> Data revision: baf7c2f4307d3a989a8fb9363df8000fda3d9310
> Source files:
- [reproduce/lab-c6-consumer-overcoupling-allowlist-20261105/README.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/lab-c6-consumer-overcoupling-allowlist-20261105/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)

# An allow-list for the budget-off consumer test: not measured

- **Finding id:** `lab-c6-consumer-overcoupling-allowlist-20261105`
- **Verdict:** Unresolved
- **Finished:** 5 November 2026
- **Comparison:** [Kogen build component](/comparisons/kogen-build/)
- **Reproduction pack:** [reproduce/lab-c6-consumer-overcoupling-allowlist-20261105/](https://github.com/KogenAI/kogen-bench/tree/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/lab-c6-consumer-overcoupling-allowlist-20261105) (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

The measured comparison of the budget-off test against an allow-list projection did not run: none of 260 checks of the two allowed Linux hosts found them quiet, so there is no false-red rate and no mutation result. A host-free replay of the four keys behind the recorded reds shows all four were red under the deny-list of their own window and one still is today, while the allow-list compares none of them and keeps any budget field. That is a design check, not a measurement, so the question stays open.

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 **Superseded**, A later test replaced this result.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).

**Finding.** The measured comparison (the budget-off test as written against an allow-list projection, 32 producer trees, quiet and loaded, plus a mutation control) did not run: neither Linux host allowed to run it was ever quiet. In 260 checks of the two hosts between 11:06 and 12:42 UTC on 5 Nov 2026, not one found a host free of delivery Builds and other Lab runs with a load average under 1. The first host was held by another Lab until about noon and then ran a delivery Build; the second was released by the build-speed Lab at 11:05 (which never ran its own loaded arms either) but was then in use by a third Lab and delivery Builds at a load average near 18. A key-level replay of the recorded red diffs, which needs no host, shows what the measured run would have tested: all 4 differing keys behind the reported reds were red under the deny-list of their own window's test version, one of the 4 (a toolchain-probe status) is still red under today's deny-list, and none of the 4 is compared by the allow-list; the allow-list keeps any key that mentions a budget, so the planted-budget mutation key stays visible. This is a design check, not a measured false-red rate: the verdict is unresolved.

## Question
Does comparing an allow-list projection (the requests, the order and role of events, the checkpoint keys, the result outcome, the commit message, and every budget field) in `TestBuildRequestBudgetOff`, instead of the whole record minus a growing deny-list of volatile keys, remove its false reds without losing real ones? Registered rule (Studio Lab lever `lab-c6-consumer-overcoupling-allowlist-20261105`): 0 false reds for the allow-list across the trees and load levels, and the planted budget field caught in every run. Assumption `build-failures-c6-consumer-overcoupling-20261105`.

## What ran
- **Static replay (done, model-free, no host):** the four keys that differed in the saved red land logs and the classified table (`request_offset_ms`, `load_average`, step-sweep `ms`, toolchain-probe `status`) were checked against the deny-list of each kogen-build version of the test that was live in that window (V0 initial 2 Nov, V1 772102a, V2 10f190b, V3 edb7e91, head 956b151) and against the allow-list. `data/deny-lists.json` holds each version's volatile keys and suffix rules, with the link to the public file at that commit.
- **Measured part (prepared, not run):** `harness/` holds everything: 12 red-window trees (consumer = kogen-build main at the window time, producer = the window's candidate where a copy still exists, 3 of 12, otherwise the producer main it was landing onto) and 20 green producer mains (`trees.tsv`), the allow-list test `budget_allow_test.go` built into the same test binary as the tree's own test, a one-line mutation of a copy, the loaded level (Lab E2's four-stream replay, reused), and `analyze_c6.py` for the tables. Design size 32 trees x 2 arms x 2 load levels x 3 reps = 384 probes plus 12 for the mutation.

## Results
| check | result |
|---|---|
| recorded differing keys behind the reds | 4 (`request_offset_ms`, `load_average`, `ms`, `status`) |
| red under the deny-list of their own window's version | 4 of 4 |
| still red under today's deny-list (956b151) | 1 of 4 (`status`) |
| compared (so able to turn red) by the allow-list | 0 of 4 |
| versions of the test that had to change the mask during the window | 2 (V1, V3) |
| planted budget key kept by the allow-list | yes |
| measured probes (arms x trees x levels) | 0 of 396 designed; no false-red rate, no mutation result |
| guard checks of the two Linux hosts for a quiet moment | 260, quiet 0 |

- The red windows are 12 of the 14 reported; the other 2 are not identifiable from the saved land logs. For 9 of the 12, the candidate commit no longer exists in any host mirror, so those trees would have been base-only proxies.
- The static replay says nothing about false reds the deny-list could still produce on keys not yet seen, nor about whether the allow-list misses a real defect outside its fields; only the measured run and its mutation control test that.

## What this decides
Unresolved. No request is filed for the budget-off test: the allow-list is not proven, and kogen-build already masks 3 of the 4 keys. The other consumer-test pins named in the same analysis were listed without measurement; one, the scripted-executor fixture that fails when a producer adds a correction round, is a plain fixture change a request can make without a Lab. Run the measured part on a quiet Linux host (about 2 to 3 hours of host time) with the steps below.

## Rerun
```
cd reproduce/lab-c6-consumer-overcoupling-allowlist-20261105
python3 -I static_replay.py
```
This prints the four rows above and a JSON line with `keys 4, red_under_window_version 4, still_red_at_head 1, red_under_allow_list 0, mutation_key_kept_by_allow_list true`. The measured part: on a quiet Linux host with Go 1.27.2 and local mirrors of the public KogenAI repositories, as the lanes user, copy `harness/` and the Lab E2 files (`e2.py`, `sched.json`, `heavy-lab.sh` from its published pack) into one directory, then `bash driver.sh 25 12 20` (setup, quiet and load minutes) and `python3 -I analyze_c6.py <work>/out/probes.jsonl <work>/out/setup.tsv`. The test source lives in the public [kogen-build](https://github.com/KogenAI/kogen-build/blob/956b151/build/build_request_budget_test.go) repository.
