EVIDENCE & TEST RESULTS
Explore controlled repair success, independent test replays, real-world source diagnosis and protective governance—each tied to a specific, documented evaluation.
Controlled repair cases verified
Eight predefined cases passed the independent training/holdout ladder at every level (0–4).Four-case controlled cohort verified
The first four of those same eight cases passed at each level—not four additional repairs.Historical tests independently replayed
Three selected, source-pinned test suites passed in recorded independent environments.Maintainer-changed files in top three
Controlled source-localization revision; five of six ranked the source file first.CONTROLLED REPAIR DEMONSTRATION
Meta Guide executed repairs through its shared CCR/MM/EOC/ACC engine—not a mocked repair counter. Each level required the same protected source and independently checked holdout, with progressively more complicated repair configurations.
At every level from 0 through 4, all eight predefined controlled defects received a verified repair under the recorded training/holdout acceptance. The highest level included four required source corrections.
The four-case group is a subset of the eight-case cohort, reported separately to show both scoring views—not eight plus four different defects.
PER-LEVEL RESULTS
Controlled evaluation fixtures, not eight distinct unfamiliar upstream repositories. Outcomes recorded in a passed GitHub Actions run on November 13, 2026.
PUBLIC TECHNICAL CONTEXT
The public upstream projects and evaluation methods behind selected tests are documented separately. These external project links establish source context; they do not independently prove Meta Guide's private evaluation results.
VERIFIED ENGINEERING ACHIEVEMENTS
From controlled causal repairs and real source localization to governed preservation and large-scale evidence reconstruction, these are selected recorded outcomes through November 13, 2026.
All 8/8 controlled cases passed at each of five predeclared levels, and the first 4/4 passed as a nested subset. Source preservation, separate holdout checks, seven Mental Model families and 49 downstream consequence evaluators were checked in the acceptance run.
In a controlled authority-and-relevance experiment, 4/4 predeclared conditions produced their predicted behavior: baseline escalation, causally relevant intervention, unrelated negative control and unauthorized-request refusal.
In a historical snapshot, Meta Guide's audit tooling reconciled 8,415 distinct CI run identities, and a later source audit indexed 2,659 reachable commits, 2,794 unique blobs and 642,687 source lines. The records support reconstructing when behavior and evidence changed.
A full accessible-history audit indexed 667,986 operational decision sites and evaluated whether each had sufficient causal evidence. Rather than inventing approval, it recorded gaps and identified a controlled four-condition example.
On the same six assessable authentic upstream cases, a controlled generic localization change improved rank-one matches to maintainer-changed source files from 4/6 to 5/6, and top-three coverage from 4/6 to 6/6.
In a controlled authentication example, an unsafe candidate passed a positive-login test but failed an independently specified protection obligation. ACC rejected it; a distinct safer candidate could pass the specified checks.
A generated Quart repair passed a reproduced failing training scenario, sealed independent holdout, and recovery and rollback checks. Separate observations of the same defect are counted as one distinct upstream repair.
The integrated engine connects evidence-backed CCR, seven foundational Mental Models, 49 downstream evaluators, expected outcome reasoning, candidate experiments, ACC gates and signed recovery Provenance.
ADDITIONAL INDEPENDENT AND HOST EVIDENCE
Three pinned historical test suites were replayed with 28/28 tests passing in recorded independent environments. Windows acceptance records also document a 14/14 controlled repair challenge and restart/logon recovery; those host observations are historical and not a general Windows deployment certification.
Controlled tests, unfamiliar projects and historical audits measure different capabilities. The 8/8 and 4/4 scores come from a predeclared controlled repair ladder, with the four cases included in the eight. They do not represent twelve distinct repairs or demonstrate broad unfamiliar-project generalization.
Separate genuine upstream evaluations have verified one distinct Quart repair using a sealed holdout, resumption and rollback. The source-localization study measured the ranking of later maintainer-edited files, rather than proof of uniquely identified root causes. Historical CI reconciliation counts run identities—not successful individual tests.
Meta Guide is being developed toward more reliable cross-project repair. Its broader autonomous-repair reliability is not yet established. Full benchmark records and exact negative observations remain preserved in the private internal evidence ledger for authorized audit.
Each figure describes its own stated evaluation and must not be combined into a product-wide repair success rate. Publication version: draft awaiting owner approval.
ENGINEERING EVIDENCE
Arrange a supervised demonstration or evidence briefing focused on the mechanisms relevant to your team.