Day 8 of 1060 minShell Walker
Day 8: Incident Investigation and Reporting
Taught by Mr Elite Updated
By the end of today you can
- Open a case with intake notes, a scope and a hashed evidence copy
- Build a timeline that separates file evidence from platform claims
- Word findings with a stated confidence level and what the evidence does not show
- Write a short report with a recommended response
Warm-up
Answer from memory first, then tap to check. Recalling yesterday is what makes it stick.
What does a hash mismatch on a C2PA manifest tell you?
The content changed after it was signed. It does not mean the change was malicious.
What does a missing content credential mean?
Very little. Stripping and never signing are both common, so it is a neutral finding.
Someone in finance forwards you a voice message that sounds like the managing director asking for an urgent payment, plus a photo of a signed approval. They want one answer: is it fake? You have run hashes, metadata, image checks, spectrograms and credential checks over the last seven days. Today you stop running tools one at a time and turn them into an investigation someone else can rely on.
The skill here is not a new command. It is the discipline around the commands: what you collect, what you record, how you word a conclusion you cannot prove. A tidy report with honest confidence beats a confident guess every time. Everything today runs on your own machine with files you create.
Investigation is a process, not a tool run
Every media case follows the same shape: intake, scope, preserve, analyse, timeline, conclude, recommend. Skipping the early steps is how analysts end up unable to defend their findings. If you cannot show the file you analysed is the file you were given, the rest does not matter.
Think of the seven stages as two halves. The first three (intake, scope, preserve) protect the evidence and fix the question. They cost ten minutes and cannot be done properly afterwards. The last four (analyse, timeline, conclude, recommend) turn the evidence into something a manager can act on. Most beginners spend nearly all their time on analysis and almost none on the other six stages, and that is where reports fall apart.
- Intake
- Recording who reported what, when, how the media reached you, and what they want to know.
- Scope
- The specific questions you will answer, and the ones you will not. For example: is this audio likely synthetic, not who made it.
- Chain of custody
- A written record of who held the evidence, when, and what was done to it.
- Confidence statement
- A plain claim of how strongly the evidence supports a conclusion, and why.
Scope matters because media analysis answers narrow questions. Tools can say a clip shows synthetic traits. They cannot tell you who sent it or why. Write the questions down before you open a single file, so the report answers what was asked.
Scope for case 08
In scope:
1. Does voice_msg.wav show signs of speech synthesis?
2. Does approval.jpg show signs of editing or carry a content credential?
Out of scope:
- Who created or sent either file
- Whether the payment request itself is legitimateQuick check
A manager asks you to prove who created a suspicious voice message. What is the best scoping response?
Scoping to what the evidence can support keeps the report honest and defensible.
Watch me do it: open a case and preserve evidence
We will build a fictional case. You will create the evidence yourself, so nothing touches a real person or system. The report says a voice message and a photo arrived by chat. I will make stand-in files and set up the case folder the way I would for real. Use a Linux machine or a WSL terminal. Every command below runs only inside the case08 folder on your own computer.
First, check your tools are present. Run the block below. Each line should print a version or a path. If a line says command not found, install that tool before going on: on Debian or Ubuntu, sudo apt install espeak-ng ffmpeg libimage-exiftool-perl python3-pil covers most of it. c2patool comes from its own project releases, and if you skipped it on an earlier day, you can skip that one command later and note the gap in your report.
espeak-ng --version | head -n 1
ffmpeg -version | head -n 1
exiftool -ver
python3 -c "import PIL; print('Pillow', PIL.__version__)"
c2patool --versionmkdir -p case08/{original,working,notes,output} && cd case08
lsespeak-ng -w original/voice_msg.wav "Please approve the transfer today. It is urgent."
python3 -c "from PIL import Image; Image.new('RGB',(640,480),(120,140,160)).save('original/approval.jpg')"
ls -l originalsha256sum original/* | tee notes/hashes.txt
chmod a-w original/*
cp original/* working/
ls -l original workingsha256sum -c notes/hashes.txtecho "$(date -u +%FT%TZ) | received 2 files via chat report | hashed SHA-256 | copies made to working/" >> notes/custody.log
cat notes/custody.logLook at what you now have. The originals are fingerprinted and locked. The copies are free for analysis. The log says what happened and when. If anyone later asks whether you changed the evidence, you can re-run the hash check and show the answer. That is the whole point of this stage.
Building a timeline from evidence
A timeline lines events up in time order and labels where each one came from. The key habit is separating what the file says, what the platform says, and what a person claims. File metadata is editable. A platform delivery time is usually harder to alter. A witness statement is a claim until something supports it.
exiftool -time:all -G1 working/approval.jpg | head
ffprobe -hide_banner working/voice_msg.wav 2>&1 | head -n 12If exiftool prints only a few file system lines such as File Modification Date/Time, that is the correct result for our plain test image. Real phone photos usually show Date/Time Original and similar fields. Their absence on a file claimed to be a camera photo would be worth a line in your findings, though never proof on its own.
| Time (UTC) | Event | Source | Reliability |
|------------------|-------------------------------|----------|-------------|
| 09:12 (reported) | Voice message received in chat| platform | medium |
| 09:14 (reported) | Photo received in chat | platform | medium |
| 09:40 | Finance forwards both to SOC | claim | medium |
| 10:05 | Files hashed, copies made | custody | high |
| unknown | File creation | file | low |Notice the last row. When you do not know a time, write unknown rather than guessing. An honest gap is more useful to a reader than a made-up entry, and it shows where further evidence, such as the platform's own logs, could help.
Quick check
A JPEG's metadata says it was created at 08:00, but the chat says it arrived at 09:14. What do you record?
Record both with their sources and reliability. Contradictions are findings, not problems to hide.
Run the analysis and word your confidence
Now apply your earlier techniques to the working copies. For the image, check metadata, structure and any credentials. For the audio, look at the spectrogram and basic features. Then comes the hard part, which is the wording. Each finding gets a strength, and the overall conclusion gets a confidence level with reasons.
exiftool working/approval.jpg | head -n 20
c2patool working/approval.jpg
ffmpeg -hide_banner -i working/voice_msg.wav -lavfi showspectrumpic=s=1000x400 output/voice_spec.pngSave what you see. Redirect each tool's output into a text file in output/, for example exiftool working/approval.jpg > output/exif.txt, so another analyst can read exactly what you read. The ffmpeg command should finish with a line about writing output/voice_spec.png. If it refuses because the file exists, add -y after ffmpeg to overwrite.
When you open the spectrogram, time runs left to right and pitch runs bottom to top. Natural recordings usually show a speckled background floor, breath and room noise between words. Our espeak-ng clip will show clean bands with an almost empty background. Real clips vary a great deal, which is why this one signal never decides a case alone.
Use a fixed scale so your wording stays consistent. I use three levels, each with a stated basis.
- Low confidence
- Weak or single signals, or signals with common innocent explanations.
- Moderate confidence
- Several independent signals agree, with no strong contradiction.
- High confidence
- Strong, independent evidence agrees, such as a failed signature plus clear artefacts, with alternatives considered.
Finding: voice_msg.wav shows a clean, uniform spectrum with no room noise.
Confidence: moderate that the audio is synthetic.
Basis: flat background and no natural noise floor; consistent with speech synthesis.
Limits: clean recordings and noise removal can look similar. Attribution is not assessed.Every finding you write should have those four parts: the observation, the level, the basis and the limits. If you cannot fill in the limits line, you probably have not thought hard enough about innocent explanations. Go back and ask what else could produce the same result.
Your practice: write the report
No lab today, so you work on your own machine in the case08 folder. A good report is short enough to be read by a busy manager and detailed enough for another analyst to repeat. Create notes/report.md in any text editor, such as nano notes/report.md, and use this structure.
# Case 08 report
## 1. Summary (2-3 sentences, with overall confidence)
## 2. Request and scope
## 3. Evidence and custody (file names, SHA-256, log reference)
## 4. Timeline (table)
## 5. Findings (each with strength, confidence, limits)
## 6. What this does not show
## 7. Recommended responseWrite the summary last, even though it sits first. It should name the question, give the overall confidence and give the main recommendation in plain words. A manager may read only that section, so it must stand alone.
Summary: The voice message shows traits consistent with speech synthesis (moderate confidence). The image carries no content credential, which is neutral. The media alone cannot show who sent either file. We recommend verifying the payment request by phone on a known number before any action.The recommended response should match the evidence. Moderate confidence of synthesis suggests verifying the request through a second channel, such as phoning the director on a known number, and holding the payment. It does not justify accusing anyone. Recommend steps that are cheap, reversible and aimed at the risk.
When you finish, check yourself with three tests. Can another analyst repeat every hash from your notes? Does every recommendation point back to a finding? Does section 6 say plainly that attribution was not assessed? If any answer is no, fix that before you call the report done.
If it did not work
Analysing the original file directly
Hash first, make the original read-only, and work only on copies in working/. Re-run sha256sum on the original if you doubt it.
Treating metadata times as fact
Label them as file-sourced and low reliability. Compare with platform times and note contradictions.
Overstating certainty in the summary
Use your three-level scale and always add a limits line saying what the evidence cannot show.
Recommending action beyond the evidence
Tie each recommendation to a finding. Prefer verification steps over accusations.
Recap
- Open every case with intake, a written scope, hashes and a custody log
- Work on copies and keep originals read-only
- Build timelines in UTC and label each row as file, platform or claim
- State confidence on a fixed scale with the basis and the limits
- Say what the evidence does not show, including attribution
- Match the recommended response to the strength of the evidence
Cheat sheet
sha256sum original/*- Records the evidence hash before analysis
sha256sum -c notes/hashes.txt- Checks files still match the saved hashes
chmod a-w original/*- Makes originals read-only
date -u +%FT%TZ- UTC timestamp for log entries
exiftool -time:all -G1 file- Lists embedded time fields with group names
ffmpeg -i a.wav -lavfi showspectrumpic=s=1000x400 out.png- Makes a spectrogram image of audio
c2patool file- Reads and validates a C2PA manifest
Low / moderate / high- Confidence scale with stated basis
Before the next day
Mission: Voice message and image: full case file
Take the fictional report of a suspicious voice message and approval image and investigate both, using the stand-in files from today or ones you make yourself. Treat it as if another analyst must repeat your work. The aim is a report where every claim traces back to evidence and every conclusion carries honest confidence.
About 40 min
You are done when: A case08 folder containing hashes.txt, custody.log, timeline.md, saved tool outputs and a completed report.md with confidence statements and a recommended response.
Want more? Swap your folder with a peer or reread it a day later and check you can reproduce every hash and finding from the notes alone.
Quiz
4 questions. Your score is saved with your progress.
1. Why do you hash the evidence before any analysis?
The hash anchors the file's identity, so you can show it was unchanged through the investigation.
2. Your analysis shows two weak signals of synthesis, each with common innocent explanations. What confidence fits best?
Weak signals with innocent explanations support only low confidence, and the report should say so.
3. Which timeline entry is the least reliable on its own?
Embedded dates are editable text, so they need corroboration from other sources.
4. Which recommended response best matches moderate confidence that a voice message is synthetic?
Out-of-band verification is cheap, reversible and addresses the risk without overreaching the evidence.
Finish the day
Mark the day complete to save your progress.
Next: Day 9 is the capstone: you investigate an image, a video and an audio clip as one case and deliver a full report. Today's case structure, timeline and confidence wording become your template.
Your notes
Private to this device. Write what you want to remember from today.