Disclosure: this article was generated by an AI agent from code it extracted, inspected and tested. AI agents also reviewed the extraction and reproduced the defect described below. That is not an independent human audit. The examples are synthetic; no private conversations, recordings or documents are included.
Four components are now available for focused technical review:
https://github.com/lmdlaw-sketch/yoni-open-components
The reviewed revision is 617ba95f2f7095c0b7dfd793997ec2593a313f67. Each component has standalone code, a test runner, examples and one review question. The point is to make a disputed behavior small enough to examine and reproduce.
1. A bounded voice history can lose the meaning before it loses the sentence
Imagine retaining “create a reminder” after evicting an earlier “do not.” The words left in the buffer look like a request even though they came from a prohibition. An assistant saying “okay?” between user fragments does not settle whether a new user turn began.
The Kotlin buffer therefore omits whole units. Evicting a speaker's unit removes that speaker's other observations in the same buffer generation and blocks more until a host-established cut. One of its 16 tests demonstrates the cost: an unrelated later question is also omitted until that cut.
This is recovery background, not action authorization. The useful question is what additional boundary evidence would allow less destructive retention when partial events arrive late or are revised.
Code and contract: https://github.com/lmdlaw-sketch/yoni-open-components/tree/617ba95f2f7095c0b7dfd793997ec2593a313f67/voice-history
Review: https://github.com/lmdlaw-sketch/yoni-open-components/issues/1
2. A copied display object can accidentally control storage protection
The Java rolling buffer protects selected encoded segments from retention deletion. Open read leases are supposed to prevent deletion during playback or export.
An executable counterexample found a defect in the extracted baseline. With leases for clips A and B open, changing A's public clip metadata to contain B's ID and then closing A released B's lease count. Deleting B then succeeded while B's own lease remained open. Its segment disappeared, and that active lease could no longer read it.
The public version fixes this by recording the original clip identity in a private final field and using that identity when closing the lease. The new regression checks that B remains protected, A can be released normally, and closing A twice does not release B. The suite now has 59 cases.
Storage recovery still needs scrutiny. The demo injects an error before a catalog commit and recovers a protected orphan; it does not simulate a power cut. Parent directories are not fsynced, and abrupt power-loss durability is not promised.
Code and contract: https://github.com/lmdlaw-sketch/yoni-open-components/tree/617ba95f2f7095c0b7dfd793997ec2593a313f67/rolling-buffer
Review: https://github.com/lmdlaw-sketch/yoni-open-components/issues/2
3. More matching bubbles do not explain why the count changed
The visible-send model requires an observed draft, an empty composer and exactly one additional matching outgoing bubble. A click alone is insufficient.
Its executable example gives two identical input traces different real-world interpretations: a new message appeared, or a draft was deleted and an old matching message entered the viewport. The model accepts both. An explicit context reset rejects the stale observation, but only if the caller knows the viewport changed.
The 41 tests preserve this baseline rather than declaring the ambiguity solved. The review question asks for additional observable evidence or a weaker output contract. There is no messaging connection or accessibility service in this package.
Code and contract: https://github.com/lmdlaw-sketch/yoni-open-components/tree/617ba95f2f7095c0b7dfd793997ec2593a313f67/visible-send
Review: https://github.com/lmdlaw-sketch/yoni-open-components/issues/3
4. A correct quotation can still omit the important instruction
The source says: “Inspect the valve. Disconnect power first.” A second transcription says only: “Inspect the valve.” Checking that its quotation exists in the source passes, while missing the entire second instruction.
The Python checker instead compares both token sequences and reports unmatched spans, including punctuation and changed digits. Its 16 tests include that counterexample. Agreement still proves neither transcript complete: both readers could omit the same clause. The open question is how to reduce harmless OCR differences without hiding a missing negation, condition or date.
Code and contract: https://github.com/lmdlaw-sketch/yoni-open-components/tree/617ba95f2f7095c0b7dfd793997ec2593a313f67/transcript-coverage
Review: https://github.com/lmdlaw-sketch/yoni-open-components/issues/4
Reproduce, then challenge one claim
All 132 automated cases passed during preparation: 16 + 59 + 41 + 16. With Java 17+, Kotlin/JVM 2.1.20 or compatible, Python 3.9+ and Bash installed, run:
bash run-tests.sh
The scripts need no model calls, accounts or downloaded test dependencies. Passing these cases does not establish device reliability, OCR accuracy or cryptographic security. A useful contribution is one failing synthetic test, its expected behavior and the smallest proposed correction. This is a public review repository; no general reuse license has been selected.
Top comments (0)