SkillCort Public Governance Report

Source manifest:

v1.15.0

Source verified:

11 September 2026

Yesterday this report said enforced withdrawal from a live session was active; today it says verified, and the route there is worth publishing rather than summarising. A candidate stopped a live AI interview four times this evening. The first three times the platform recorded its own failure: the decision was saved, the session ended, nothing was evaluated and no score was written — but the server could not terminate the provider call, so the enforcement record said FAILED and opened a high-severity case about itself. The cause was that the call had been created with the browser's temporary credential, and a provider binds a call to whoever created it, so our own termination requests addressed a session we could not see. The fourth run, after that was corrected: the AI was stopped 623 milliseconds after the candidate asked, nothing arrived from it afterwards, the session's tamper-evident log verified intact, the attempt stayed open and unscored, and a named reviewer was notified. That run is the evidence behind the word verified, and the three failures are kept with it — a record that reports its own failure is what made the cause findable. What remains bounded is unchanged and stated in the entry itself: this control exists in the live AI interview; in AI case simulations and AI-scored assessments the choice is made before the step starts. No control, capture behaviour, consent copy or retention rule changed.

Source manifest:

v1.14.0

Source verified:

11 September 2026

This report said enforced withdrawal from an in-progress live session was deferred to a future implementation. It shipped on 11 September 2026 and is enabled in production: in a live AI interview the candidate can stop the AI, and the platform enforces the stop server-side rather than trusting the browser — the decision is recorded first, the AI is cancelled and the call ended, the session is not evaluated, no score is written, and the case is routed to a named reviewer who is notified in the product. Two things were fixed before the control was turned on, and both are worth stating plainly. The queue those cases route to had existed for two months with no page linking to it, so a routed case was a record nobody could open; it now has one. And the candidate-facing text promised that a person would review what the candidate had done so far and arrange an alternative way to complete the step — neither is something this product does, because a withdrawn session's recording is deliberately never kept and an AI-free alternative is, as this report states elsewhere, not guaranteed by SkillCort. That text now says what actually happens. The boundary that remains is the surface: AI case simulations and AI-scored assessments have no mid-session control, only the choice made before the step starts. Governance limitation L-01 is narrowed accordingly rather than retired, and this report does not claim production verification — no candidate has used the control yet.

Source manifest:

v1.12.0

Source verified:

8 September 2026

The previous entry said a finding was recorded and waited in a queue because the worker that would deliver it was switched off. That worker is now on: it reads every notification source including the daily detection run, and the full chain — a finding recorded, a notification queued, delivered in the product with platform-operator fan-out — was verified in production on 8 September 2026 by a controlled dry run followed by a real run that delivered all four queued notifications without a failure. A daily delivery schedule is configured to run each morning after the detection run; at publication its first scheduled run had not yet fired, and this entry says so rather than rounding up. The Notification worker and Detection alert delivery limitations are closed, and the in-app delivery entry now describes the verified chain. What remains absent is stated on its own terms: no email, SMS or other external channel exists — delivery is in-app only. In the governance manifest (v1.12.0) limitation L-02 is retired in favour of L-16, which records exactly that residue, and R-02 is narrowed onto it. No control, capture behaviour, consent copy or retention rule changed.

Source manifest:

v1.10.0

Source verified:

7 September 2026

Two days ago this report said scheduled automated detection was not active; that was corrected on 6 September to say it ran daily but read-only. On 7 September the daily run was switched to writing, and this report still said it could not write. The same statement has now been wrong in both directions within two days, which is what happens when the automation and the record are changed by different hands. What is true today: the daily run records what it finds and queues a notification for staff, and the worker that would deliver that notification is switched off — so a finding is recorded and waits rather than reaching a person. The entry is renamed to what is actually not active, which is the delivery, not the detection. Governance limitations L-03 and R-03 are retired in manifest v1.10.0 because their content was that findings could not be recorded; the remaining gap is the disabled delivery worker, which L-02 already records and now states explicitly. No control, capture behaviour, consent copy or retention rule changed, and no delivery, email or retention automation was started.

Source manifest:

v1.8.0

Source verified:

6 September 2026

This report said scheduled automated detection was not active. It has run every day at 06:15 UTC since 10 August 2026, when the recurring schedule was approved for read-only runs, and the governance manifest published alongside this report was updated to say so in v1.8.0 — while this page was not. The correction is narrower than "now active", because what runs cannot alert: a scheduled run evaluates every check and records what it counted, but writes no finding and sends no notification, so a signal is seen when the run log is read rather than when it appears. The limitation is therefore narrowed rather than retired, matching governance limitation L-03. Separately, this report described a structured in-product candidate appeal flow as in progress. It is not: no such flow exists, none is being built, and whether SkillCort should provide one is an undecided product question. That was published here, on the Trust Center overview, in the solutions pages and in the exported governance summary, and in the AI manifesto it was stated in the present tense as an existing capability. All of those now say what is true — candidates raise concerns through the organization running the assessment, and no single integrity signal rejects a candidate on its own. No control, capture behaviour, consent copy or retention rule changed.

Source manifest:

v1.7.0

Source verified:

10 August 2026

SkillCort's Terms of Use, Privacy Policy, Data Processing Addendum, Cookie Policy and subprocessor list went into force on 3 August 2026, but this report, the Trust Center overview, the Security & Privacy and AI Governance pages and the pricing page all still described them as in preparation and not yet generally available. That was wrong on every one of those surfaces and is corrected here; legal documents move from a documented limitation to a verified control. Governance risk R-13 is narrowed in manifest v1.7.0 to what is actually still open — a candidate's acceptance is not recorded against the version of the disclosure they were shown, because the disclosure-version machinery has no active production producer (limitation L-12). No control, capture behaviour, consent copy or retention rule changed.

Source manifest:

v1.6.0

Source verified:

5 August 2026

Automated retention execution now covers assessment records (responses, scores, evaluator notes, decisions — 24-month default, organization-configurable), operational logs, and organization offboarding (30-day export window, then deletion from active systems with backup-cycle aging), alongside the existing monitoring-media sweep. Governance limitation L-13 (broad retention execution not active) is retired in manifest v1.6.0 with production evidence; the published privacy policies and Terms of Use state the same schedule.

Source manifest:

v1.5.0

Source verified:

3 August 2026

The third parties behind AI-guided assessment sections are now named to candidates and on this site: OpenAI processes candidate speech (understanding and the interviewer's voice) and powers the AI-assisted proctoring review; HeyGen generates the interviewer's on-screen presenter and receives only the interviewer's lines — never candidate audio or camera. The per-question interviewer clip renderer moved from D-ID to HeyGen; that path renders authored question text only and carries no candidate data. Governance risk R-12 (third parties not yet named) is retired in manifest v1.5.0.

Source manifest:

v1.4.0

Source verified:

2 August 2026

This statement now binds to governance manifest v1.4.0. The manifest records that the human-adjudicated AI review surfaces (webcam, screen, audio, ID name-check) have each been exercised by reviewers on real production sessions, with the runs held in the provenance ledger. No public control, limitation or consent statement changes — the source binding and verification date are refreshed.

Source manifest:

v1.3.0

Source verified:

26 July 2026

The human-adjudicated AI review now also transcribes a session's microphone recording (when the organization enabled mic capture and the candidate consented) and flags speech segments worth listening to — the question read aloud, dialogue or answer-dictation patterns. Speaker identity is never claimed, the reviewer always listens to the actual audio, every flag receives an explicit human verdict, and notes never quote personal content. The candidate consent screen names AI-assisted transcription in lockstep.

Source manifest:

v1.3.0

Source verified:

26 July 2026

The human-adjudicated AI review now also scans a session's screen captures — flagging visible AI-assistant windows, search engines, messaging apps, remote-control tools and unrelated windows. Same guarantees as the webcam review: runs only in reviewer context, never decides, every flag human-adjudicated; AI notes never quote personal message content.

Source manifest:

v1.3.0

Source verified:

26 July 2026

The AI webcam scan now starts automatically when a reviewer opens a session (previously click-to-run). It still never runs unattended and never decides: each flagged moment receives an explicit, audited human verdict — confirmed or dismissed.

Source manifest:

v1.3.0

Source verified:

26 July 2026

Discloses the org-provided identity reference photo: candidate-notified on the consent screen, human-reviewer comparison only, never fed to automated matching, and removable as the current reference at any time; review-pinned versions follow the configured media-retention window.

Source manifest:

v1.3.0

Source verified:

21 July 2026

Names microphone capture (organization-enabled, candidate-consented) alongside camera and screen, and discloses staff-triggered, human-in-the-loop AI review of webcam images that never makes an automated decision.

Source manifest:

v1.1.0

Source verified:

20 July 2026

Public publication of verified governance controls and documented limitations.