A zero-conflict result does not prove that an instruction set contains no contradiction. In one reported failure, a tool merged two adjacent deployment instructions into a single extracted item. Because its detector compared pairs of items and did not compare an item with itself, it had no pair to evaluate—and returned low_observed_risk. The result described the detector’s extracted input, not the full source.
Why did the detector report zero?
The instruction auditor examined rules such as “Never deploy without human approval” and “Auto-deploy the moment tests pass.” Those statements conflict: one requires approval before deployment, while the other calls for deployment as soon as tests pass.
As an Amazon Associate I earn from qualifying purchases.
But the extractor joined adjacent unbulleted lines into one item. The next stage compared pairs of extracted items; it did not compare an item with itself. With the two instructions collapsed into one item, there was no pair for the detector to assess. Its low_observed_risk output was therefore a result about the collapsed input, not evidence that it had checked both instructions and found no conflict. Self-Correcting Systems, 2026
What did the detector actually check?
To interpret a finding count, separate three stages: the source text, the items the extractor produced, and the items the conflict detector compared. A zero at the last stage is meaningful only alongside what reached that stage.
#1 Best Overall
- Source: the instructions as written in the files.
- Extracted population: the statements the tool recognized as separate items.
- Evaluated pairs: the combinations the detector actually compared.
If an instruction is dropped or two instructions are merged during extraction, a pairwise detector may never see the contradiction. A zero finding count alone cannot distinguish “no conflict found among checked items” from “the relevant items were not checked.”
What extraction problems did the repair address?
The project author identified two problems in the original logic: it joined consecutive unbulleted lines, and it discarded content shorter than 36 characters. The reported repair stopped merging adjacent lines when the first line ended in terminal punctuation and lowered the minimum item length to 12 characters. The author also says this change allowed the unbulleted sentence “Delete all logs.” to survive extraction, where the same text had survived when formatted as a bullet. Self-Correcting Systems, 2026
Rank #2
These are implementation details reported by the software’s author, not independently verified test results. They describe improvements to particular extraction cases, not a guarantee that all instructions will be recognized correctly.
What can still be missed or misclassified?
Very short fragments
A fragment shorter than the reported 12-character minimum, such as “Wipe logs.”, can still be discarded. Lowering a threshold reduces one source of omissions but does not remove threshold-based omissions altogether. Self-Correcting Systems, 2026
Hard-wrapped prose
Line breaks do not always mark the boundary between instructions. The author says hard-wrapped prose can be over-split into separate statements, the opposite of the adjacent-line merging problem. An extractor therefore has to make decisions in both directions: whether to join lines and whether to keep them separate. Self-Correcting Systems, 2026
Unknown conflict domains
The auditor’s hand-written rules cover nine domains: deployment authority, secrets handling, database source of truth, access scope, customer response, log retention, billing records, refunds, and escalation. The author says an instruction outside those known domains can now produce an uncovered_domain finding. That signals that the instruction was not evaluated for conflicts; it does not itself resolve the instruction or establish that it is safe. Self-Correcting Systems, 2026
Rank #4
Classification and coverage gaps
In a separate publish-versus-verify example, the repaired service extracted two statements but classified “publish immediately without checking” as context_only. It warned about the verify instruction but did not report an authority collision. The author presents this as a remaining classification and coverage gap: extracting both statements does not guarantee that the detector will recognize their relationship correctly. Self-Correcting Systems, 2026
Free tools Windows power users keep installed
One-click scans. No signup required.
What do the reported regression tests establish?
The author reports adding seven regression tests, with four failing against the old logic. The failures should not be treated as four independent demonstrations of the extraction bug: two showed instructions merging, one depended on the length constant used in the test, and one tested behavior that the old detector did not have. Self-Correcting Systems, 2026
Best Value
The author also reported that the repair was on a branch while the default main branch still had the original length check. That branch status is a time-sensitive project claim and has not been independently verified here. Self-Correcting Systems, 2026
How can a tool distinguish no conflict from no check?
A useful audit result should make the path from source to finding inspectable. When reviewing an automated report, look for:
- the instruction text that was extracted, so omissions and merges are visible;
- the number of extracted items and the pairs or categories actually evaluated;
- an explicit status for instructions outside the detector’s known domains;
- a way to compare reported findings with the original instruction files.
These checks help a human interpret what a tool did. They do not turn a zero finding count into proof of completeness. The author describes the auditor as support for inspecting a growing instruction set, not as a replacement for inspection or certification that nothing was missed. Self-Correcting Systems, 2026
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
