Recruitment Visible World Demo

Target

Visible World Lab · https://visible.shiroand.io/lab/recommendation

Observation period

2026-09-20T04:26:16.900Z → 2026-09-20T04:26:33.832Z

Included runs

A · Senior Marketing / AI Interest · cmu9bdehh0001l204k3d1sfy5

B · Senior Marketing / Security Interest · cmu9bdelc0005l2047ynx1ik9

C · C — Neutral Control · cmu9bdeoc0009l2043iw2olv0

Scope is Top 10. Historical runs are not mixed into these figures unless listed above.

In the controlled recruitment experiment, personas began from an identical visible world.

Following different observed interaction sequences, Persona A and Persona B received measurably different refreshed exposure sets, while the neutral control remained unchanged.

Persona A’s AI / GenAI representation increased from 30% to 60% in the observed Top 10.

Persona A vs Neutral produced Visible World Divergence 0.400.

Persona B vs Neutral produced Visible World Divergence 0.415.

The neutral control returned Visible World Divergence 0.000.

• Three synthetic personas started from an identical visible world of 24 JOB-* listings.
• Persona A recorded AI / GenAI clicks, then refreshed.
• Persona B recorded Cybersecurity clicks, then refreshed.
• Persona C recorded no item clicks and refreshed as an untouched control.
• Exposure, actions, and comparisons are stored as evidence records.

Personas began from an identical visible world of 24 JOB-* listings before recorded interactions.

• OBSERVEDNOT TESTED
  3 listings were clicked.obs cmu9bditk002 · act cmu9bdlhb003 · act cmu9bdmwm004 · act cmu9bdoq5004
• OBSERVEDNOT TESTED
  3 listings were clicked.obs cmu9bdit3001 · act cmu9bdko7003 · act cmu9bdm3g003 · act cmu9bdniv004

A — AI Interest began with the common baseline Top 10. 3 AI / GenAI listings were then clicked. After the visible world was refreshed, 3 additional listings entered the observed Top 10. AI / GenAI representation increased from 30% to 60%. Top-10 Jaccard overlap between the initial and refreshed exposure sets was 0.54. Visible World Divergence was 0.400. These observations describe a change in exposure. They do not establish an unknown external ranking mechanism.

B — Security Interest began with the common baseline Top 10. 3 Cybersecurity listings were then clicked. After the visible world was refreshed, 3 additional listings entered the observed Top 10. AI / GenAI representation decreased from 30% to 20%. Top-10 Jaccard overlap between the initial and refreshed exposure sets was 0.54. Visible World Divergence was 0.415. These observations describe a change in exposure. They do not establish an unknown external ranking mechanism.

Persona A and Persona B shared 43% Jaccard overlap in the observed Top 10.

Persona A and the neutral control produced VWD 0.400.

Persona B and the neutral control produced VWD 0.415.

C — Neutral Control began from the common baseline Top 10. No recorded item clicks preceded the refresh. The untouched control retained an identical observed visible world. Top-10 Jaccard overlap between the initial and refreshed exposure sets was 1.00. Visible World Divergence was 0.000. These observations describe a change in exposure. They do not establish an unknown external ranking mechanism.

Jaccard

1.00

Churn

0

Rank displacement

0.00

TVD

0.000

VWD

0.000

• OBSERVEDNOT TESTED
  An email click was recorded before the web observation run.obs cmu9bditk002
• OBSERVEDNOT TESTED
  3 listings were clicked.obs cmu9bditk002 · act cmu9bdlhb003 · act cmu9bdmwm004 · act cmu9bdoq5004
• OBSERVEDNOT TESTED
  3 new items entered the observed Top 10.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  Top-10 overlap was 54%.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  AI / GenAI representation increased by 30 percentage points.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  Cybersecurity representation decreased by 10 percentage points.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  General B2B SaaS representation decreased by 10 percentage points.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  Industrial / Edge / Physical AI representation decreased by 10 percentage points.obs cmu9bditk002 · obs cmu9bdrka006 · cmp cmu9bdrm3007
• OBSERVEDNOT TESTED
  An email click was recorded before the web observation run.obs cmu9bdit3001
• OBSERVEDNOT TESTED
  3 listings were clicked.obs cmu9bdit3001 · act cmu9bdko7003 · act cmu9bdm3g003 · act cmu9bdniv004
• OBSERVEDNOT TESTED
  3 new items entered the observed Top 10.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  Top-10 overlap was 54%.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  AI / GenAI representation decreased by 10 percentage points.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  Cybersecurity representation increased by 30 percentage points.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  General B2B SaaS representation decreased by 10 percentage points.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  Industrial / Edge / Physical AI representation decreased by 10 percentage points.obs cmu9bdit3001 · obs cmu9bdqen005 · cmp cmu9bdqgp006
• OBSERVEDNOT TESTED
  Top-10 overlap was 100%.obs cmu9bdiwp003 · obs cmu9bdn2m004 · cmp cmu9bdn53004
• OBSERVEDNOT TESTED
  Persona A and Persona B shared 43% Jaccard overlap in the observed Top 10.obs cmu9bdrka006 · obs cmu9bdqen005
• OBSERVEDNOT TESTED
  Persona A and the neutral control produced VWD 0.400.obs cmu9bdrka006 · obs cmu9bdn2m004 · cmp cmu9bdvzx000
• OBSERVEDNOT TESTED
  Persona B and the neutral control produced VWD 0.415.obs cmu9bdqen005 · obs cmu9bdn2m004 · cmp cmu9bdw2m000

• Known lab mechanism
  finalScore = baseScore + sum(personaInterestWeight[tag] * jobTagWeight)
• Known lab mechanism
  VIEW = 0
• Known lab mechanism
  CLICK = +1.0
• Known lab mechanism
  SAVE = +1.5
• Known lab mechanism
  DISMISS = -0.75
• Known lab mechanism
  Equal scores break ties by externalItemId
• Known lab mechanism
  In this controlled environment, CLICK adds +1.0 to the corresponding observed interaction weight.
• Known lab mechanism
  Displayed ranking changes only after REFRESH VISIBLE WORLD.
• Known lab mechanism
  This information is known because SHIRO & Co. controls the test environment.

• INFERENCENOT ESTABLISHED
  The observed interaction sequence is associated with the subsequent divergence.obs cmu9bditk002 · obs cmu9bdrka006 · obs cmu9bdit3001 · obs cmu9bdqen005 · obs cmu9bdiwp003 · obs cmu9bdn2m004 · act cmu9bdlhb003 · act cmu9bdmwm004 · act cmu9bdoq5004 · act cmu9bdko7003 · act cmu9bdm3g003 · act cmu9bdniv004 · cmp cmu9bdrm3007 · cmp cmu9bdqgp006 · cmp cmu9bdn53004 · cmp cmu9bdvzx000 · cmp cmu9bdw2m000
• INFERENCENOT ESTABLISHED
  The events appear to form one exposure sequence.

• UNKNOWNNOT ESTABLISHED
  How an external recruitment platform would weight equivalent interactions.
• UNKNOWNNOT ESTABLISHED
  Whether the same sequence would produce a persistent effect over multiple days.
• UNKNOWNNOT ESTABLISHED
  Whether exposure would recover after behavior changes.

• Known lab mechanism
  KNOWN IMPLEMENTATION RELATIONSHIP: the controlled lab recommendation engine is explicitly coded to use interaction state.
• UNKNOWNNOT ESTABLISHED
  This known implementation relationship is not an experimental causal claim about an external platform.
• UNKNOWNNOT ESTABLISHED
  For external systems, causal status defaults to NOT ESTABLISHED.
• UNKNOWNNOT ESTABLISHED
  The email click is not established as the cause of later ranking changes.

• This is a controlled synthetic environment.
• The mock recommendation engine is deterministic and owned by SHIRO & Co.
• Personas are synthetic experimental variables, not claims about real individuals.
• No inference about real individuals is licensed by this report.
• No external-platform ranking mechanism was observed.
• OPENED is an observed email-open event. It does not necessarily represent human reading or engagement.
• Visible World Divergence is an internal experimental comparison metric. It is not a validated industry standard.
• Association language in this report must not be generalized into causal claims about external systems.
• Observations cover the defined audit window only.
• Visible World Audit observes outputs, not proprietary source code.
• Unseen candidates cannot be assumed to have been deliberately suppressed.
• Semantic classifications are inferences from observed text.
• Causal relationships are not established unless separately demonstrated.
• Visible World Divergence is an experimental internal metric that compares observed exposure-set overlap, rank movement, and category distribution. It is not a validated industry standard. It is not a judgment that a system is good, bad, safe, unsafe, biased, or unbiased.
• OBSERVED EMAIL OPEN EVENT. OPENED does not necessarily represent human reading. Some email clients may preload or proxy content. OPENED is not equivalent to engagement.
• This engagement used a controlled synthetic target, not an external production ranking system.

Visible World Divergence is an internal experimental comparison metric. It is not a validated industry standard.