good anchor but broad | Slightly over-detailed | Mostly consistent | Keep as GI hub; trim satellite detail |
eczema-diverticular-connection.md | 14.9 / 189 | PMIDs: 18; action terms: 6; meta/process hits: 0 | Evidence outweighs action | Over-detailed | Review-like | Compress hard with evidence tiers |
elevated-iga-workup.md | 9.4 / 111 | PMIDs: 1; action terms: 34; meta/process hits: 2 | Excellent urgent workup | Right-sized | Decision-first | Keep as model |
ferritin-iron-workup.md | 6.6 / 86 | PMIDs: 6; action terms: 23; meta/process hits: 1 | Good diagnostic anchor | Right-sized | Clinical | Keep; avoid overlap with oral iron |
hdl.md | 9.1 / 126 | PMIDs: 6; action terms: 8; meta/process hits: 0 | Appropriate monitor topic | Right-sized | Plain | Keep; maybe compress if CVD router expands |
inflammatory-thrombotic-axis.md | 9.8 / 138 | PMIDs: 0; action terms: 6; meta/process hits: 0 | Useful but less actionable | Moderate | Mechanism-heavy; watch overlap | Keep as interpretation bridge |
ldncp-advanced-imaging.md | 8.9 / 104 | PMIDs: 17; action terms: 67; meta/process hits: 1 | Good specialty satellite | Right-sized after compaction | Consistent | Keep |
lp-a.md | 12.6 / 168 | PMIDs: 6; action terms: 9; meta/process hits: 2 | Anchor carrying satellite jobs | Over-detailed | Mixed roadmap/watchlist/anchor | Split/trim into stable anchor |
lpa-therapy-watchlist.md | 7.8 / 88 | PMIDs: 0; action terms: 17; meta/process hits: 2 | Correct live watchlist | Right-sized | Watchlist style appropriate | Keep; absorb pipeline detail from Lp(a) |
medication-avoid-list.md | 6.4 / 102 | PMIDs: 0; action terms: 20; meta/process hits: 0 | Good safety topic | Right-sized | Action style | Keep |
occult-stool-blood-workup.md | 8.1 / 128 | PMIDs: 7; action terms: 48; meta/process hits: 1 | Canonical escalation tree | Right-sized | Decision-first | Keep as model |
omega-3.md | 2.5 / 52 | PMIDs: 7; action terms: 1; meta/process hits: 0 | Supplement note | Right-sized | Plain | Keep compact |
oral-iron-repletion.md | 6.8 / 107 | PMIDs: 5; action terms: 21; meta/process hits: 2 | Good treatment satellite | Right-sized | Decision-first | Keep |
phnom-penh-medical-access.md | 6.9 / 122 | PMIDs: 0; action terms: 19; meta/process hits: 0 | Useful but under-structured | Under-detailed for access role | List-like | Reformat service matrix |
planned-blood-tests.md | 10.1 / 109 | PMIDs: 0; action terms: 34; meta/process hits: 0 | Operational ledger, not clinical article | Detailed but role-appropriate | Table/logistics | Keep as live ledger; archive later |
prevention-status-cvd-burden.md | 9.8 / 102 | PMIDs: 4; action terms: 57; meta/process hits: 2 | Good router but overlaps | Slightly over-detailed | Framework style | Compress as CVD router |
probiotics.md | 2.9 / 70 | PMIDs: 5; action terms: 1; meta/process hits: 0 | Slightly shallow beside CBM588/SIBO | Under-detailed for decisions | Plain but generic | Clarify or merge into supplement hub |
psa-kinetics.md | 4.4 / 74 | PMIDs: 0; action terms: 18; meta/process hits: 0 | Appropriate monitor topic | Right-sized | Plain | Keep |
rebleeding-risk-alcohol.md | 7.2 / 156 | PMIDs: 3; action terms: 13; meta/process hits: 0 | Status/action synthesis | Right-sized | Action-oriented | Keep; avoid hub duplication |
recurrence-action-plan.md | 5.9 / 83 | PMIDs: 2; action terms: 27; meta/process hits: 2 | Good urgent action plan | Right-sized | Safety/action | Keep |
research-queue.md | 10.0 / 102 | PMIDs: 0; action terms: 36; meta/process hits: 0 | Planning file | Role-specific | Workflow | Keep outside clinical standards |
sibo-mmc.md | 10.7 / 122 | PMIDs: 13; action terms: 35; meta/process hits: 1 | Near target after compaction | Right-sized | Logistics/action | Keep; minor duplication trim |
smoking-alcohol-relapse-risk.md | 6.9 / 117 | PMIDs: 10; action terms: 5; meta/process hits: 1 | Good behavior-risk synthesis | Right-sized | Clinical/action | Keep |
thirty-day-experiment.md | 14.3 / 205 | PMIDs: 0; action terms: 53; meta/process hits: 0 | Operational protocol now separated from app/lab implementation | Long but role-appropriate | Cleaner protocol style | Fixed; monitor after day 30 |
thrombocytosis-lpa-thrombosis.md | 13.0 / 134 | PMIDs: 6; action terms: 41; meta/process hits: 1 | Good workup, title mismatch | Moderately over-detailed | Synthesis-heavy | Clarify role and compress |
tsh-thyroid-trend.md | 5.0 / 92 | PMIDs: 0; action terms: 11; meta/process hits: 0 | Appropriate monitor topic | Right-sized | Plain | Keep |
vitamin-d-k2.md | 2.1 / 48 | PMIDs: 2; action terms: 1; meta/process hits: 0 | Maintenance note | Possibly shallow for priority | Plain | Keep compact or link calcium/PTH |
Recommended Rewrite Sequence
- bloating-vs-bleeding-risk.md — compress into router; it will reduce GI duplication immediately.
- lp-a.md — turn into stable anchor and push live/specialized detail to satellites. Highest drift-risk reduction.
- eczema-diverticular-connection.md — compress evidence review into evidence tiers. Biggest evidence/action imbalance.
- thrombocytosis-lpa-thrombosis.md — clarify whether it is a platelet-workup topic or combined-risk synthesis.
- phnom-penh-medical-access.md — reformat into service matrix. High practical value per token.
Audit Method
Inputs were all source markdown files under topics/*.md. Metrics such as size, line count, heading count, table density, PMID count, action-term density, and process/meta-label hits were used only as prompts. The final classification is LLM judgment over article role, not a formula. Three passes were used: style/language consistency, depth/actionability consistency, and cross-topic ownership/duplication.
This report is a planning artifact. It should not be folded into the main health report unless explicitly requested.