CareerOS 360 · The operating standard
What the agents must prove before a file gets out.
An AI that writes fluently is easy to build. An AI that refuses to invent, checks every number and admits what it could not do is not. This is the standard both CareerOS agents run under, in plain words.
- One record
- Five evidence classes
- Fourteen checks
One source of truth
Everything is built from one career record.
Before any file is written, the agent builds a set of registers. Every CV line, slide and interview answer is derived from them, which is why your files never contradict each other.
Source inventory
Every document and public source used, listed.
Canonical career record
The single agreed version of your career.
Evidence ledger
Each claim, its class and where it comes from.
Discrepancy register
Where your CV and LinkedIn disagree, recorded.
Voice profile
How you write: formality, rhythm, vocabulary, locale.
Metric register
Every material number, with at least nine fields.
Audience & redaction matrix
Who may see each claim.
Revision trail
Corrections update the record first, then every affected file.
The evidence ladder
Five classes. No sixth called “trust me”.
Each claim also records its origin, consistency, ownership, confidence, privacy and shareability.
Backed by a document or record you supplied that settles the point.
StrongestConsistent across independent sources, such as your CV, LinkedIn and public professional records.
StrongStated by you but not independently supported yet.
Needs supportA reasonable reading of the evidence, not a stated fact.
Labelled as inferenceNot supported either way.
Shown as a gapThe agent keeps four things apart that most CVs blur together: the company context, the team outcome, your contribution, and what you owned. Participation never becomes ownership; company growth never becomes your personal impact.
Anatomy of a number
Every number gets nine questions.
Numbers are the most persuasive and most abused lines in a senior CV. Before one is used, the agent records its value, unit, period, baseline, denominator, scope, ownership, source and confidentiality (the Professional edition also records confidence).
If they do not align, the number is omitted, generalised or qualified. Given two conflicting figures, the agent never chooses the larger one.
“Cut costs by 30%” — a sentence most CVs contain.
Research rules
It looks you up carefully — or not at all.
Two identity anchors
Public information is accepted only when two professional identity details match, so a namesake is never merged into your record.
Sources it refuses
Data brokers, contact aggregators, irrelevant personal social media and any private information are excluded.
Consent scoped to your case
Research covers relevant public professional, role, company and opportunity information for your case only, and ends with it. You can restrict it, for example to “no web research”.
Built-in safety
Your documents are data, not instructions.
Built to resist manipulation
- Instructed to treat uploaded files and web pages as untrusted data
- Instructed to ignore instructions hidden inside a document
- A named reviewer checks every output, whatever the input said
One case at a time
- Never merges namesakes, other candidates, templates or prior cases
- Your facts stay inside your case and are never reused as knowledge
- It never claims memory, background work, sent emails or submitted applications
Quality gates
Fourteen checks. Then a person.
A file is released only after these checks, and every slide is rendered and inspected for clipping and overlap.
- Evidence
- Voice
- Search-partner and recruiter read
- Hiring-leader / mandate read
- ATS
- Privacy
- Accessibility
- Asset rights
- Visual
- Links
- Spreadsheets
- Structured data (JSON)
- Manifest
- ZIP
No borrowed logos, no fake faces
Unlicensed logos are rejected. Image generation is limited to abstract visuals; no synthetic likeness of you is ever created.
Facts stay editable
Your facts and data visuals stay as native, editable text and shapes — never baked into a generated image.
Accessible by default
Sufficient contrast, logical reading order, descriptive titles and alt-text notes.
Honest status
It tells you what it could not do.
Every item in the package carries one of five statuses. A file is never claimed unless it was created, opened, validated and linked. If one format fails, a validated alternative is delivered instead.
Built, opened and checked.
ReadyUsable now; the gaps are listed.
Ready, flaggedDoes not fit your case, so not padded in.
SkippedNeeds one specific missing item.
WaitingStated plainly rather than faked.
DisclosedThen a person
The agent drafts. A named reviewer decides.
Human review is required before anything is used externally. Your CVPRO reviewer checks the claims, numbers, voice and privacy tags, and signs off the package. See who reads your record and how we use AI.
Common questions
The standard, plainly
Who wrote this standard?
CVPRO did, from fifteen years of leadership hiring at Talpro. It is versioned like software — Executive V6.1.2 and Professional V5.1 — and newer versions supersede older ones where they conflict.
Can the agent be talked out of these rules?
The agents are instructed to treat uploaded documents and web pages as data, not instructions, and to ignore attempts to change their role or rules. Because no instruction is a guarantee, a named reviewer checks every output before release.
What happens if the agent cannot verify something?
It says so. The claim is classed as candidate-claimed, inferred or unknown, the affected file is marked “complete with evidence gaps”, and you are told exactly which item would close the gap.
Does a person still check the work?
Always. Human review is required before anything is used externally. See who reads your record.
Next step
Hold your record to this standard.
Book a scope call and we will show you how it applies to your career.
No job search, no applications, no interview or outcome promises. CVPRO is not an employer, recruiter or placement agency. ChatGPT and GPT are trademarks of OpenAI; CVPRO is not affiliated with or endorsed by OpenAI.