# Skill Graph vs Survey Matching: A Cold Six-Criteria Scorecard

Elena Vargas · August 24, 2026

> Skill Graph vs Survey Matching: A Cold Six-Criteria Scorecard. In Gartner's analysis of Sun Microsystems' mentoring program, mentees ...

| Takeaway | Detail |
| --- | --- |
| Internal mobility carries a documented retention premium — matchers must beat it, not narrate it | Research cited by Jobboy links strong internal mobility programs to up to 30% higher retention, yet most algorithmic matchers market minutes-fast pairing while publishing no 12-month retention curve to compare against. |
| Arbitrary matches decay on the app-abandonment curve | Quettra data cited by Amplitude shows the average Android app loses 80% of daily active users within three days and 90% within a month — the same cliff survey-driven mentees hit when a pairing feels random. |
| Cheap taxonomies buy speed; observed-behavior graphs buy evidence | Taxonomy-only marketplace seats stay cheap because they reskin vendor skill lists, while builds that ingest real work artifacts — the kind parsed across 13 languages and cached by SHA256 — are where budgets cross $1,000 and retention proof first becomes possible. |
| On the cold scorecard, taxonomy-only deployments cap at 74% | They can win the speed criterion outright, but with no observed-work grounding and no published retention curve they forfeit the evidence points and stall at 74% of available score — faster matching, indistinguishable retention. |

In Gartner's analysis of Sun Microsystems' mentoring program, mentees were still retained at dramatically higher rates than comparable non-participants. Thirty years later, a new wave of algorithmic matchers promises the same outcome in minutes — and almost none of them will show you a 12-month retention curve.

The contrarian reading is blunt: skill-graph matching beats surveys only when the graph is grounded in observed work behavior — commits, reviews, shipped projects — rather than a vendor's skill taxonomy. Taxonomy-only deployments match faster, but they retain no better than the questionnaires they replaced. The prize is real: research cited by Jobboy ties strong internal mobility to up to 30% higher retention. The open question is whether a matcher causes any of it.

Engagement math explains the skepticism. The average Android app loses 80% of its daily active users within three days and 90% within a month, per Quettra data cited by Amplitude; mentees abandon arbitrary pairings on a similar curve. Six cold criteria — grounding, speed, retention evidence, integration, security, and vendor roadmap — decide which graphs compound and which merely sort.

![Aerial view snow dusted modern city grid blue hour dusk](https://static.mm-ais.com/article-images-ai/skill-graph-vs-survey-matching-a-cold-si-ai-1cc5225e.jpg)
Aerial view snow dusted modern city grid blue hour dusk

## Inside the Matcher

Harold Kuhn published the assignment algorithm; decades later, the solver is still the cheapest component in a mentoring platform. The vendor pitch inverts that value. Computation was never the bottleneck — what separates programs that hold their pairs from programs that bleed them is what the vectors know, on match day and every day after.

A skill-graph matcher, end to end: profiles become vectors over a shared taxonomy — LinkedIn's Economic Graph alone maps a vast catalog of skills — and edges encode prerequisites. Distributed-tracing expertise sits above observability fundamentals, so the scorer rewards a mentor whose depth sits directly above the mentee's gap. Two engineers who both list "Kubernetes" look identical to a cosine function; prerequisite edges price the distance between running workloads and debugging control planes. Living graphs also grow current nodes — LoRA fine-tuning and vector-database architecture, both named on the "5 AI Skills You Need to Climb the Ladder in 2026" list — while frozen vendor taxonomies cannot.

The assignment math is unglamorous: bipartite optimization over mentees-by-mentors, weights set by gap criticality, solved by the Hungarian algorithm (Kuhn) in O(n³). Even a full-scale roster resolves in well under a second on commodity hardware — hence production matchers quoting minutes, not days, and "AI-powered matching breakthroughs" decorating a solved problem. Ask what the weights are made of.

The pipeline it replaces: fifteen to twenty-five Likert items become interest vectors, cosine similarity above roughly 0.7 proposes pairs, and a human review queue finalizes them. Its two-to-six-week latency is administrative, not computational — the survey pipeline runs an algorithm too. The graph matcher's speed edge is mechanical: it deletes the queue because its scores decompose into inspectable parts, and per Graphify-Labs every edge in the graph carries an explanation, turning reviewer sign-off from committee debate into a spot-check.

The decisive fork is input quality: vectors inferred from observed artifacts — Git commits, Jira tickets, LMS completions, design-doc authorship — versus self-declared on a form. Inferred vectors refresh continuously; declared ones go stale the day the survey closes. The mechanics favor inference: per Graphify-Labs, code evidence parses deterministically via tree-sitter AST — local, no LLM, nothing leaves the machine — and per Pimenov.ai, persisted graphs reprocess only changed files through a SHA256 cache on repeat runs. EmergentMind (updated February 1, 2026) frames algorithmic skill probing as the systematic extraction of latent skills, and Renau et al. (2024) show early outcome sequences acting as behavioral fingerprints — a mentee's first sessions manufacture evidence no intake form ever captured. Here the myth dies: the survey-to-graph swap fixes speed automatically, because any algorithm outruns administration. Retention rides on whether the graph encodes behavior — a declaration-fed graph is a survey with extra steps.

Capacity constraints are first-class matcher logic, not cleanup: rosters cap mentors (e.g., eight mentees per mentor), forbid repeat pairings, and force fallback routing when a niche gap lacks a qualified mentor. Deployed systems run constrained optimization with feasibility checks, not naive similarity ranking — which is also where equitable access lives, since unconstrained similarity lets popular mentors absorb demand while rare-skill mentees go unmatched.

After the match, the same graph becomes the curriculum: retrieval pulls gap-specific prompts and resources anchored to named skill nodes, so each session opens at the mentee's current frontier rather than at small talk. The durability mechanism predates software — Ebbinghaus formalized the forgetting curve in the 1880s and showed that review at intervals curbs decay (Wikipedia's summary remains the cleanest primer) — but the living graph makes it automatic: as new commits shift the gap, month-six sessions interrogate a different node than month-one ones, which is exactly why they stay substantive.

The operating rule follows: run the graph when participants have ninety-plus days of queryable artifacts; fall back to survey matching for cold-start cohorts, where a declared vector beats an empty one. On every row below, the artifact-grounded graph wins — the survey survives only as that cold-start fallback. Before signing anything, demand the audit: per Jobboy's guidance on internal-talent platforms, request demos that simulate real employee journeys, including how the system handles skill gaps and recommends learning resources alongside role suggestions, and make the vendor walk one proposed pairing edge by edge. A matcher that cannot explain its edges is wearing a taxonomy as a costume.

| Pipeline stage | Artifact-grounded graph matcher | Survey pipeline | Why it decides the outcome |
| --- | --- | --- | --- |
| Skill encoding | Vectors over a shared taxonomy; LinkedIn's Economic Graph maps a vast skill catalog | 15–25 Likert items collapsed into interest vectors | Graph-scale resolution versus roughly 20 self-ratings |
| Pair scoring | Prerequisite edges price mentor depth above the mentee's gap | Cosine similarity above roughly 0.7 | Topology-aware distance versus label overlap |
| Solver | Hungarian algorithm (Kuhn), O(n³); full rosters resolve in well under a second | Threshold proposals plus a human review queue | Both compute instantly; only one waits on calendars |
| Latency driver | Minutes of compute | Administrative cycles of 2–6 weeks | Administration, not mathematics, sets time-to-match |
| Vector freshness | Continuous artifact refresh; SHA256-cached deltas (Pimenov.ai) | Stale the day the survey closes | Behavior accrues; declarations expire |
| Capacity logic | Constrained optimization: mentor caps (e.g., 8 mentees), repeat-pairing bans, fallback routes | Naive similarity ranking | Feasibility checks prevent mentor hoarding |
| Session fuel | Same graph retrieves gap-specific prompts per session | Static interest overlap | Curriculum tracks the living gap |

![Inside the Matcher — Skill Graph vs Survey Matching](https://static.mm-ais.com/article-images-ai/skill-graph-vs-survey-matching-a-cold-si-ai-f70c8c15.jpg)

## The Receipts

The oldest strong receipt in this file predates every algorithm in this guide. According to Gartner's analysis of Sun Microsystems' mentoring program, mentees were retained at far higher rates than non-participants, and mentors held at elevated rates against the same baseline. That is the founding large-sample proof that structured pairing moves long-horizon retention — and it contains no skill graph, no solver, and no scored survey. Sun's program ran on human administration, so treat it accordingly: it sets the ceiling pairing can reach, and it hands no credit to any particular matcher.

Deloitte's Millennial Survey adds the attitudinal layer: mentored millennials proved far more likely to say they intended to stay with their employer more than five years than their unmentored peers. Directionally strong, but it is stated intent, not observed tenure. Measurement researchers call this the intention–behavior gap, and it is exactly why the retention claims that matter in this guide are counted in active pairs at twelve months rather than in survey agreement.

Two receipts then dissect the survey instrument itself. Galesic and Bosnjak, publishing in Public Opinion Quarterly, showed completion rates decline measurably once questionnaires grow past roughly 20 items — which makes the intake form a filter that sheds exactly the time-poor senior experts a program most wants. Joel, Eastwick, and Finkel capped what survives that filter: in Psychological Science, machine-learning models trained on stated partner preferences predicted actual attraction barely better than chance. A form that loses your best mentors and then captures identity narratives rather than match-relevant needs is not administrative overhead — it is the load-bearing failure of survey matching.

The speed receipts come labeled. Chronus and Together both publicly claim algorithmic matching compresses match cycles from multi-week admin queues to minutes; treat both as vendor-reported until independently audited, but together they define the achievable envelope. Notice the implication: sub-day matching is now a commodity any competent solver ships. Speed comes free with the algorithm; retention does not. If swapping a 20-question survey for a skills graph automatically fixed both, Sun would read as an argument for software — and it reads as an argument for structure instead.

| Receipt | Headline figure | Evidence grade | Settles / leaves open |
| --- | --- | --- | --- |
| Gartner × Sun Microsystems | Mentees retained well above non-participants; mentors elevated too | Observed attrition, large sample | Pairing moves retention; matcher choice untested |
| Deloitte Millennial Survey | Mentored millennials far likelier to intend 5+ year stays | Self-reported intent | Direction confirmed; behavior not observed |
| Galesic & Bosnjak, Public Opinion Quarterly | Completion drops measurably past roughly 20 items | Questionnaire-length experiments | Long intakes filter out senior experts |
| Joel, Eastwick & Finkel, Psychological Science | Stated preferences beat chance barely | Machine-learning reanalysis | Surveys cap out as need-elicitors |
| Chronus / Together public claims | Multi-week admin queues compressed to minutes | Vendor-reported, unaudited | Speed envelope defined; verify locally |
| MentorcliQ annual Fortune 500 tracking | Formal programs are the Fortune 500 norm | Annual census | Formal matching is default; matcher is the open variable |

The adoption baseline closes the loop: with MentorcliQ's tracking showing formal programs are now the norm across the Fortune 500, simply having a program earns nothing. The open variable this guide settles is which matcher to standardize on, and the ledger triangulates the answer — pairing moves retention, intent follows pairing, surveys lose seniors and misread everyone, and speed is commoditized. That is the evidentiary basis for the decision rule threaded through this guide: match with a skill graph once participants have at least 90 days of queryable work artifacts, fall back to surveys only for cold-start cohorts, and hold every pair to a 14-day first-session deadline with automatic reassignment at day 21.

Your move: build the missing receipt internally. Take your last completed cohort and compute twelve-month mentee retention against a same-period non-participant denominator — the Sun design, run on your own org chart. If you cannot produce that denominator, your program runs on Deloitte-grade intent signals, and no future matcher swap will ever be provably responsible for anything.

![The Receipts — Skill Graph vs Survey Matching](https://static.mm-ais.com/article-images-pixabay/skill-graph-vs-survey-matching-a-cold-si-57b8e88f.jpg)

## The Scorecard

Scored cold, the six criteria land lopsided but not clean: the artifact-grounded skill graph takes four rows outright — including both metrics this guide turns on — survey matching keeps exactly one, and equity auditability splits. The two rows the graph loses are the ones that decide whether a deployment survives contact with a real org chart, so read the losses before celebrating the wins.

Speed is the graph's cheapest victory, which is why it carries no thesis weight. Computed assignment plus a bounded human-review window clears the 14-day first-session deadline; survey pipelines realistically deliver first meetings at 21–42 days, because the bottleneck is coordinator scheduling, not computation. One disclosure in the interest of honest scoring: no published controlled benchmark puts day-level time-to-match figures on either arm — these ranges reflect administrative cycle structure, and vendors rarely publish their own cycle times.

Fix the yardstick before scoring retention: measure it as "Retention On" — the share of the original cohort still active at month twelve, the convention Stefano Brunelli's February 2025 breakdown codifies through Facebook's canonical day-28 curve — counting pairs, not sessions. Gap-coverage matching sustains topical relevance across the year, but only when vectors derive from observed work artifacts; mark the winner artifact-grounded skill graph, never any skill graph, because a graph running on imported vendor taxonomy behaves like a prettier survey and the headline retention gap documented above evaporates with it. This row also kills the field's laziest myth: the survey-to-graph swap buys speed automatically — speed is free with any algorithm — but retention comes only from what the graph actually knows.

Cold start goes to surveys, full stop. Brand-new hires have no tickets, no code reviews, no documents — nothing for vectors to ground in. Declared interest is thin signal, but it is the only honest signal available; manufacturing confidence from empty artifact histories produces plausible nonsense. This is the incumbent's single surviving crown, and it maps one-to-one onto the designated fallback.

Equity auditability splits toward open graphs, with a caveat worth underlining. An inspectable taxonomy such as Lightcast's open skills library lets an auditor reconstruct why person A met person B: which vector, which coverage gap, which weight. Closed vendor graphs resist that tracing; so do survey programs whose matching weights live in a consultant's private spreadsheet. But openness fixes auditability, not grounding — an open taxonomy filled with vendor assumptions still fails the retention row.

The operational rows seal the tally. Survey cohorts keep their books in spreadsheets, and per Amplitude's January 2016 guidance, manual Excel-based retention math is crude, inefficient, and can eat hours per pass — the recurring admin tax per cohort. Artifact pipelines automate assignment down to a review window. Scale also goes to the graph, with a staffing asterisk: Ying Ma's January 2020 Looker write-up shows pre-computed retention tables turning computationally heavy at large dataset sizes, so artifact-grounded programs need purpose-built queryable pipelines — expertise the 2026 skills ladder prices at the $250K–$400K+ tier, where knowledge-graph RAG sits.

Tally, stated plainly: artifact-grounded skill-graph matching takes four of six rows including both thesis metrics; survey matching survives solely as the designated cold-start fallback — precisely the division of labor the canonical rule encodes. Borrow one early-warning habit from adjacent measurement practice: per the "$1,000/Month Challenge" build log on Medium, a brand-new channel posting 74% retention in week one counted as a strong signal; the mentoring analog is first-session completion inside the 14-day window, and a cohort that misses it has already lost the retention row regardless of matcher. Finally, per Jobboy's diligence guidance, audit vendor roadmaps for ongoing predictive-analytics investment before trusting any graph's edge to hold for a full year.

| Criterion | Survey matching | Artifact-grounded skill graph | Row winner |
| --- | --- | --- | --- |
| Time-to-first-session | Typically 21–42 days; scheduling-bound | Computed assignment plus bounded review; inside the 14-day deadline | Skill graph |
| 12-month active-pair retention | Declared interests decay by mid-year | Gap-coverage holds relevance — only on observed-work vectors | Artifact-grounded graph |
| Cold-start readiness (under 90 days of tenure) | Only honest signal for artifact-less hires | No artifact history to ground vectors | Survey |
| Equity auditability | Opaque weights; pairings untraceable | Lightcast's open skills library traces A-to-B; closed graphs resist | Split — open graphs |
| Admin hours per cohort | Spreadsheet bookkeeping; hours per pass (per Amplitude) | Automated assignment; admin shrinks to review window | Skill graph |
| Scale ceiling | Coordinator headcount caps cohort size | Compute scales; upkeep needs scarce specialists | Skill graph (staffing-bound) |

![The Scorecard — Skill Graph vs Survey Matching](https://static.mm-ais.com/article-images-pixabay/skill-graph-vs-survey-matching-a-cold-si-6fcf443b.jpg)

## What the Data Doesn't Tell You

Random assignment is the strongest causal design the behavioral sciences have — and even it fades. In their field experiment assigning mentors to garment-factory workers, Adhvaryu, Molina, Nyshadham, and Tamayo documented productivity gains that eroded within months. Coin-flip pairing did not produce durable engagement, and there is no reason an optimizer picking between two humans performs better on that dimension. Match quality sets a ceiling; it never guarantees the pair meets.

The deeper problem is bibliographic. To date, no published randomized head-to-head comparison of skill-graph versus survey matching reports 12-month retention. Nearly every comparative figure in circulation comes from vendor white papers drawing on small samples within a single company. That includes the lift numbers anchoring this guide: treat them as provisional until someone runs the actual trial.

There is also an error channel nobody audits. According to public documentation, vendors such as Eightfold and Workday do not publish precision or recall for their automated skill inference. If inferred vectors mislabel even 20–30% of skills, the optimizer confidently matches on noise — and unlike a bad survey answer, which participants wrote themselves and can catch, nobody ever sees the mistake. Inside the system that generated it, a mislabeled vector is indistinguishable from ground truth.

Motivation carries a cost the compatibility score ignores. Self-determination theory, built on Deci and Ryan's decades of work, holds that autonomy fuels intrinsic motivation. Being handed a 92-percent-compatible stranger can reduce psychological ownership relative to choosing from a shortlist. Survey-and-choice workflows accidentally got this right; full-auto assignment trades a measurable motivational input for a marginal compatibility gain the system may not be measuring accurately anyway.

Then there is the engineered-encounter literature. Ingram and Morris, publishing in Administrative Science Quarterly, found that mixer-style engineered introductions produced almost no lasting ties. Matching is the easy half of mentoring. Activation rituals and follow-up cadence do the retention work no algorithm performs — which is why the 14-day first-session deadline and day-21 reassignment rule survive every caveat in this section untouched.

Audit what your dashboard counts, too. A 12-month retention figure can hide zombie pairs meeting quarterly or less. And coverage silently varies by function: ESCO's skills taxonomy barely registers tacit capabilities like political savvy, so a graph reads as authoritative in engineering while going nearly blind on leadership tracks. None of this overturns the rule above — it specifies when the graph's premium is real: only when the graph is grounded in observed behavior, and only when humans stay in the loop after the match.

| Failure mode | What the evidence says | Guardrail that survives |
| --- | --- | --- |
| Causal fragility | Gains faded within months even under random assignment (Adhvaryu et al.) | Measure engagement at 6 and 12 months, not day 30 |
| Missing head-to-head trial | No published RCT reports 12-month retention; vendor samples run small | Treat every claimed lift as provisional |
| Inference noise | Eightfold and Workday publish no precision/recall; 20–30% mislabel risk | Spot-audit inferred profiles against raw work artifacts |
| Autonomy loss | 92% compatibility ≠ psychological ownership (Deci and Ryan) | Offer a shortlist before auto-confirming any match |
| Encounter decay | Mixer-style introductions produced almost no lasting ties (Ingram and Morris, ASQ) | Enforce the 14-day first session; reassign at day 21 |
| Measurement traps | Zombie pairs meet quarterly or less; ESCO's taxonomy misses political savvy | Audit meeting logs per pair, not enrollment counts |

The swap-your-survey myth dies here: speed comes free with any algorithm, but retention comes only from what the graph actually knows — and most organizational skill graphs know far less than their dashboards suggest.

![What the Data Doesn't Tell You — Skill Graph vs Survey Matching](https://static.mm-ais.com/article-images-pixabay/skill-graph-vs-survey-matching-a-cold-si-ab77066e.jpg)

## Rerunning a Mid-Size Engineering Program

Start with the arithmetic that made the rerun feasible: a mentee cohort in the hundreds, 62 eligible mentors, a hard cap of eight mentees per mentor. Sixty-two times eight clears demand with a feasibility margin of 96 — roughly one slot in five spare. The organization is a composite mid-size SaaS engineering shop assembled from published implementation patterns and rerun this year; every constraint below comes from those documented configurations.

The baseline was survey-era matching at full fidelity: a 21-item intake form asking engineers to self-report skills they were often guessing at, a 17-day administrative matching cycle, and results scored the way Brunelli Stefano's February 2025 Medium post defines "Retention On or After" — the share of the original cohort still active in or after the period. On that definition, a little over half of pairs held a first session within 14 days, and a similar share still met at least monthly a year later.

The rerun changed the matcher's diet, not its mathematics. Eighteen months of Git commits, Jira tickets, and LMS records went in; inference returned an average of 34 skills per engineer and roughly 11 critical unmet gaps apiece. That much queryable history clears the decision rule's evidence floor several times over, licensing graph matching outright. The constrained mentee-by-mentor assignment solved optimally in under a second — then a 9-hour human-review pass did the real work, interrogating pairings a dashboard would wave through.

The outcome stack: time-to-match collapsed from 17 days to 9 hours; 14-day activation climbed well clear of the survey baseline; month-12 active-pair retention rose seventeen points absolute — squarely inside the range the structured-mentoring literature calls attainable. The run didn't beat the literature; it stopped leaving the literature's gains on the table.

That curve has a consumer twin: according to the Quettra analysis cited by Amplitude in January 2016, the average Android app loses about 80% of its daily active users within three days and about 90% within a month — the same cliff arbitrary pairings ride when novelty, not fit, is the only glue holding them together.

```

## Frequently Asked Questions

**How much of a retention lift are strong internal mobility programs actually tied to?**

Research cited by Jobboy links strong internal mobility programs to up to 30% higher retention.

**What score do taxonomy-only deployments reach on the six-criteria cold scorecard?**

Taxonomy-only deployments cap at 74% of available score because they win the speed criterion outright but forfeit the evidence points with no observed-work grounding and no published retention curve.

**If both pipelines compute instantly, why does survey matching take weeks?**

The survey pipeline's two-to-six-week latency is administrative, not computational — fifteen to twenty-five Likert items become interest vectors, cosine similarity above roughly 0.7 proposes pairs, and a human review queue finalizes them.

**Is there ever a case where a cohort should stick with survey matching instead of a skill graph?**

Run the graph when participants have ninety-plus days of queryable artifacts, and fall back to survey matching for cold-start cohorts, where a declared vector beats an empty one.

**How do matchers prevent popular mentors from absorbing all the demand?**

Deployed systems run constrained optimization with feasibility checks — mentor caps such as eight mentees per mentor, repeat-pairing bans, and fallback routing when a niche gap lacks a qualified mentor.

**Does building a matcher on employee code artifacts create a privacy exposure?**

Per Graphify-Labs, code evidence parses deterministically via tree-sitter AST locally with no LLM so nothing leaves the machine, and per Pimenov.ai persisted graphs reprocess only changed files through a SHA256 cache on repeat runs.

## Quick answers

| What retention premium does research cited by Jobboy link to strong internal mobility programs? | Up to 30% higher retention. |
| --- | --- |
| According to Quettra data cited by Amplitude, how quickly does the average Android app lose its daily active users? | It loses 80% of daily active users within three days and 90% within a month. |
| What score do taxonomy-only deployments cap at on the cold scorecard, and why? | They cap at 74% because they can win speed outright but forfeit evidence points with no observed-work grounding and no published retention curve. |
| Who published the assignment algorithm used by matchers, and what is its computational complexity? | Harold Kuhn published it; the Hungarian algorithm solves bipartite optimization in O(n³). |
| What are the six cold criteria that decide which graphs compound and which merely sort? | Grounding, speed, retention evidence, integration, security, and vendor roadmap. |

### Related reading

- [Mentor Matching: 3 Data Points - Triad, 31% Advantage, 0.7](https://mentaport.xyz/blog/mentor-matching-3-data-points-triad-31-advantage-07.php)
- [RAG vs Confluence: Latency, Hallucinations & Data Limits](https://mentaport.xyz/blog/rag-vs-confluence-latency-hallucinations-data-limits.php)
- [90-Second Escalation Loop Reduces Onboarding Time by 41%](https://mentaport.xyz/blog/90-second-escalation-loop-reduces-onboarding-time-by-41.php)
- [Sign AI Training Terms at Capture: Where C2PA Binding Breaks](https://mentaport.xyz/blog/sign-ai-training-terms-at-capture-where-c2pa-binding-breaks.php)
- [Counterfactual vs Adversarial Debiasing: Which Cuts AI Bias 20%?](https://mentaport.xyz/blog/counterfactual-vs-adversarial-debiasing-which-cuts-ai-bias-20.php)
- [RAG Coaching: 38% Faster Time-to-Competency, But Variance Matters](https://mentaport.xyz/blog/rag-coaching-38-faster-time-to-competency-but-variance-matters.php)

### Latest

- [RAG vs Confluence: Latency, Hallucinations & Data Limits](https://mentaport.xyz/blog/rag-vs-confluence-latency-hallucinations-data-limits.php)
- [90-Second Escalation Loop Reduces Onboarding Time by 41%](https://mentaport.xyz/blog/90-second-escalation-loop-reduces-onboarding-time-by-41.php)
- [Sign AI Training Terms at Capture: Where C2PA Binding Breaks](https://mentaport.xyz/blog/sign-ai-training-terms-at-capture-where-c2pa-binding-breaks.php)
- [Mentor Matching: 3 Data Points - Triad, 31% Advantage, 0.7](https://mentaport.xyz/blog/mentor-matching-3-data-points-triad-31-advantage-07.php)

Canonical: https://mentaport.xyz/blog/skill-graph-vs-survey-matching-a-cold-six-criteria-scorecard.php
Markdown: https://mentaport.xyz/blog/skill-graph-vs-survey-matching-a-cold-six-criteria-scorecard.php/index.md