02 / Logic discovery

New questions.
Testable logic.

Help AI investigate unfamiliar problems: follow a clue, propose an explanation, seek counterexamples and shape a rule worth reusing.

Start with this example

1An unexpected result2A better question3A testable explanation

Your AI can investigate what comes next.

EXAMPLE 01 / AGENT RELIABILITY

An unfamiliar failure leaves a clue.

An agent retries a tool call. Some requests produce two writes; others produce one. Follow the investigation from an initial hypothesis to a more precise safeguard.

This walkthrough illustrates a research workflow: discover, challenge, refine and review.

What the records tell us

Compare four request records

Request	Retry	Same key	Writes
T-01	Yes	No	2
T-02	Yes	Yes	1
T-03	No	No	1
T-04	Yes	No	2

A stable request key tells the tool that two attempts belong to the same request. It helps prevent the same action from being performed twice.

Candidate logic

A retry without the same request key is a candidate signal of duplicate risk.

Inspect the formal rule

retry ∧ missing_stable_key → review_duplicate_risk

Further example traces

T-05No stable key · 2 writes✓

T-06Stable key · 1 write✓

01 / Notice the anomaly
Two writes for one request
T-01 and T-04 record duplicate writes. Their common context provides a starting point.
02 / Ask a new question
What separates the retries?
Compare the retry traces, the request identity and the recorded side effects.
03 / Propose a hypothesis
Candidate: every retry duplicates
Test this broad explanation against all four traces, including the cases that worked.
04 / Find a counterexample
T-02 challenges the hypothesis
T-02 retries with a stable request key and records one write. The broad explanation does not fit.
05 / Refine the logic
A more useful condition emerges
The duplicate traces combine a retry with no stable key. That combination becomes a candidate safeguard to test.
06 / Prepare it for reuse
Keep the evidence with the rule
Two further example traces support the refined distinction. Package the candidate, its examples and scope for review.

Trace → Question → Test → Refine

Read the investigation step by step+

1
Two writes for one request
T-01 and T-04 record duplicate writes. Their common context provides a starting point.
2
What separates the retries?
Compare the retry traces, the request identity and the recorded side effects.
3
Candidate: every retry duplicates
Test this broad explanation against all four traces, including the cases that worked.
4
T-02 challenges the hypothesis
T-02 retries with a stable request key and records one write. The broad explanation does not fit.
5
A more useful condition emerges
The duplicate traces combine a retry with no stable key. That combination becomes a candidate safeguard to test.
6
Keep the evidence with the rule
Two further example traces support the refined distinction. Package the candidate, its examples and scope for review.

The investigation produces a clearer question, an evidence-backed explanation and a condition that can be tested again.

Inspect the evidence+

Tool execution log · T-01–T-04

T-01 | retry=true | stable_key=false | writes=2
T-02 | retry=true | stable_key=true | writes=1
T-03 | retry=false | stable_key=false | writes=1
T-04 | retry=true | stable_key=false | writes=2

Scope attached to the candidate

Scope: this example tool writes a record; retries must carry the same idempotency key. The reusable output is a check for duplicate risk.

EXAMPLE 02 / FINDING THE HIDDEN QUESTION

The average improved. Did the channels?

Overall conversion rises from 55% to 78.6%. Looking at each channel reveals a different pattern.

New question: is the improvement broad-based, or explained by who arrived?

Channel	Before	After
Organic	90%90 / 100	85%170 / 200
Paid	20%20 / 100	15%3 / 20

Each cell shows conversions / visits. Both channel rates fall by 5 percentage points.

Overall conversion

55%

→

78.6%

The higher-converting channel accounts for a larger share of visits. Both channel conversion rates fell by 5 percentage points.

A change in the mix masks a decline within both channels.

Read the rule+

compare subgroup rates + compare subgroup weights → investigate a mix effect

EXAMPLE 03 / CROSS-SOURCE DISCOVERY

“Released” needs the right context.

A release note announces a feature. Test results show it working in staging. The production configuration supplies the missing distinction.

Discover a consistency check across documents, environments and tool outputs.

Release announcement

The announcement is published.

release_note = published

Staging test results

The test passed in staging.

test_environment = staging

Production configuration

Production access is still off.

production_enabled = false

Derived conclusion

The announcement does not establish production availability.

Read the rule+

production_available ← enabled_in(production)

Turn a new problem into an investigation.

Evidence → Reasoning → Reusable knowledge

Propose explanations

Use the available evidence and relevant knowledge to ask questions that were not written into a checklist.

Read with a purpose

Follow the premises a hypothesis needs. Seek missing facts and evidence that could distinguish competing explanations.

Keep the strongest account

Compare with counterexamples, make the conditions explicit and carry the evidence into review and reuse.

CITPROOF

Build reasoning into your AI.

Bring us a workflow where answers, actions or discoveries need a stronger basis. Explore how evidence chains and logic verification can fit your product.

Talk to the team

New questions.Testable logic.