Axiom Cortex™ Method Map — v3.0.0 (Science Only)
Purpose: Give CTOs a single, accurate view of how Axiom Cortex™ works — the math and process only.
Not included: No public API, no vendor integration guides, no security docs. Axiom Cortex™ runs inside Nearshore IT Co-Pilot™.
1) Signal System (44 total) — what we measure
Grouped, audit-ready signals extracted from interview evidence and work samples:
- Reasoning & Problem Decomposition — goal factoring, constraint handling, trade-off narration, hypothesis sequencing.
- Solution Design & Architecture Thinking — interface contracts, boundary placement, failure modes, latency/cost awareness.
- Code Quality Behaviors — test intent, invariants, naming semantics, refactor triggers, complexity control.
- Debugging Strategy — fault localization steps, observability usage, minimal reproduction, rollback discipline.
- Data & API Literacy — schema inference, paging/limits, idempotency, consistency guarantees.
- Concurrency & Performance — race awareness, backpressure, caching safety, throughput/latency balancing.
- Security & Reliability Cues — input hardening, key/secret hygiene (conceptual), blast-radius thinking.
- Collaboration Signals — turn-taking, clarification prompts, disagreement framing, commit hygiene.
- Language-Fair Communication — idea clarity under L2 variance, ambiguity repair, evidence referencing.
- Meta-cognition — uncertainty surfacing, plan revision moments, verification loops.
The 44 signals live across these families; each has a rubric, evidence extractor, and stability test. (Full catalog appears as a separate Signal Catalog page in the next step.)
2) L2-Aware Fairness Layer — language-fairness calibration
- Proficiency-normalized scoring: ( C=\alpha C_{\text{sem}} + \beta C_{\text{form}} ) with (\beta \to 0) as L2 uncertainty rises.
- Cross-lingual semantic fidelity (FSD): Fréchet-style distance on multilingual embeddings.
- Optimal transport with code-switch mask: W2 with neutral costs for bilingual markers (Sinkhorn).
- DIF checks: Mantel–Haenszel / logistic DIF; biased items adjusted or dropped.
- Calibration: reliability diagrams, ECE; paraphrase/word-order stress tests.
3) Evidence Orchestration — semantic chunking & staged prompting
- Semantic chunking in RAG: split transcripts into meaning-bearing units aligned to rubrics and Blueprints of Ideal Answers.
- Staged / multi-step prompting: extract → evaluate → normalize → aggregate; adversarial paraphrase checks at each stage.
- Flag policy: any uncertainty or contradiction is flagged for expert review of flags before roll-up.
4) Measurement Models — from evidence to traits
- Non-parametric link: isotonic regression to keep monotonic relations (more evidence → higher trait).
- Monotone neural / lattice layers: partial-order constraints where applicable.
- Network psychometrics: Gaussian Graphical Model + stability selection → skill connectivity map.
5) Scoring & BARS mapping — explainable outcomes
- Trait estimates carry uncertainty bounds and stability notes.
- BARS = Behaviorally Anchored Rating Scales (ratings tied to observable behaviors), with role-specific anchors L1–L4.
- Report artifacts: per-question evidence, per-signal scores, rubric hits, and BARS summaries.
6) Decision Layer — constrained Bayesian utility
[ \max_{\mathcal{R}} \mathbb{E}[U(\mathcal{R}\mid \mathbf{e})]\quad \text{s.t. } \Pr[\text{Collab}<\tau_c]\le\epsilon_c,\; \text{DIF}k\le\delta,\; G\ge G{\min} ] Outputs include pass/hold/fail with justification, fairness/reliability gates, and confidence intervals.
7) Reliability & Monitoring — trust math
- Generalizability (G) across rater/question/time facets.
- Random Matrix Theory guard: remove non-replicating spikes outside Marchenko–Pastur support.
- Drift & stability: test–retest, bootstrap CIs, calibration-within-groups.
8) Protocol — how we evaluate (operational)
- Data & splits: anonymized interviews; stratified by role & locale; nested CV; seeded folds.
- Metrics: AUC/PR, Kendall’s ( \tau ), Brier score, ECE/MCE; fairness gaps (parity, equalized odds), DIF counts.
- Ablations: remove L2 layer / skill-graph / non-parametric link and observe degradations.
- Red-team: adversarial paraphrase, negation/hedging, verbosity/terseness controls; out-of-domain → “insufficient evidence”.
9) Evidence Locker — audit trail
Transcripts, embeddings, partial-corr matrices, calibration plots, gate statuses, and a provenance manifest (versions, seeds, hashes). Rollbacks available by bundle hash.
10) Where to read the math
- Scientific Foundations (full math/audit): /docs/axiom-cortex/scientific-foundations/
- Publications & Research: /docs/publications/
SSRN: 5165433 • 5253470 • 5188490
Platform context (no API exposed)
Axiom Cortex™ powers the Nearshore IT Co-Pilot™ — one platform to hire, equip, secure, and pay LATAM engineers under one SLA.
Authority links: