Engineering Methodology

Development Philosophy

Condrox AI was built on four principles that guided every decision from the first line of code to the final deployment manifest.

• Determinism over probability. Every response is the product of a traceable pipeline. The same input produces the same output. There is no sampling, no temperature, no hidden state that drifts between runs. This property is verified by the test suite on every change.
• Transparency over fluency. The system shows its work. Every stage of the 18-step pipeline is observable. Every decision is logged. When the system produces a response, you can trace exactly which stages contributed to it and what knowledge was retrieved.
• Verification over assertion. Claims about the system are not promises. They are measurements. The Evidence Engine regenerates academic-grade reports from the live system on every run, covering conversation quality, reasoning, agent usage, and full-system integration.
• Architecture over scale. The system does not require more parameters to solve harder problems. It requires better structure. 21,917 knowledge items and 185 word identities are sufficient when the pipeline is correct. The 18-stage architecture does the heavy lifting, not the data volume.

Testing Standards

The project maintains 63 mandatory testing rules (R-TS-01 through R-TS-63) documented in Report/rules/testing_standards.md. These rules are derived from academic, corporate, and laboratory standards.

12 acceptance criteria (A1 through A12) must all pass for any test to be considered valid:

Current test suite: 938 tests across 43 files.

All passing. Zero failures. ~210 seconds execution time.

Test-First Development

New features start with failing tests that define expected behavior. The implementation is then written to make those tests pass. This ensures that every feature has a test before it exists, not as an afterthought.

The Evidence Engine followed this pattern across all 7 phases. Each phase started with test scenarios that defined what the evidence should look like, then the evaluator was built to produce that evidence.

Evidence-Driven Improvement

Every change to the system is measured by the Evidence Engine before and after. The engine runs 11 scenarios across 7 evidence dimensions:

1. System verification (boot, API health, scenario runs)
2. Conversational evidence (empathy, relevance, consistency, multi-turn stability)
3. Reasoning evidence (step-by-step structure, logical coherence, planning)
4. Agent and pipeline evidence (agent usage, RAG, memory, personality)
5. Full-system integration (end-to-end conversation, reasoning, agents, RAG, API)
6. Academic reports (Lab, Academic, Peer Review editions)
7. Evidence portal (score history, report library)

If a change reduces the evidence score, it is not merged. The current score is 93.87% (WARN), with full-system release status READY and a complete-session score of 92.94%. The remaining WARN is driven by StepByStepStructure and ExplanationQuality on the code_analysis reasoning task.

Code Standards

• Python 3.11+ with type hints on all public functions
• Dataclasses use @dataclass(frozen=True, slots=True) for value objects
• BrainState is immutable; updates use with_field() method
• O() complexity documented in all docstrings
• No Hugging Face datasets dependency; official sources used directly
• All new modules must be wired into BrainController and tested
• All new code must have real algorithms, no placeholders

Report Standards

All professional reports follow rules in Report/rules/:

• report_writing_rules.md: 40 mandatory writing rules (R-WR-01 to R-WR-40)
• testing_standards.md: 63 mandatory testing rules (R-TS-01 to R-TS-63)
• academic_standards.md: Academic and peer-review standards
• corporate_standards.md: Corporate, ISO, CMMI standards
• lab_standards.md: Lab, NIST, FDA standards

Every report has a unique ID: RPT-<domain>-<seq>-<year>. Domains: SYS, PERF, CI, AGT, CREA, TEST, SEC, REL, RES.

Phase-Based Development

Work is organized into phases with clear entry and exit criteria. Each phase has:

• A defined scope (what will be built)
• Entry criteria (what must be true before starting)
• Exit criteria (what must be true before declaring complete)
• Test requirements (how many new tests, what they cover)
• Documentation requirements (what must be written)

16 phases have been completed. Each is documented on the Development Phases page.

Reproducibility

Every claim about the system can be independently verified:

• python -m pytest tests/ -v: Runs all 938 tests
• python scripts/generate_evidence.py: Regenerates all evidence reports
• condrox verify: Runs system verification probes
• condrox evidence-portal: Opens the evidence portal

The evidence reports include methodology sections, references, signatures, and reproducibility instructions. Any expert can run the same commands and get the same results.