SHIRO & Co.

Visible World Audit

Internal audit view · DRAFT

Audit session

Visible World Audit Report · v0.1

Recruitment Visible World Demo

Generated 2026-09-20 07:53:39Z

Scope

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.

Executive summary

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.

Channel sequence

Persona A

Email

1 delivered · 1 opened · 1 clicked

Web

1 landing exposure · 3 item clicks · 1 refreshed ranking

Persona B

Email

1 delivered · 0 opened · 1 clicked

Web

1 landing exposure · 3 item clicks · 1 refreshed ranking

Persona C

Email

1 delivered · 0 opened · 0 clicked

Web

1 landing exposure · 0 item clicks · 1 refreshed ranking

Observation design

  • 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.

Baseline

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

Observed 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

Visible world changes

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 comparison

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

Control comparison

Persona A and the neutral control produced VWD 0.400.

Persona B and the neutral control produced VWD 0.415.

Neutral control

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

Category shifts

A

CategorySourceTargetDelta
AI / GenAI30%60%+30pp
Cybersecurity20%10%-10pp
General B2B SaaS30%20%-10pp
Industrial / Edge / Physical AI20%10%-10pp

B

CategorySourceTargetDelta
AI / GenAI30%20%-10pp
Cybersecurity20%50%+30pp
General B2B SaaS30%20%-10pp
Industrial / Edge / Physical AI20%10%-10pp

C

CategorySourceTargetDelta
AI / GenAI30%30%0pp
Cybersecurity20%20%0pp
General B2B SaaS30%30%0pp
Industrial / Edge / Physical AI20%20%0pp

Exposure chain

    A

  1. 04:20:13

    OBSERVEDEMAIL DELIVEREDPossible link

    Channel event EMAIL DELIVERED: Synthetic A production validation notice

    Senior Marketing / AI Interest · CHANNEL EVENT

    Evidence

    ↓

  2. 04:20:21

    OBSERVEDEMAIL OPENEDPossible link

    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.

    Senior Marketing / AI Interest · CHANNEL EVENT

    Evidence

    ↓

  3. 04:20:29

    OBSERVEDEMAIL CLICKEDPossible link

    Channel event EMAIL CLICKED: Synthetic A production validation notice

    Senior Marketing / AI Interest · CHANNEL EVENT

    Evidence

    ↓

  4. 04:26:22

    OBSERVEDINITIAL WORLD

    Senior Marketing / AI Interest was shown JOB-AI-001 at rank 1 in the INITIAL visible world.

    Senior Marketing / AI Interest · INITIAL

    Evidence

    ↓

  5. 04:26:25

    OBSERVEDOBSERVED EVENT

    Senior Marketing / AI Interest clicked JOB-AI-001.

    Senior Marketing / AI Interest · ACTION

    Evidence

    ↓

  6. 04:26:27

    OBSERVEDOBSERVED EVENT

    Senior Marketing / AI Interest clicked JOB-AI-002.

    Senior Marketing / AI Interest · ACTION

    Evidence

    ↓

  7. 04:26:30

    OBSERVEDOBSERVED EVENT

    Senior Marketing / AI Interest clicked JOB-AI-003.

    Senior Marketing / AI Interest · ACTION

    Evidence

    ↓

  8. 04:26:31

    OBSERVEDOBSERVED EVENT

    Senior Marketing / AI Interest refreshed the visible world.

    Senior Marketing / AI Interest · REFRESH

    Evidence

    ↓

  9. 04:26:33

    OBSERVEDREFRESHED WORLD

    Senior Marketing / AI Interest was shown JOB-AI-001 at rank 1 in the REFRESHED visible world.

    Senior Marketing / AI Interest · REFRESHED

    Evidence

    ↓

  10. 04:26:33

    OBSERVEDNOT ESTABLISHEDOBSERVED DIFFERENCE

    Observed difference between recorded exposure sets. Top-10 Jaccard overlap was 0.54. Visible World Divergence was 0.400.

    Senior Marketing / AI Interest · INITIAL vs REFRESHED

    Evidence

    B

  1. 04:20:33

    OBSERVEDEMAIL DELIVEREDPossible link

    Channel event EMAIL DELIVERED: Synthetic B production validation notice

    Senior Marketing / Security Interest · CHANNEL EVENT

    Evidence

    ↓

  2. 04:20:45

    OBSERVEDEMAIL CLICKEDPossible link

    Channel event EMAIL CLICKED: Synthetic B production validation notice

    Senior Marketing / Security Interest · CHANNEL EVENT

    Evidence

    ↓

  3. 04:26:22

    OBSERVEDINITIAL WORLD

    Senior Marketing / Security Interest was shown JOB-AI-001 at rank 1 in the INITIAL visible world.

    Senior Marketing / Security Interest · INITIAL

    Evidence

    ↓

  4. 04:26:24

    OBSERVEDOBSERVED EVENT

    Senior Marketing / Security Interest clicked JOB-SEC-001.

    Senior Marketing / Security Interest · ACTION

    Evidence

    ↓

  5. 04:26:26

    OBSERVEDOBSERVED EVENT

    Senior Marketing / Security Interest clicked JOB-SEC-002.

    Senior Marketing / Security Interest · ACTION

    Evidence

    ↓

  6. 04:26:28

    OBSERVEDOBSERVED EVENT

    Senior Marketing / Security Interest clicked JOB-SEC-003.

    Senior Marketing / Security Interest · ACTION

    Evidence

    ↓

  7. 04:26:30

    OBSERVEDOBSERVED EVENT

    Senior Marketing / Security Interest refreshed the visible world.

    Senior Marketing / Security Interest · REFRESH

    Evidence

    ↓

  8. 04:26:32

    OBSERVEDREFRESHED WORLD

    Senior Marketing / Security Interest was shown JOB-SEC-001 at rank 1 in the REFRESHED visible world.

    Senior Marketing / Security Interest · REFRESHED

    Evidence

    ↓

  9. 04:26:32

    OBSERVEDNOT ESTABLISHEDOBSERVED DIFFERENCE

    Observed difference between recorded exposure sets. Top-10 Jaccard overlap was 0.54. Visible World Divergence was 0.415.

    Senior Marketing / Security Interest · INITIAL vs REFRESHED

    Evidence

    C

  1. 04:20:53

    OBSERVEDEMAIL DELIVEREDPossible link

    Channel event EMAIL DELIVERED: Synthetic C production validation notice

    C — Neutral Control · CHANNEL EVENT

    Evidence

    ↓

  2. 04:26:22

    OBSERVEDINITIAL WORLD

    C — Neutral Control was shown JOB-AI-001 at rank 1 in the INITIAL visible world.

    C — Neutral Control · INITIAL

    Evidence

    ↓

  3. 04:26:26

    OBSERVEDOBSERVED EVENT

    C — Neutral Control refreshed the visible world.

    C — Neutral Control · REFRESH

    Evidence

    ↓

  4. 04:26:27

    OBSERVEDREFRESHED WORLD

    C — Neutral Control was shown JOB-AI-001 at rank 1 in the REFRESHED visible world.

    C — Neutral Control · REFRESHED

    Evidence

    ↓

  5. 04:26:28

    OBSERVEDNOT ESTABLISHEDOBSERVED DIFFERENCE

    Observed difference between recorded exposure sets. Top-10 Jaccard overlap was 1.00. Visible World Divergence was 0.000.

    C — Neutral Control · INITIAL vs REFRESHED

    Evidence

What was observed

  • 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

What is known

  • 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.

What is inferred

  • 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.

What remains unknown

  • 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.

Causal status

  • 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.

Limitations

  • 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.

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

Evidence index

IDRecordEvidence
E-001A — AI Interest INITIAL · Senior Marketing / AI InterestOpen
E-002CLICK JOB-AI-001 · Senior Marketing / AI InterestOpen
E-003CLICK JOB-AI-002 · Senior Marketing / AI InterestOpen
E-004CLICK JOB-AI-003 · Senior Marketing / AI InterestOpen
E-005A — AI Interest REFRESHED · Senior Marketing / AI InterestOpen
E-006A — AI Interest INITIAL vs REFRESHED · Senior Marketing / AI InterestOpen
E-007B — Security Interest INITIAL · Senior Marketing / Security InterestOpen
E-008CLICK JOB-SEC-001 · Senior Marketing / Security InterestOpen
E-009CLICK JOB-SEC-002 · Senior Marketing / Security InterestOpen
E-010CLICK JOB-SEC-003 · Senior Marketing / Security InterestOpen
E-011B — Security Interest REFRESHED · Senior Marketing / Security InterestOpen
E-012B — Security Interest INITIAL vs REFRESHED · Senior Marketing / Security InterestOpen
E-013C — Neutral Control INITIAL · C — Neutral ControlOpen
E-014C — Neutral Control REFRESHED · C — Neutral ControlOpen
E-015C — Neutral Control INITIAL vs REFRESHED · C — Neutral ControlOpen