> Data revision: baf7c2f4307d3a989a8fb9363df8000fda3d9310
> Source files:
- [reproduce/lab-shaping-assumption-guidance-screen-20261108/README.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/lab-shaping-assumption-guidance-screen-20261108/README.md)
- [categories/harness-levers.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/categories/harness-levers.md)
- [site/content/finding-pages.json](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/site/content/finding-pages.json)

# Shaping guidance for assumptions and evidence

- **Finding id:** `lab-shaping-assumption-guidance-screen-20261108`
- **Verdict:** No
- **Finished:** 8 November 2026
- **Comparison:** [Kogen shaping component](/comparisons/kogen-shaping/)
- **Reproduction pack:** [reproduce/lab-shaping-assumption-guidance-screen-20261108/](https://github.com/KogenAI/kogen-bench/tree/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/lab-shaping-assumption-guidance-screen-20261108) (data, analysis script and rerun steps)
- **Category page:** [categories/harness-levers.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/categories/harness-levers.md)

## In plain terms

Assumption and evidence guidance reached 4 of 10 ready, compared with 6 in the stored release, and was dropped.

Verdict key: **Yes**, The result supports the change or claim that was tested. **No**, The result does not support it, so nothing changes. **Unresolved**, The evidence is not strong enough either way. **Superseded**, A later test replaced this result.

## Full text of the reproduction pack

The text below is the pack's README as published in the repository (a system folder name may show without its leading slash, so that the site's privacy scan passes).

On ten small requests, the stored release reached 6 ready Intents and a candidate reached 4. One case changed from not ready to ready and three from ready to not ready. Two unpatched repetitions reached 6 and 7, with two and three paired flips respectively. The candidate was dropped because the score fell and the stored comparison had regressions. No questions were asked. This is a screen, not a qualified delivery change.

All four discordances occurred in shaping; their cause is unclear, because the instructions may have changed the draw and variation is also observed. In two losses the candidate cited tracked files with line ranges and the evidence check refused those paths; the third ended on decision-check refusals of claims about existing behavior. In the gain the stored run exhausted decision checks while the candidate recovered after decision and evidence refusals. The instruction to cite inspected files appeared in the candidate, but the trace does not establish a causal gain. The decision, evidence and readiness gates were unchanged.

The candidate instructed the shaper to mark genuinely unresolved technical choices as assumptions and cite inspected tracked files for existing behavior, while prohibiting invented facts or contradictions as assumptions and retaining concrete acceptance checks. Both arms used the same model, low effort and no answerer; the control was stored. Times include failed attempts. The shared credit balance did not move visibly; attributable spend is unknown. No held-out or complex requests, Build stage or hidden checks ran; landing and hidden outcomes are unmeasured. Small Go and Node workspaces were reconstructed; byte identity with the earlier study is unverified.

The release uses [Kogen source](https://github.com/KogenAI/kogen/tree/47236bbb93385bc960fd2ac353c9b000bd6ed571) and [shaping source](https://github.com/KogenAI/kogen-shaping/tree/0484edf817f6dc74f691f44d5234cf7f6ae0435e). The candidate adds only the instruction described above.

## Rerun

Run `python3 report.py` to reproduce paired counts. To repeat the model screen, use the [small-request setup](/findings/lab-shaping-small-requests-20261107/), pinned release and same ten requests; add the described instruction, keep all gates, and shape each request with the same model, low effort and no answerer. Record readiness, questions, refusal causes and elapsed time.
