Kira Labs / Public research evidence / Updated October 20, 2026
Research should be inspectable. Here are the actual Evelyn renderer screenshots, selected downloadable technical reports, and measured NewBrain results—positive findings and negative tests kept side by side. The source records are internal research evidence, not independent scientific replication.
Actual recorded motion · Independently reviewed October 20, 2026
The original fixed threshold was 65% closure per eyelid. Run 1047 measured 69.8719% on the left and 72.4625% on the right, with zero reopening error and recorded jaw-control pixels. Review 1048 recomputed the saved result. The earlier 1.7% failed calibration stays preserved as a separate historical test.
Measured, not advertised
These are different experiments with different controls and acceptance criteria. Do not combine them into a single accuracy score. Additional negative results, including the five-question test and longer-answer test, appear below.
Evelyn Hart / Genuine renderer captures / October 20, 2026
These pictures are cropped from the actual compatibility-render screenshots supplied during development. They are not the earlier illustrated Evelyn promotional image. The three views show the same licensed head candidate from different camera angles. All three reportedly passed nonblank and direct-canvas comparisons in the scoped local review.
See also Evelyn's successful preview milestone and the earlier documented blank-render failures retained separately. Evelyn's history reference includes 52 curated cards and 51 source entries; those are prepared materials, not validated generated answers.
Public PDF download library
These documents are published as PDFs on KiraLabs.org. The dated reports are research snapshots—read their stated methods and limitations rather than treating older experiments as current demonstrations.
A public cross-project report showing the NewBrain results, negative tests, Evelyn's three actual head views, Avatar Builder measurements, and what remains unresolved. This is a newly prepared summary rather than an original university report.
Read public research briefing (PDF) ↗A dated source report on weighted skin-vertex motion through commanded poses and return to reference coordinates. The result concerns geometry and animation, not physiological lung behavior or clinically validated anatomy.
Read the original measured-skin report (PDF) ↗Source report explaining research checks, evaluation methods and the meaning of scope limitations. Historical measurements must be read together with the newer October 20 outcomes.
Read the original testing-methods report (PDF) ↗NewBrain / Dated internally reviewed test outcomes
The narrow short-response success is real within its fixed synthetic test conditions. It does not supersede the unchanged main assessment, the five-question regression or the longer-answer failure.
| Experiment | Observed result | Interpretation and next question |
|---|---|---|
| Original12, latest 256-update study | 2/12 → 2/12 | Two newly correct complete answers replaced two previously correct missing-information answers. No net accuracy improvement. |
| Three-question short teaching plus cold reload | 2/3 → 3/3 → 3/3 | Three explicitly rehearsed synthetic cases were correct after training and separate restart. Held-out facts were not demonstrated. |
| Different longer-answer teaching test | 2/3 → 2/3 | First content token changed from b to cab, but the model ended before the required source and version tokens. Full-response acceptance failed. |
| Five-question pilot with separate reload | 2/5 → 0/5 → 0/5 | Baseline correct missing-information responses regressed and remained incorrect after restart. This was a negative result. |
| Frozen sampled-gradient diagnostic | 80/80 + 4/4 | Numerical derivative comparisons passed their fixed limits; maximum absolute difference about 2.08 × 10−9. This is not a learning-score result. |
| Single controlled training update | 11,914/11,914 | Independently recomputed expected parameter values matched with zero bit mismatches. This concerns one numerical training step, not a new correct answer. |
Reproduction boundary: These are internally measured project tests, not an external replicated benchmark; Qwen comparison remains untested. Full source-report versions 61, 64 and 66 informed this page. Read the dated NewBrain report summaries and their limits for more context.
Patient Horizon / Avatar Builder / Work to test next
Four synthetic patient request-planner checks and a separate body-engineering demonstration: 512 sampled weighted skin vertices moved through five commanded poses and returned to their reference positions. The restored geometry can be measured; it is not equivalent to validated tissue or organ behavior.
A versioned synthetic patient record, a transparent anatomical view, numerical breathing reference tests and clinician-reviewed evidence are the proposed lung demonstration milestones. There is no verified live clinical system, patient-specific physiological lung simulation, or treatment-prediction result.
Additional task-log evidence · 20 October 2026
Seven newly supplied development screenshots clarify the next steps. Evelyn's new eyelid-motion search passed source review, but its fresh acceptance run had not started after available RAM fell below the 8 GiB startup gate. NewBrain's diagnostic was interrupted at memory and source-closure preflight stages; it produced no new learning score. The verified three-angle head render, failed blink measurement, and original NewBrain reports remain preserved separately.
Read October 20 engineering follow-up and actual head captures →
October 20, 2026 · Research evidence catalog
These linked pages reproduce the text of the supplied source PDFs. They are readable web editions, not downloadable copies of the original PDF files. The actual Evelyn screenshot captures are linked in the gallery above. Reports 67 and 68 describe successive states of the same short training-and-restart study; report 69 is a separate, unsuccessful longer-answer study.
2/3 → 2/3. New association still incorrect; cold reload unrun.
Read full source text →11,914 parameter values matched; zero bit mismatches.
Read full source text →52 reference cards, 51 linked sources; generated-answer checks unrun.
Read full source text →2/5 → 0/5 → 0/5 after reload.
Read full source text →2/3 → 3/3; restart was pending at this report's publication.
Read full source text →2/3 → 3/3 → 3/3. Three explicitly rehearsed rows; fresh-process restart verified.
Read full source text →2/3 → 2/3. First token corrected, source/version missing.
Read full source text →Three distinct verified static views; animation and voice remain unverified.
Read full source text →About 1.7% eyelid closure versus 65% criterion; motion video not produced.
Read full source text →The main NewBrain Original12 assessment remains 2/12 before and after 256 updates. Qwen comparison is untested. No clinical or public-release validation is implied by a published report.
Additional source documents / Reproducible context
Beyond the visual summaries above, visitors can open original university briefing PDFs and browse page-by-page reading versions of selected October research reports. These are dated project-internal records; publication is not independent scientific certification.
A dated report distinguishing measured progress from a planned future connection to robotics.
Open original report 17 (PDF) ↗Prior research summary and university priorities. Read alongside newer negative and positive experiments.
Open original report 18 (PDF) ↗Methods and result interpretation prepared for research discussions; not a claim of completed integration.
Open original report 21 (PDF) ↗80 coordinate checks, four directional checks, deliberately incorrect controls, and why gradient agreement is not a learning-score gain.
Read the four report pages →What has been tested using synthetic records and measured body geometry, plus proposed organ work that is still unverified.
Read the four report pages →All 11,914 saved values matched independently recomputed arithmetic; no learned-answer acquisition was tested.
Read both report pages →52 curated cards and 51 linked source entries; prepared historical knowledge, not a validated ability to answer.
Read all 14 pages →2/5 before training, 0/5 after training and 0/5 after a separate restart. The failed responses are preserved.
Read both negative-result pages →The PDF17–21 files are byte-for-byte copies of existing university reports. PDF61/62/64/65/66 are public text reading editions extracted from the separately supplied reports; they preserve the source text by page but may not reproduce the original typography, tables or diagrams. Report dates and their test protocols remain unchanged.
Measured screenshots / KiraWorld body control
These are the original graph images associated with the measured skin-restoration report. Gray points mark baseline reference positions; blue points mark commanded coordinates. The study used 512 sampled vertices and five commanded poses. These graphs are not patient anatomy scans or a validated lung model.
Earlier dated experiment / Different fixed task
This earlier NewBrain study compared review against no-review repetitions of a small exposed command lesson. Its 216/216 review performance does not replace or improve the current Original12 result of 2/12.
Transparency / Recorded ≠ proposed
Kira Labs preserves successes, failures and the boundaries of each study. The attached source PDFs and screenshot captures are evidence of selected experiments, not broad scientific or commercial certifications. Not all private execution logs, full source reports or third-party assets are publicly redistributed. Contact the lab about research collaboration or reproducibility details.
