← TALENT2ROLE

Trust: AI, security and assessment integrity

Exactly what AI does in Talent2Role and what it cannot do, where your data goes, how the platform is secured, how sittings are kept honest, how to contest a result, and what we will never claim. Last updated 5 September 2026.

AI & model governance Security posture Assessment integrity Dispute a mapping Employment use Claims policy

1. AI and model governance

Transparency notice. Talent2Role uses an AI model (Anthropic Claude, via the API) for specific, named tasks. Scoring is not one of them. Every readiness score, coverage figure, role ranking and credential is produced by a deterministic engine whose method is published at /method.html; identical inputs always produce identical outputs.

What AI does

Task	What the model receives	What happens to its output
Résumé analysis (Capture)	The résumé / profile text you import	Structured fields (skills, knowledge, capabilities, roles held) are shown for your review and become part of your evidence only when you apply them.
Statement mapping (Map)	Your evidence text and a list of "near-miss" statements the lexical matcher could not decide	The model proposes matches. Each proposal must quote your evidence verbatim (checked), meet a confidence threshold (per statement tier), and either fire a rule that reaches the statement or lexically hit it; anything else is rejected or held at half weight. Accepted proposals add a bounded credit (0.10 × weight) to the engine's own score — they cannot flip a statement that has no lexical basis. The engine, not the model, makes the final match decision.
Writing tasks (gap summaries, interview prep, LinkedIn and résumé copy, Profile Compare, Decision Studio, CPE knowledge-check questions)	The relevant profile fields, engine scores and any text you paste for that task	Drafts for you to edit. They never change a score.

Model, determinism and versions

Model. Mapping runs on a pinned model string (currently claude-haiku-4-5-20251001) at temperature 0 with candidates sorted by statement ID in fixed batches; writing tasks run on the model shown in the app's AI status. The model name is recorded on every mapping and credential.
Prompt version. The mapping prompt is versioned (map-v2) and stamped on results; changing it bumps the version.
Cache. Mapping results are stored under a key derived from the evidence, the candidate statements, the prompt version, the rule-set version and the ontology version. The same inputs return the stored result — no new model call, no drift.
Consensus. Before an alignment certificate is issued, the mapping is re-run three times independently and only statements agreed by at least two runs are kept.
Explainability. Every matched statement shows "why": the lexical hit, the rule that reached it, the quoted evidence and the AI credit. You can dispute any match (see §4).

Data sent, retention, training

Only the text needed for the task is sent, from our server; your API credentials are never involved and the API key never reaches the browser.
Anthropic processes API requests under its commercial API terms, which exclude the use of API inputs and outputs to train models. Results are stored on your profile so they are not regenerated needlessly; deleting your profile deletes them.
No profile data is used to train any model, by us or by anyone else. Aggregate, de-identified statistics may be used to validate the method (see the technical note).
Free accounts and Professional accounts have monthly AI-credit allowances; admins can top up. Credit usage is logged per task, not per content.

2. Security posture

Listed on the CSA STAR Registry — Level 1 self-assessment. Quantum Task AI's entry for Talent2Role is public at cloudsecurityalliance.org/star/registry/quantumtask-ai, with the full CAIQ v4.0.3 answers. Level 1 is a published self-assessment reviewed by CSA for completeness, not a third-party certification; it is renewed annually and re-submitted whenever practices change.

Talent2Role completed the Cloud Security Alliance Consensus Assessments Initiative Questionnaire (CAIQ) v4.0.3 in September 2026: 258 of 262 applicable questions answered "Yes". The three "No" answers are listed below rather than hidden, because they are the honest state of a young product. Security contact: security@quantumtaskai.com.

Area	Position
Hosting & architecture	Static site plus serverless functions on Netlify (AWS infrastructure); operating-system, network and virtualisation security are inherited from the provider. Environments are separated; production deploys go through a git-based pipeline.
Encryption	TLS 1.2+ for all traffic (a wildcard RSA certificate is used because some regions' clients cannot negotiate ECDSA-only chains); encryption at rest is provided by the managed data stores. Credential signing uses ECDSA P-256 keys; the private key is held in the environment or a protected store, never in the browser.
Identity & access	Netlify Identity with per-user identifiers and Google sign-in; server functions validate every request against the identity service. Administrative access is limited to named Quantum Task AI personnel with MFA.
Data	Profiles live in per-user server storage with a browser cache; data is classified (account, assessment, telemetry) and minimised. No card data is handled by Talent2Role — payments run on Ziina. See the privacy policy for retention and your rights.
Logging & monitoring	Function and access logs via the platform; administrative actions and credential events are recorded in an activity store; server errors are captured for diagnosis.
Vulnerability management	Automated dependency scanning in CI; continuous deployment of fixes; platform patching inherited from the provider.
Incident response	A documented lightweight plan (detect, triage, contain, notify, review). Report security issues to security@quantumtaskai.com.

Not yet done (CAIQ "No" answers): independent third-party audit and assurance assessment (A&A-02, A&A-03) and external penetration testing (TVM-06). Both are planned; compensating controls are continuous dependency scanning, provider-inherited platform security and the serverless attack surface. We will update this page when either is completed.

3. Assessment integrity policy

The knowledge check is an online, unproctored sitting. We say so on every credential. What we do to keep it honest, and what we deliberately do not do:

What we do

Honour statement before the first credential: the evidence is your own, you are 18 or older, you sit the check yourself.
Randomised, stratified sampling from a larger item bank per role; options are shuffled; no two sittings are alike; items in bank order are never shown.
Timing: one item at a time, 75 seconds each, no going back; unanswered items count as wrong.
Cooldown: 14 days between attempts and 3 attempts per 90 days, enforced on the server.
Signals: tab switches away from the sitting and paste attempts are counted and recorded on the result. Three or more flags mark the sitting provisional; they do not fail it on their own.
Consistency checks on evidence (thin evidence with a high score, advanced-tier claims at an early stage, senior certifications against junior experience, evidence and knowledge results that disagree) mark a credential provisional on its face and in the registry.
Revocation for false evidence or a sitting taken by someone else; see the issuing policy.
Item statistics: as sittings accumulate, item difficulty and discrimination will be computed from the sitting records; items that stop discriminating will be retired and the bank version bumped. Every credential records the bank version it was sat against.

What we do not do

No webcam, screen recording, keystroke logging or browser lockdown. No biometric identity checks.
No collection of protected attributes. No use of sitting data for anything other than scoring, integrity and method validation.
No claim that a pass proves professional competence — it proves knowledge of the role's competency statements under time pressure.

4. Dispute a mapping or a result

Every matched statement in the Gap Analyzer and the mapping panel has a why button showing exactly how it was matched. If you think a match is wrong, click Dispute: the statement immediately stops counting toward any score or credential, the dispute is recorded, and you can withdraw it at any time. If you think a knowledge-check result or a revocation is wrong, e-mail info@quantumtaskai.com from your account address; a person reviews it within 10 working days and replies with reasons. Nothing in Talent2Role is decided by an algorithm without a route to a person.

5. Use in employment decisions

Talent2Role is a career-development and workforce-planning tool. It is not designed or validated for employment selection. Scores, credentials and reports must not be used as the sole basis for any hiring, promotion, pay or termination decision, and organisations must not use them to screen candidates without a written agreement that assigns responsibility for bias testing and legal compliance in the relevant jurisdiction (for example New York City Local Law 144 and the EU AI Act's rules for employment-related systems). Our terms say the same.

6. Claims policy — what we will never say

Our public claims follow a written policy that is enforced automatically at build time. In short:

We say aligned to, mapped to or based on the public-domain NICE Framework. We never say NIST-certified, NICE-certified, accredited, endorsed or recognised by NIST or any government. See framework attribution.
We say credential, certificate (of mapped skills, of skills alignment, of completion) and knowledge-verified. We never confer a title ("Certified Threat Hunter") and we never call the product a certification, licence or accreditation.
We say what a credential rests on — evidence-aligned or knowledge-verified — on its face, every time.
We describe AI as assistive and say where it is used; we never describe scores as AI-generated, because they are not.