1. Candidate evidence
The validator receives only the source text the tool supplies, such as resume context, job details, or stated preferences. Missing context cannot be verified.
The honest architecture
How ApplyArc checks AI-generated claims
ApplyArc uses deterministic metric and rewrite guards on selected AI paths. They catch specific failure modes. They do not make AI infallible, prove authorship, or replace your final factual check.
What the checks do, and where they stop
A validator can compare output with supplied evidence. It cannot know facts you never provided or judge whether a softer claim is misleading.
2. Deterministic check
metricVerifier compares generated numbers and durations with that source text. Bullet guards separately check wording, scope, leadership, and evidence changes.
3. Tool-specific response
The tool may retry, clean, withhold, or record a validation failure. The response depends on the output format and whether the tool streams.
metricVerifier.js · simplified
const check = verifyMetrics(output, candidateSources);
if (!check.ok) {
return { ok: false, reason: check.reason, hint: check.hint };
}
return { ok: true };235+
language patterns catalogued
11
structural text detectors
9
Tools share unified memory
2
maximum Resume Tailor repairs
1
final factual reviewer: you
What ApplyArc checks
- Unsupported numeric tokens and spelled-out durations where metricVerifier is wired
- Invented leadership, expanded scope, and weakened evidence in bullet rewrites
- A fixed catalogue of generic wording and rhythm patterns
- Tool-specific validation failures in telemetry
What you still check
- Whether every claim is true in its full context
- Dates, titles, names, and outcomes that were absent from your source text
- Whether the wording sounds like you
- Whether the final document matches the employer's instructions
Frequently Asked Questions
No universal claim would be honest. Narrative tools wire output validation according to their format. Resume Tailor and Outreach can reject and retry unsupported metrics. Streaming tools use tool-specific post-processing or telemetry because replaying a stream would flicker. Bullet rewrites use separate deterministic claim guards.
metricVerifier extracts numeric tokens and spelled-out durations from generated text, then compares them with candidate-provided source text. It returns a failure reason and a corrective prompt when a figure is unsupported. The calling tool decides whether to retry, clean, withhold, or surface the issue.
Yes. A numeric verifier cannot prove a qualitative claim, infer missing context, or replace your final review. ApplyArc adds deterministic checks, but you must confirm that every fact, date, title, and outcome is yours before sending.
Voice Guard scores a fixed set of generic wording and rhythm patterns from 0 to 10. Separate bullet guards check scope expansion, invented leadership, weakened evidence, and unsupported metrics. These are editing checks, not proof of authorship or a prediction of ATS treatment.
It depends on the tool. Resume Tailor allows up to two bounded repair attempts. Outreach allows one. The anonymous bullet rewriter withholds a rejected rewrite. Some streamed tools clean safe issues after generation and record validation failures for monitoring instead of replaying the stream.
Where a tool has a validator or guard, that code path is part of the tool rather than a paid safety add-on. The free plan includes 5 AI generations each month. Pro changes limits and, for selected tools, the model used.
Use AI. Keep the final say.
5 AI generations each month on the free plan. Check every claim before you send it.
No credit card · Pro is £19/month
Related Resources
Anti-AI Resume Rewriter→
See what Voice Guard checks and what it cannot prove
ApplyArc vs Teal→
Compare current features, limits, and pricing
ApplyArc vs Huntr→
Compare job tracking and AI workflows
Pricing→
£19/month · 18 AI tools · cancel anytime
Make an AI Resume Sound Human→
Replace generic wording with accurate evidence