stemedb/uat/consumer-health/glp1-semantic-decay.md

UAT: Semantic Decay (Knowledge Half-Life)

Date: YYYY-MM-DD Feature: Source-Aware Decay Status: [ ] PASS / [ ] FAIL / [ ] BLOCKED

Scenario

Medical knowledge decays at different rates based on source class:

• Regulatory (Tier 0): Never decays
• Clinical (Tier 1): 730-day half-life (2 years)
• Anecdotal (Tier 5): 30-day half-life

A 200-day-old Clinical study should have effective confidence ~0.45x original. A 200-day-old Reddit post should have effective confidence ~0.01x original (essentially expired).

Acceptance Criteria

Criterion	Expected	Met?
Regulatory never decays	conf unchanged at 200 days	[ ]
Clinical decay correct	~0.45x at 200 days	[ ]
Anecdotal decay correct	~0.01x at 200 days	[ ]
Decay applied at query	Not at write time	[ ]

Test Matrix

Step	Action	Expected	Actual	Status
1	Ingest Regulatory assertion	Hash returned		[ ]
2	Ingest Clinical assertion	Hash returned		[ ]
3	Ingest Anecdotal assertion	Hash returned		[ ]
4	Query with decay (200 days)	Decayed confidences		[ ]
5	Verify Regulatory unchanged	0.95		[ ]
6	Verify Clinical decayed	~0.43		[ ]
7	Verify Anecdotal expired	~0.01		[ ]

Decay Formula

Confidence decay follows exponential half-life:

effective_conf = original_conf * (0.5 ^ (age_days / half_life_days))

Source Class	Half-Life (days)	200-day decay factor
Regulatory	Infinite	1.0
Clinical	730	0.826
Observational	365	0.682
Expert	180	0.464
Community	90	0.214
Anecdotal	30	0.010

Setup Commands

# Start StemeDB
cargo run --bin stemedb-api &
sleep 2

Test Commands

Step 1: Ingest Regulatory Assertion (FDA)

# Use a timestamp 200 days ago
TIMESTAMP_200_DAYS_AGO=$(($(date +%s) - 17280000))

curl -X POST http://localhost:18180/v1/assertions \
  -H "Content-Type: application/json" \
  -d "{
    \"subject\": \"Semaglutide\",
    \"predicate\": \"approved_for\",
    \"object\": {\"Text\": \"Type2Diabetes\"},
    \"confidence\": 0.95,
    \"source_class\": \"Regulatory\",
    \"timestamp\": $TIMESTAMP_200_DAYS_AGO
  }"

Expected: Hash returned Actual: Status: [ ]

Step 2: Ingest Clinical Assertion

curl -X POST http://localhost:18180/v1/assertions \
  -H "Content-Type: application/json" \
  -d "{
    \"subject\": \"Semaglutide\",
    \"predicate\": \"hba1c_reduction_percent\",
    \"object\": {\"Number\": 1.5},
    \"confidence\": 0.95,
    \"source_class\": \"Clinical\",
    \"timestamp\": $TIMESTAMP_200_DAYS_AGO
  }"

Expected: Hash returned Actual: Status: [ ]

Step 3: Ingest Anecdotal Assertion (Reddit)

curl -X POST http://localhost:18180/v1/assertions \
  -H "Content-Type: application/json" \
  -d "{
    \"subject\": \"Semaglutide\",
    \"predicate\": \"user_experience\",
    \"object\": {\"Text\": \"Lost 20 lbs in 3 months\"},
    \"confidence\": 0.95,
    \"source_class\": \"Anecdotal\",
    \"timestamp\": $TIMESTAMP_200_DAYS_AGO
  }"

Expected: Hash returned Actual: Status: [ ]

Step 4: Query with Decay Lens

# Query all Semaglutide assertions with decay applied
curl "http://localhost:18180/v1/query?subject=Semaglutide&lens=decay"

Expected: Three assertions with decayed effective_confidence values Actual: Status: [ ]

Step 5: Verify Regulatory Unchanged

From Step 4 response, find the approved_for assertion.

Expected: effective_confidence = 0.95 (unchanged) Actual: Status: [ ]

Step 6: Verify Clinical Decayed

From Step 4 response, find the hba1c_reduction_percent assertion.

Expected: effective_confidence ~ 0.79 (0.95 * 0.826) Actual: Status: [ ]

Step 7: Verify Anecdotal Expired

From Step 4 response, find the user_experience assertion.

Expected: effective_confidence ~ 0.01 (essentially expired) Actual: Status: [ ]

Sign-Off Checklist

• All three source classes ingested
• Decay applied at query time (not write time)
• Regulatory sources never decay
• Clinical sources decay correctly
• Anecdotal sources expire quickly
• Decay factors match documented half-lives

Notes

The "73-day half-life" mentioned in the GLP-1 Living Review use case refers to the overall effective knowledge decay for mixed-source queries. Individual source classes have different rates.

---

Tester: Date: Result: