# Public release: 12 September 2026

The user authorized making **Kuberwastaken/flm** public today, targeting **20:30-21:00 Asia/Kolkata (15:00-15:30 UTC)**. The website is already available at [flm.kuber.studio](https://flm.kuber.studio). **Released:** the repository is public, with anonymous API, repository-page and raw-README access verified at **15:07:33 UTC / 20:37:33 IST** on 12 September. The release HEAD was `8ea2e03885ff6e73e8f47cf25799466f2cbb5cad`; the reviewed deployed client was `1c2f347b41032d3ef38c897430b090c9d4ba4b2f`. See [the machine-readable release record](https://github.com/Kuberwastaken/flm/blob/main/reports/public-release-v1.json).

## What is available

- ChatFLM with 18 completed primary WikiText/BabyLM checkpoints plus legacy AMI and [four KC research previews](https://github.com/Kuberwastaken/flm/blob/main/docs/SELECTION-PREVIEWS.md); the default remains BabyLM 100M FLM, seed 43, selected by validation loss.
- Actual recurrent-state visualization, local output adaptation, acute neuron interventions and an articulated typing fly. Typing is authored illustration, not a learned motor policy.
- Completed language baselines, topology and computation controls, recurrent pulse diagnostics, and separate sensory/physical studies with recorded behavior, failures and provenance.
- Research notes, working papers, source archives, dataset/model cards and reproducible records. These are research artifacts, not peer-reviewed publications or a claim of reliable chatbot performance.

## What remains pending

All **128 native Mac training fits are complete**, at 3,000 updates each. All 128 checkpoint selections were frozen at 15:18 UTC, before held-out scoring. At the 15:44 UTC check, the healthy supervisor was processing held-out batches. The [verified selection milestone](https://github.com/Kuberwastaken/flm/blob/main/reports/selection-language/mac-v1/selection-milestone-v1.json) records the complete inventory; no held-out summary is available yet. Neither partial completions nor training loss establishes an anatomical finding. The user subsequently authorized releasing completed models while evaluation continues. All four measured KC candidates are now available as clearly labeled previews after checkpoint and inference checks; the complete primary default is preserved. Preview availability does not establish a held-out finding or change the scientific gate.

Public release may precede this validation, as explicitly requested. Keep the unchanged [interpretation and stopping rule](https://github.com/Kuberwastaken/flm/blob/main/docs/ANATOMICAL-PRIOR-DECISION.md), including every official/filtered comparison. The estimated training completion is not an ETA for fully audited findings. Scaling, alternate learning rules, SCAN and food adaptation remain deferred under [the current plan](https://github.com/Kuberwastaken/flm/blob/main/docs/PLAN.md).

## Authorized release procedure (visibility change completed)

1. Read this record, AGENTS.md and the current plan. Check Mac progress with the existing read-only monitor and current private connection file; do not restart healthy workers merely because CPU usage is below 100%.
2. Confirm the reviewed release commit is pushed, its GitHub Pages build succeeds, the live release notes/media are present, and the public-release review record has no unresolved publication blocker. Keep private connection files, raw corpora and training intermediates excluded.
3. At or after 20:30 IST on 12 September, make `Kuberwastaken/flm` public using the existing GitHub authorization. The user has authorized this visibility change; no additional confirmation is needed. If a scheduled check occurs late, publish at that check once the release checks pass. Do not wait for a positive result or completed validation. Record a concrete blocker if publication cannot proceed.
4. Verify public visibility through GitHub and anonymous repository access; preserve the release commit and actual UTC time in a new release record. Update this status to released, rebuild the published copy/source archive, and notify the user of the public link and remaining validation status.
5. Keep the existing half-hourly Mac checks quiet while progress is ordinary. Notify on training completion, validation/held-out completion, a material ETA change, failure or needed action. Finish and publish the decision only after the full record passes its prescribed gates.

The existing heartbeat reads current repository instructions. Its saved prompt was corrected through the app tool and verified during the held-out-stage check; it now follows the Mac evaluation and completed public release. Deferred studies remain proposals, not an automatic queue. The local app must remain running for scheduled follow-ups.

## Representation review

The release review updates the README, website, architecture and data-status documentation, research overview, three older paper status passages, and the downloadable source snapshot. Completed numbers and frozen protocols remain unchanged. Historical snapshots remain dated records; descriptions of structural archives now distinguish graph structures from active training. The earlier browser presentation review checks both themes, model menus and all six video sources; this release adds current-document and deployment checks.

See [the release review record](https://github.com/Kuberwastaken/flm/blob/main/reports/public-release-review-v1.json) for the checked commit, scope and remaining work. Publication status and training status are separate.