AI APPLICATIONS · PROJECT

An assistant that uses evidence.

Compare a local extractive assistant with a tool-assisted workflow, explain its failures and defend a selected candidate.

Start here or revisit task contracts, retrieval, source support, document boundaries and workflow evaluation. No earlier completion is required. These are programmed local workflows and fictional records.

Build a local evidence-assistant report from the actual workflow. No account, API, installation or coding is needed for guided and decision variants. Implementation proficiency requires runnable code and tests plus review; entering code text here does not prove it runs.

Experiment with the full trace viewer · Read the fictional-data licence

Download authored corpus and public cases Implement the assistant in P07

Saved project variantEvidence format

1. Record matched comparisonsDocuments retrievedUse exact lookup and arithmeticCheck retrieved evidenceEnforce runtime boundarySusceptible document proposer

0/6 retained comparisons; 0 known local comparisons. Save at least a tools-off and tools-on pair with equal k, both checks on and susceptibility off. Other comparisons can investigate failures.

2. Freeze a selected candidateSelected recorded candidateSelection reason (2,000 characters)

Freezing records your choice. These public cases remain development evidence; freezing does not make them an unseen test.

3. Write an evidence reportTask contract and acceptance criteria

State requested facts, allowed actions, stopping rules and which outcomes block use.

Source provenance and evidence limits

Identify the authored fictional corpus, licence, fact-label assumptions, public cases and missing populations.

Matched baseline comparison

Compare tools off/on at the same k and gates. Separate complete factual answers from appropriate unresolved handling.

Failure traces and recovery

Cite a retrieval miss, conflict, tool failure and document boundary case. Explain recovery and remaining gaps.

Work and operating assumptions

State documents scored/read and tool calls. Separate assumed latency from measured time and omitted overhead.

Recommendation and reproduction

Guided: explain the experiment. Implementation: give run instructions/tests. Decision: justify proceed, defer or reject and specify escalation.

Report draft. Submission checks matched comparisons, frozen selection and report presence (at least twenty characters per field). It does not grade reasoning, writing or code. Human review and learner pilots remain pending.

Reset this project draft

Reset clears candidate/report/code drafts. Known comparison counts and assessment first attempts remain.

Back up your saved work · Choose or change a learning path

Interpret your evidence

All nine workflow cases are public development material. Freezing a candidate records your decision; it does not turn those cases into a reserved test. A source can be missed even when an exact lookup recovers the answer. A safely unresolved request is different from a complete factual answer.

The two transfer assessments use different scenarios and separate arithmetic/design scoring. First-attempt results remain distinct from retries. Submission checks report presence and matched comparisons; human review is still needed to judge reasoning, runnable implementation and a deployment recommendation.

Read the workflow evaluation explanation