← kondi home
examplesFour pipelines, proven end-to-end
Run with real models, every step's output inspected and verdicted. Each dossier shows the step table (what it does + expected output), excerpts from the actual run, and what the verification process caught — the failures each example surfaced became engine fixes. The receipts ship in the repo.
01
Market Brief
VERIFIED · 3/3 STEPS
From a product README on disk to a committed positioning and a one-page market brief. A CLI worker with read tools explores the actual repo; an analysis council commits to exactly one opportunity; a writing council turns the decision into the deliverable.
- enrich
- analysis
- council
- input: file
- models: NIM nemotron ×2 + claude CLI
step 2 · the council commits — no hedging
“DECISION: Commit to building the Cost Management Dashboard as the primary market opportunity … RATIONALE: … addresses the most urgent and universal user pain point: accidental overspend during AI experimentation … REJECTED: General-purpose AI productivity tool … Developer-centric LLM IDE …”
step 3 · the brief, grounded in that decision
“Stop AI Spend Surprises, Start Confident Experimentation — Real-time tracking & alerts · Zero-architecture lift (built on the existing ledger that already token-tracks every LLM call) · can automatically block LLM calls at budget limits.”
what verification caught
The final brief once contradicted the upstream decision — invented product, invented features. Workers only saw the manager's directive, two summarization hops from the real input. Kondi now re-attaches the original source material to every worker prompt. The bug was found because we read the output instead of trusting “completed”.
02
Code Factory
VERIFIED · 7/7 STEPS
Ship a small working utility with tests: plan, human gate, real implementation by the claude CLI, mechanical test verification, a retry loop, a documentation review council, and release notes. The script step actually runs the tests — a model saying “tests pass” counts for nothing.
- code_planning
- gate
- coding
- script
- condition · loop
- review
- agent
- input: text spec
step 4 · the machine's verdict, not the model's — re-run by hand afterwards, same result
ALL TESTS PASS
SCRIPT-VERIFIED
step 3 · the coding council debugged for real before handing off
“I found the issue! The test.js file is using a relative path 'wordfreq.js' which resolves relative to where the test runner is executed from … I need to fix the path to use __dirname … The fix works from any location.” — the worker, mid-debug-cycle
step 6 · the review is evidence, not vibes
“Tokenise on whitespace after removing all non-alphanumeric characters — Satisfied? Yes — Evidence: line 59: text.toLowerCase().replace(/[^a-z0-9\s]/g, ' ') …” — review.md, a per-requirement adherence table with line numbers, written to disk alongside README.md and four docs/ files
what verification caught
An earlier run failed the exact-string gate, looped, and failed again — the retry's feedback was just the test output, which read as success (✓ All tests passed!), so it fixed the wrong things. Loop feedback now leads with the condition's expression and verdict; the next run passed the gate on the first attempt. An agent once claimed it wrote a file that didn't exist — text-output agents are no longer offered write tools, so the deliverable lands in the artifact where the pipeline can check it.
03
Policy Debate
VERIFIED · 3/3 STEPS
A genuinely adversarial council: a chair, an advocate, and a skeptic debate a real question over two rounds — should a 40-person software company adopt a 4-day work week? — then a guard verifies a recommendation exists before an agent distills the executive summary.
- council · 4 personas · 2 rounds
- condition
- agent
- input: text question
- models: NIM nemotron super + nano
the debate was real — positions opposed, dissent recorded
“Consultant A argued … structured expert communication reduces dominance bias … Consultant B cautioned that its iterative, time-intensive nature makes it unsuitable for urgent decisions …” — opposing consultants, round 2, before the chair's decision
what verification caught
Conditions and gates used to replace the data flowing through them with bookkeeping text (“Condition evaluated: TRUE”), so the summary step summarized the bookkeeping. They are now pass-through nodes: upstream content flows on, the evaluation rides in metadata.
04
Content Refinery
VERIFIED · 6/6 STEPS · 1 REAL LOOP-BACK
A live URL in, a length-controlled explainer out. An agent extracts strict JSON from a fetched Wikipedia summary; two analysis lenses run in parallel on individual JSON fields; a council synthesizes; a script counts words; a condition loops the synthesis back — with the measured count as feedback — until the length lands.
- agent · json
- analysis ×2 · parallel
- council
- script
- condition · loop
- input: url · live fetch
step 1 · strict json — and factually right beyond the fetched page
{ "topic": "Delphi method",
"origin": "Developed in the 1950s by Olaf Helmer, Norman Dalkey,
and Nicholas Rescher at the RAND Corporation …", … }the loop fired for real
[log] ↻ Looping back to "Stage 3" (iteration via Length gate)
LENGTH OK (446 words)
what verification caught
Models offered tools they can't run will roleplay them — the extractor once emitted a fake fetch call as its “output”. Tool advertising is now matched to what a persona can actually execute.
Every example ships in the repo under examples/ — full pipeline.json (importable via Pipelines → Import), every step's real output, and an honest per-step VERIFICATION.md. The failures each example surfaced became engine fixes — the examples are the test suite that hardened the product.
Browse the verification dossiers →
← back to kondi home
