SR&ED Submission Package · FY2025 · CHAI — InfiniTEA Labs Inc.

Scientific Research &
Experimental Development

CHAI (Compliance Helper for AI) is a BC-incorporated CCPC developing novel AI-assisted compliance infrastructure for the community care sector. This package documents qualifying SR&ED work conducted April–December 2025.

Entity InfiniTEA Labs Inc.
Fiscal Year April – December 2025
R&D Start June 17, 2025
Standards BC RCR + CARF 2023/2017
Status READY FOR SUBMISSION
15
Hypotheses
Investigated
22
Experiments
Conducted
9
Technical
Advancements
0.92
Avg Hardy
Confidence
148
Git Commits
Audited
CRA Form T776 — Part 3

Technical Advancements

Findings that achieved Hardy PRESTIGE state (≥0.90 confidence via 7-chamber Byzantine consensus). Each represents a resolved technological uncertainty with reproducible git-versioned evidence.

FIND002 · H003
PRESTIGE
245 regulatory rules extracted to executable YAML — zero-LLM compliance routing confirmed across BC RCR and CARF 2023
Hardy
0.91
H003: Deterministic rule routing for regulatory compliance without LLM
EXP004 EXP005 8adac02 7e7fd39 d03cb16
FIND004 · H005
PRESTIGE
Autonomy tiers (AUTONOMOUS / LOG / QUEUE / ALERT / BLOCK) — governance matrix verifies zero unilateral compliance-affecting actions in audit log
Hardy
0.93
H005: Autonomy tier architecture preventing unilateral LLM compliance actions
EXP008 EXP009 a527940 41b336b 6ca3da9
FIND005 · H006
PRESTIGE
Source authority ranking (OPERATOR > BINDER > PDF > DRIVE > EMAIL) — deterministic conflict resolution with 100% accuracy on non-ambiguous regulatory conflicts
Hardy
0.90
H006: Source authority reconciliation for conflicting regulatory documentation
EXP010 EXP011 e1ea113 41b336b
FIND007 · H008
PRESTIGE
Hybrid TF-IDF + FAISS + LLM matcher resolves BC care compliance vocabulary mismatch — finds "personal care assistance" when staff writes "ADL support". Single-method approaches insufficient.
Hardy
0.90
H008: Hybrid semantic matching resolves regulatory vocabulary gap
EXP014
FIND008 · H009
PRESTIGE
475 executable YAML rules extracted from 40,883-line CARF 2017 manual without domain expert annotation. Coverage gap analysis revealed 67–93% field gaps, requiring development of iterative extraction methodology. Extended to 818 rules across 4 governance domains (CARF, RCR, CLBC, Fraser Health).
Hardy
0.91
H009: Natural language regulatory standards extractable to executable YAML without domain expert annotation
EXP015
FIND009 · H010
PRESTIGE
Source Authority Matrix (OPERATOR:5 > BINDER:4 > PDF:3 > DRIVE:2 > EMAIL:1) resolves conflicts across CLBC, Fraser Health, CARF, RCR. Novel: dignity-of-risk exception to override logic is itself a regulatory requirement, implemented without manual intervention.
Hardy
0.90
H010: Weighted multi-authority conflict resolution with dignity-of-risk exceptions
EXP016
FIND010 · H011
PRESTIGE
FOIPPA-compliant dual-LLM routing: Claude API for non-PII reasoning, Ollama for unredacted resident records. chai/deidentifier.py as mandatory pre-processing gate. No existing healthcare AI framework addressed FOIPPA + AI capability trade-off.
Hardy
0.90
H011: FOIPPA-compliant dual-LLM routing for BC data sovereignty
EXP017
FIND011 · H012
PRESTIGE
QLoRA fine-tuning achieves mean quality 0.926 on care compliance reasoning using 92 manually verified examples on RTX 5060 Ti 16GB consumer GPU (8-minute training run). Prior automated approach failed — hallucinated 13 fictional file paths, requiring validate_training_data.py as systematic correction.
Hardy
0.93
H012: QLoRA fine-tuning achieves production-quality compliance reasoning on consumer GPU
EXP019
FIND012 · H013
PRESTIGE
NL2SQL pipeline converts care operations queries ("what is overdue?", "bowel care last week?") to SQL without LLM invocation at query time. 25+ intent patterns with fuzzy typo tolerance (difflib). Achieves 80/80 (100%) on care compliance query benchmark as of Dec 2025.
Hardy
0.92
H013: NL2SQL intent detection converts care queries to SQL without LLM at query time
EXP020 8cba097 e5fc0e0
FIND014 · H015
PRESTIGE
20-scenario real-world benchmark library constructed from Fraser Health and CLBC published sources. Establishes ground truth for compliance analyzer rule coverage testing. Baseline: 0% standard detection (December 2025). Benchmark enables ongoing measurement of rule coverage improvements against authentic care staff language.
Hardy
0.91
H015: Real-world scenario benchmark for compliance analyzer rule coverage
EXP022
CRA Form T776 — Part 2

Technological Uncertainties

Active investigations where the outcome cannot be determined through standard practice. Each represents genuine scientific or technological uncertainty requiring systematic study.

ACTIVE INVESTIGATION

Bradley-Terry peer ranking improves multi-agent consensus quality

Whether probabilistic pair-comparison models generalize to LLM output quality ranking in healthcare compliance contexts.

EXP002 EXP003
ACTIVE · PHASE 3 COMPLETE

NL2SQL parsing of unstructured caregiver voice/text notes into compliance events

Whether LLM-assisted pipeline achieves sufficient precision for audit trail use on raw caregiver language. Phase 4 (production validation) in progress.

EXP006 EXP007
ACTIVE · BC COMPLETE

Jurisdiction-portable rule externalization via YAML configuration

Whether compliance rules can be externalized so adding a new jurisdiction requires only configuration changes, not code changes. BC done; second province (ON or AB) pending.

EXP012 EXP013
ACTIVE INVESTIGATION

FOIPPA-compliant dual-LLM routing — phase validation in progress

Whether the de-identification gate + dual-routing architecture meets BC data sovereignty requirements under operational load. Architecture validated; production testing ongoing.

EXP017
ACTIVE INVESTIGATION

Canadian healthcare optimization benchmark from public data

Whether a Canadian-specific AI benchmark (30-day readmission, HSMR) is constructable from CIHI + ICES + PopulationDataBC with ICD-10-CA normalization. No published public benchmark currently exists for Canada.

EXP021
ACTIVE INVESTIGATION

NL2SQL caregiver notes — Phase 4 production validation

Whether structured compliance event extraction from voice/text notes achieves audit-ready precision under production conditions with diverse care staff writing styles.

EXP006 EXP007
Research Infrastructure

Systematic Investigation Methodology

CHAI's R&D uses two interlocked methodologies: Hardy confidence states for epistemological rigor, and the UEP 4-phase framework for systematic documentation — both designed to satisfy CRA's systematic investigation requirement.

Hardy 3-State → CRA Mapping

Hardy State Confidence CRA Category
PRESTIGE ≥ 0.90 Technical Advancement (T776 Part 3)
HYPOTHETICAL 0.65–0.89 Systematic Investigation (T776 Part 2)
LATENT < 0.65 Technological Uncertainty (T776 Part 1)

Source: agents/hardy_standalone.py:43–47
Validation: 7-chamber Byzantine consensus

UEP 4-Phase → CRA Form T776

  • 1
    Archaeological Discovery
    Prior art search, existing implementation inventory. Maps to T776: "Why not routine engineering?"
  • 2
    Implementation
    Hypothesis-driven development with contemporaneous git commits as evidence.
  • 3
    Validation
    Hardy consensus scoring. Benchmarks. Systematic measurement of advancement.
  • 4
    Complete / PRESTIGE
    Reproducible finding registered. Maps to T776 Part 3: Technical Advancement.
CRA Compliance

Form T776 Readiness Checklist

All four CRA elements are documented and evidence-backed. Package generated from git-versioned audit trail by TEAOS SR&ED infrastructure (built November 2025).

  • ✓
    Technological Uncertainty documented
    6 active investigations with formal LATENT/HYPOTHETICAL classification. Each includes "Why not routine engineering?" justification and prior art search evidence.
  • ✓
    Systematic Investigation recorded
    22 experiments documented using UEP 4-phase methodology. Git commits provide contemporaneous, tamper-proof records. Experiment registry JSON-backed with H→E→F lineage.
  • ✓
    Technical Advancement validated
    9 findings achieved PRESTIGE state (≥0.90 Hardy confidence). Includes zero-LLM compliance routing, autonomy tier governance, QLoRA consumer-GPU training, and FOIPPA-compliant dual-LLM routing.
  • ✓
    Supporting Evidence complete
    148 CHAI git commits audited (June 17 – December 2025). Includes pre-git artifacts (SMW Hackathon technical plan, June 17 2025). Fraser Health + CLBC published source citations.
  • ✓
    CCPC eligibility confirmed
    InfiniTEA Labs Inc. — BC-incorporated CCPC. Development began April 2025. FY2025 = April–December 2025. Qualifies for 35% refundable credit on up to $3M qualifying expenditure.
About this package

This package was generated by TEAOS SR&ED infrastructure built in November 2025 as a 12-agent swarm (16/16 tests passing). The generator reads directly from the experiment registry — a JSON-backed audit trail updated via git post-commit hooks. Hardy confidence scores are computed by a 7-chamber Byzantine consensus algorithm, not manually assigned.

The git commit hashes cited in each finding are real commits in the CHAI repository. They can be independently verified via git show <hash>.
Generated: 2026-04-03 · Generator: .git-research/sred_generator.py
Registry: .git-research/experiment_registry.json
Hardy: agents/hardy_standalone.py · SOP: sops/SRED_RESEARCH_SOP.md

CHAI SR&ED FY2025 — InfiniTEA Labs Inc. · BC-incorporated CCPC · aingco@gmail.com

Package generated by TEAOS SR&ED Infrastructure · Hardy 7-chamber Byzantine consensus · Git audit trail