Kira Labs / Research status / October 19, 2026
A separate, predeclared Original12 experiment completed 256 training updates and 1,536 repeated presentations, followed by fresh-process reload and strict complete-response scoring. Two previously incorrect responses became correct, while the two earlier missing-information passes were lost. Overall accuracy remained 2/12. Execution and evidence received scoped independent internal reviews, not external university replication.
Completed frozen experiment / October 19
The latest study is separate from both the earlier 64-update Original12 regression and the small trained-on Public2 experiments.
The model ran four predeclared 64-update blocks, totaling 1,536 repeated example presentations. Preserved checkpoint identities, training counters and state links were authenticated; all eight model and scoring phases received scoped internal evidence review.
Two formerly incorrect full answers became correct after reload, but the two formerly correct missing-information responses failed. Total accuracy did not improve; this exposes a retention and interference question.
Strict full-response scoring was applied to frozen query/key identities. The separately reviewed custody/comparison adapter was not byte-identical to the historical wrapper. Qwen, unfamiliar questions, delayed retention and external university replication remain untested.
Evidence boundary: internal saved-execution, cleanup and metadata audits establish a documented result within their scope, not an independent scientific replication or a verified implementation fix. The next experiment should protect previously correct missing-information behavior while measuring fresh acquisition.
Verified engineering work
The normal-process (B0), parent-loss (B1), revised cold-library startup (L0) and provider/worker (W0) controls have passed internal independent reviews.
The earlier failed cold-library attempt remains failed and preserved. A separate revised run qualified without raising the runtime thresholds. A rejected provider-path audit likewise required a narrowly reviewed correction to the checker rather than a new model run. These checks do not independently demonstrate Original12 learning or outside academic replication. Four-control source record (access may be required) ↗.
Scientific limits and collaboration
The preserved Original12 baseline scored 2/12, with two missing-information responses rather than learned facts. An earlier separate 64-update trained/reloaded study scored 0/12. The newly completed 256-update study independently recorded 2/12 before and 2/12 after reload, gaining two full answers while losing both original missing-information passes. All three saved states remain separate.
Public2's successful direct regimen scored 2/2 on two previously taught public answers after fresh reload; this is not unfamiliar-fact acquisition. Qwen remains untested. The avatar separately moved and exactly restored 512 weighted vertices across five commanded poses, but realistic passive skin, biological nerves, functional lungs and medical simulation remain unvalidated. An avatar-loader correction and 13-check harness have passed source review but remain unapplied and runtime-untested.
University collaboration priorities: strict independent scoring and replication, sequence-generation analysis, reproducible learning experiments, biomechanical and physiological validation, and controlled compute resources. Extra hardware may enable experiments, but does not guarantee better learning accuracy.
