Optional paired challenge / Evidence cabinet

The Archive Vault

Recover the governors' evidence briefing from three connected locks.

This is extra, not a replacement. The archive's briefing cabinet has sealed its evidence record. Use what you learned in the five-stage AI Archive investigation to restore three evidence seals and open the cabinet. This is an extra challenge, not a substitute for your case-file notes. Work in pairs: one investigator reads a clue, the other connects it to the idea. Swap roles at each lock. No written response, timer, or computing lesson is needed.

Cabinet status0 of 3 locks opened

01The mixed-up methods

02The demonstration trap

03The corrupted citation board

Seal 01 / How it works

The mixed-up methods

Rebuild the methods panel, then connect each system to the clue that identifies it.

The idea behind the lock

General AI clue: three ways to build a system

Written IF/THEN rules handle situations a designer planned for. A system trained on examples finds patterns in sample data. Generative AI uses learned patterns to create new text. These descriptions explain how the examples work; they do not tell you which system is safe.

Source evidence / Read before deciding

Archive file 2 / Technician's test log

The help desk handled all 12 password requests but sent four offline-work requests to an irrelevant default page.

Archive file 4 / Training examples audit

The filter used labelled messages to learn patterns; its training examples included few genuine newsletters.

Archive file 5 / ORACLE fact box

A generated answer combined catalogue facts with a confident claim about a study that the catalogue did not confirm.

Open the original investigation: The AI Archive

Arrange these methods from written instructions to learned patterns to generated text.

The initial order is not a clue. Arrange the items using the move controls; keyboard users can tab to each control.

Make a new paragraph from learned patterns
Follow written IF/THEN instructions
Find patterns in labelled example messages

Connect each archive clue to its method.

Password requests succeed; an unexpected request reaches a default page. Newsletter mistakes may relate to limited training examples. A new paragraph confidently cites a study that was not verified.

Open a small hint

Check which system follows written instructions, which uses example messages, and which writes new text.

Lock opened / token recovered

The methods seal is restored. Knowing how a system works is only a starting point.

Seal 02 / Test the claim

The demonstration trap

Connect three test results to the conclusion each result actually supports.

The idea behind the lock

General AI clue: tests have limits

A good result on selected examples does not guarantee the same result on unfamiliar ones. A test also measures only what its observers checked. Look for both benefits and mistakes, not just the best-looking number.

Source evidence / Read before deciding

Archive file 1 / Test team's results and conversation extract

Four of six testers thought the hidden speaker was a person. They did not check its statement about library opening hours.

Archive file 3 / Atlas launch and independent test

Atlas translated 19 of 20 chosen sentences accurately at launch, but only 31 of 100 unfamiliar sentences six months later.

Archive file 4 / Trial results

Filter Two caught 17 of 20 spam messages instead of 12, but wrongly blocked five genuine messages instead of one.

Open the original investigation: The AI Archive

Match each trial with its fair conclusion.

Four of six testers thought the speaker was a person. A strong demonstration was followed by a weak unfamiliar-sentence test. More spam was caught, but more genuine messages were blocked.

Open a small hint

Do not treat a human-sounding reply as a fact check, or a chosen demonstration as a test on unfamiliar work.

Lock opened / token recovered

The trial seal is restored. A result is useful only when you know what was tested.

Seal 03 / Trace the source

The corrupted citation board

Separate confirmed details, contradicted claims and claims that have not been checked.

The idea behind the lock

General AI clue: confidence is not a source

AI-generated text can mix checked facts with unsupported claims. 'Not verified' does not automatically mean false; 'contradicted' means the case includes evidence against the claim. Both are different from a fact confirmed by a named source.

Source evidence / Read before deciding

Archive file 5 / Archive catalogue search

The catalogue confirms the date 1912 and a count of 840 maps. It does not confirm the generated answer's claimed study.

Archive file 5 / Archivist's advice

Some maps showed roads that were planned, not roads that had been completed.

Archive file 1 / Conversation extract

The speaker gave a confident library opening-hours answer, but nobody checked whether it was true.

Open the original investigation: The AI Archive

File each statement under the right evidence status.

The catalogue records 840 maps and the date 1912. The cited study establishes that every map is accurate. Every road shown on the maps had already been built. The chatbot's library opening-hours answer was correct.

Open a small hint

The archivist describes planned roads. The catalogue checks the date and count, not the claimed study.

Lock opened / token recovered

The source seal is restored. The cabinet can now distinguish evidence from confidence.

Final cabinet / All three tokens required

Open the governors' briefing

Each recovered slip contains a different case detail. Match the slips to the governor's questions below, then choose the only conclusion all three checks support. The order is not the order you opened the locks.

The slips you recovered

1. Which slip warns that a fluent fact box can include an unverified study?

2. Which slip compares a chosen launch demonstration with unfamiliar sentences?

3. Which slip reveals what happens when a request is missing from written instructions?

Which briefing claim fits all three slips?

This is an evidence check. Your separate verdict below is not graded.

Routine success, a strong demonstration and fluent wording are not enough on their own to prove reliability. A fluent answer verifies its own sources if a catalogue number in it is correct. One weak test proves that no AI system can ever be useful.

Choose your pair's verdict

Your view is yours to make. The verdict is not marked for correctness; the token sequence opens the briefing.

Hold any proposed use while important claims are unchecked Consider limited uses with testing and human correction Recommend against the current proposal

Project ORACLE / The Archive Vault

Year 9 paired extra / CodeBash / Not yet completed

Objective: The archive's briefing cabinet has sealed its evidence record. Use what you learned in the five-stage AI Archive investigation to restore three evidence seals and open the cabinet. This is an extra challenge, not a substitute for your case-file notes.

Lock 01 / The mixed-up methods

Concept: General AI clue: three ways to build a system. Written IF/THEN rules handle situations a designer planned for. A system trained on examples finds patterns in sample data. Generative AI uses learned patterns to create new text. These descriptions explain how the examples work; they do not tell you which system is safe.

Source evidence

Archive file 2 / Technician's test log
The help desk handled all 12 password requests but sent four offline-work requests to an irrelevant default page.

Archive file 4 / Training examples audit
The filter used labelled messages to learn patterns; its training examples included few genuine newsletters.

Archive file 5 / ORACLE fact box
A generated answer combined catalogue facts with a confident claim about a study that the catalogue did not confirm.

Token: Not recovered

Lock 02 / The demonstration trap

Concept: General AI clue: tests have limits. A good result on selected examples does not guarantee the same result on unfamiliar ones. A test also measures only what its observers checked. Look for both benefits and mistakes, not just the best-looking number.

Source evidence

Archive file 1 / Test team's results and conversation extract
Four of six testers thought the hidden speaker was a person. They did not check its statement about library opening hours.

Archive file 3 / Atlas launch and independent test
Atlas translated 19 of 20 chosen sentences accurately at launch, but only 31 of 100 unfamiliar sentences six months later.

Archive file 4 / Trial results
Filter Two caught 17 of 20 spam messages instead of 12, but wrongly blocked five genuine messages instead of one.

Token: Not recovered

Lock 03 / The corrupted citation board

Concept: General AI clue: confidence is not a source. AI-generated text can mix checked facts with unsupported claims. 'Not verified' does not automatically mean false; 'contradicted' means the case includes evidence against the claim. Both are different from a fact confirmed by a named source.

Source evidence

Archive file 5 / Archive catalogue search
The catalogue confirms the date 1912 and a count of 840 maps. It does not confirm the generated answer's claimed study.

Archive file 5 / Archivist's advice
Some maps showed roads that were planned, not roads that had been completed.

Archive file 1 / Conversation extract
The speaker gave a confident library opening-hours answer, but nobody checked whether it was true.

Token: Not recovered

Final briefing

Completion: The briefing has not been recovered.

Our token sequence: Not submitted

Evidence conclusion: Not submitted

Our verdict: Not submitted