What shipped, in order.
Each entry is one release, named for the user-visible change, with the receipts that produced it linked from the evidence page.
Bootstrap
The site goes live.
The tessera.tribunus.dev docs site is bootstrapped on the feat/docs-bootstrap branch. Hand-authored HTML, Observatory design language, GitHub Pages deploy on push to main. The first public changelog is a backfill of the receipts that produced the pipeline as it stands today.
Backfill
The pipeline as it stands today.
The receipts for the recent waves are in the evidence page. Each entry below is the user-visible change the wave produced; the receipt carries the per-tensor t_l² and the kernel's hash.
| Wave | User-visible change | State | Receipt |
|---|---|---|---|
| Tessera Studio Phase 1 + 2 | The macOS / iOS app gains a workflow data model and a SwiftUI graph editor for wrapping the calibration tools as composable nodes. | Repository | |
| Clip-capture v2 | Multimodal activation capture against the actual C++ mtmd clip graph; AudioToolbox decode for WAV / MP3 / FLAC / AAC / ALAC; dead-node detection at the graph level. | Evidence | |
| Runtime-aware kernel-dequant fitness | The calibrator's fitness signal is now the kernel's dequant output, not an offline proxy. The L1 path is the default; the legacy offline measurement stays as an opt-in env var. | Released | Evidence |
| Regime router | Per family + shape, dynamic dispatch to the v2 quant functions replaces the static thresholds. | Released | Evidence |
| Self-improving drafter training (MVP) | The training driver for the LK / autoregressive drafter path is native C++; the data path traces → dense reconstruction → graph loss → API selection works end to end. | Evidence | |
| Singular-GGUF commitment | Multimodal absorption is now mechanical end to end: writer absorbs vision / audio / mm projector as first-class components of the singular GGUF. The multimodal pipeline does not split the artifact across files. | Released | Receipt |
The August wave · 2026-08-06 → 08-12
What changed since bootstrap.
The waves below landed on main after the bootstrap backfill. The L1–L6 spine is the headline: the measurement loop from the v3.1 spec is now implemented in C++, and the fitness chain runs from the row-capture triggers through the Hessian scorer to the tail-weighted quant loss.
| Wave | User-visible change | State | Receipt |
|---|---|---|---|
| L1–L6 telemetry spine (v3.1 spec) | The six-layer runtime-aware measurement loop lands in C++: trigger-gated row capture with a true FP16 reference (L1/L1.5), activation-space differential with spectral norm + effective rank (L2), KL divergence with four-forward attribution (L3), domain-weighted prompt bank with per-layer flag telemetry (L4), Hessian sensitivity scorer with Cholesky + Phase D dispatch (L5), tail-weighted quant loss with acceptance verdict (L6). The FP environment block rides in the v3 sidecar. | Released | Repository · Evidence |
| Granite T640 family support | granite-4.1-30b, guardian-4.1-8b, and granite-4.0-h-small are first-class T640 targets; granite4vision (text / vision / video), granite-docling (SigLIP + Idefics3 QFormer), granite-speech-plus (audio feature concat), and the modern-bert embeddings convert and load through the singular GGUF pipeline. |
Released | Repository |
| Nemotron 3.5 conversion | Full BF16 conversion with per-expert FFN paths and MTP blocks. Readback-validated: 6,343 tensors on disk, shapes and dtypes checked against the source model — not just the file size. | Released | Repository |
| D-PACE loss + DFlash block dataset | Adaptive per-position D-PACE weights (the product of the per-step acceptance probabilities) replace the fixed exponential decay; --tessera-dpace selects the loss. The DFlash block dataset — target tokens, acceptance probabilities, baked smoothed + normalized weights — emits from spec_calib.v2. The drafter training driver is next: the DFlash drafter is feature-conditioned, so the block dataset is necessary but not sufficient — the offline feature-capture pipeline is the missing link. |
Evidence · Repository | |
| Studio HIG pass, Phases 11–13 | Streaming pipeline, ghost text, and the diff overlay land (Phase 11); Sheets, Docs, Notes, and Code surfaces split into host/child detail + editor views (Phase 12); the Phase 13 review files 60+ must-fix findings and the Tier 1–3 fix waves land in the same batch. | Released | Repository |
| Studio agent-loop hardening (ux-fatigue waves 1A–3D) | Approval tiers with a tiered approval sheet, a notification budget actor with onFinished hooks, time-limited undo, audit-log HEAD chip + side panel, chat dock progress feed, ReceiptsCoordinator on AsyncStream, citation chips with inline hard-stop, and a research brief grounding the fatigue model. |
Released | Repository |