FACT
Observed signals
The mirror points to an unapproved address for eleven minutes. Release 0.8.4 has a branch and package digest that do not match the approval.
Fictional defensive campaign · 12 stages · 4 acts
A wrong mirror, a normal automation identity and a mismatched package record meet in the same incident window. Join mail, Git, TLS, DNS, visual, HUMINT and telemetry evidence into one technical timeline.

Running systems
Evidence workspace
Loads the case files and keeps source paths visible.
Local progress
Keeps stage and notebook progress on this device.
Evidence scoring
Checks whether the minimum evidence record is ready.
Server validation
Checks flags with the server-side HMAC flow.
Report output
Builds a structured assessment from the analyst record.
Initial assessment
18 JUL 2026 · 04:17 UTC · GH-26-07
FACT
The mirror points to an unapproved address for eleven minutes. Release 0.8.4 has a branch and package digest that do not match the approval.
CLAIM
The cropped CCTV frame shows North Pier but no face. The social post is an accusation, not identity evidence.
OUTPUT
An auditable incident report that explains the technical path, evidence limits and controls that must change before service resumes.
ACT 1STAGE 01-04
Difficulty: Beginner
A believable handover message lands while the mirror briefly points somewhere it should not. Get the first technical facts down before the public narrative forms.
Follow a convincing handover message through its raw headers and find the point where trust breaks.
Header reconstruction
Email & Identity
Pull a removed handover detail back out of repository history instead of searching only the current files.
Deleted Git history
Repository Forensics
Tie a certificate observation to the mirror before the public DNS change shows up anywhere.
TLS and passive DNS correlation
Infrastructure OSINT
Measure the real exposure window by lining the DNS collectors up against the isolated redirect captures.
DNS and redirect timeline
Web Infrastructure
ACT 2STAGE 05-07
Difficulty: Intermediate
A cropped image and three conflicting voices pull the review toward one person. Check the time and the place. Protect the sources. Keep early access apart from copycat noise.
Normalize one CCTV frame using what is visible in it, the camera inventory, the clock drift and the weather record.
Chronolocation and metadata
Visual Verification
Assess three fictional sources without mixing up trust in a source with trust in a claim.
Source reliability grading
HUMINT Assessment
Separate an early technical cluster from a later copycat without trying to identify a real operator.
Provenance-led clustering
Shadow Intelligence
ACT 3STAGE 08-10
Difficulty: Advanced
The normal automation identity publishes an abnormal artifact. Follow the sessions, refs, provenance, permissions and timing without treating a service account as a person.
Test a public accusation against account, badge, camera and service records without claiming an alibi you can't prove.
Account and physical-evidence separation
Identity Correlation
Compare approval, branch history, workflow permissions and provenance to find where the release source stops matching.
Workflow provenance
Software Supply Chain
Find the machine-like cadence in harmless synthetic telemetry and stop short of a malware or exfiltration claim.
Interval and limitation analysis
Network Telemetry
ACT 4STAGE 11-12
Difficulty: Advanced
Test the competing explanations. Define what the evidence supports and what it does not. Then deliver an incident assessment the board can act on.
Test four explanations with inconsistency rather than by counting supporting evidence.
Analysis of competing hypotheses
Intelligence Assessment
Deliver the final timeline, the scope, the attribution boundary, the containment plan and what stays uncertain.
Evidence-led incident reporting
Final Synthesis
Flags aren't stored in the public client. The later stages need full Git history, a deleted blob, a separate branch, a release tag, file hashes and agreement on timing across records. No honest CTF can promise that AI will never solve it. The goal here is narrower. A campaign you can't guess from the prompt and can only finish with real evidence work.