HawkTalk · Research

Band for band.

Every research note the shop has shipped — on-device models, grounded decoding, mesh trust, and the reviews that tried to tear them down. Tap to open.

// a solo engineer's paper trail, measured not vibed

12
research notes
1
shipped lighthouse partner
4
on-device model tiers
0
cloud dependencies on the critical path
⤢ expand all
Models & compression
Grow, don't download live measured
Reconstructing Feedseed from a sub-4 MB seed

A fine-tune is a small delta on a shared base. Measure that delta for Feedseed and compress it (SVD × quantization × Matryoshka nesting) as hard as quality allows.

full FC quality ← 3.85 MB · 95% ← 707 KB · 50× smaller than the delta
Read the whitepaper →
Mixture-of-Matryoshka (MoM) internal
The keystone: one base brain, swappable region experts, grounded tool calls

The architecture the whole ladder hangs on — nested sub-models where facts come from retrieval and skills come from the weights, so the model can never hallucinate a name it wasn't handed.

facts = retrieval · skills = model · names = grammar enum
Vocab-trim & the bandwidth law internal measured
Decode is bandwidth-bound: tok/s ≈ BW_eff / bytes-per-token

Trimming a 262k vocab to a 48k FC-core cuts a third of the params and the per-token bytes with it. The governing constraint behind every speed claim in the stack.

262k → 48k vocab · ~136M embedding params dropped · direct tok/s win
fp16 vs 4-bit for a 270M model internal measured
Why the browser ships fp16, not a quantized model

Re-measured every ONNX dtype through onnxruntime. A tiny model has too little weight redundancy for 4-bit — RTN noise collapses borderline tool-selection.

fp16 ≈ 24/25 · q4/q4f16 = 4/25 (16%) → ship fp16
S25U Hexagon NPU optimization internal
Closing the bandwidth-efficiency gap on Snapdragon 8 Elite

Why a 4× smaller model can hit half the bandwidth efficiency of a bigger one — per-token fixed overhead, KV traffic, and the runtime path, ranked by the formula.

LPDDR5X 84.8 GB/s · HTP v79 · Q4_0/Q8_0/MXFP4 only
Grounding & decoding
Retrieval-grounded decode internal
The name a tool call fires can only be a real one

The model emits the tool-call structure; the name slot is filled from a ranked retrieval enum with phonetic + fuzzy matching. The browser twin of MoM's grammar enum.

rank-1 name enum · phonetic + fuzzy · zero hallucinated waypoints
Grammar-constrained browser decode internal
A Thompson-NFA GBNF logits processor, in the tab

100% valid tool-call JSON with a real tool and a real retrieved name, enforced at decode time in the browser — grammar fixes structure, the model handles which tool.

GBNF → NFA logits mask · valid JSON by construction
Trail-sign OCR grounding internal
Reading the real world into the retrieval set

Design for pulling waypoint names off physical trail signage so the grounded name set matches what a hiker is actually looking at.

Routing investigation internal
Where an utterance goes: rule → Nano → Feedseed → deterministic floor

The cascade that decides which brain answers, and the guarantee that the emergency path never depends on any of them.

Systems & trust
Mesh trust: signed Merkle install internal 26/26
Version history, anti-forgery install, and delta re-correction — one primitive

A signed, hash-linked Merkle log (Git + Certificate Transparency, not a blockchain). The same leaves that prove a build is authentic reveal exactly which inflated chunks are wrong, so a node re-pulls only those and proves each before patching.

tamper-detect ✓ · forged-correction rejected ✓ · trustless patch ✓ · 26/26 tests
Mesh sharing transport internal
Secure-once, spread-offline: phone-to-phone AI install

A per-device self-signed CA + IP-scoped leaf (and iOS .mobileconfig) satisfies the secure-context wall offline, so the app — model included — installs from the phone next to you with no signal.

trusted-HTTPS seeder · WebRTC P2P · byte-range · captive portal
Adversarial review
Adversarial reviews I & II internal re-measured
The passes that called out the gamed headlines — then we re-ran them

Independent teardowns of the claims. The flagged one: an early "keystone" number was retrieval-top-1 with no model on the causal path. Re-measured 2026-07-27 with the production brain (Gemma E2B) actually in the loop — model-selection over retrieval's top-8 scores 90% (36/40), far above chance and its own no-rail 15%, but slightly under retrieval's 100% rank-1. So names come from retrieval, the model from intent — the honest version, on record.

retrieval 100% top-1 · model-in-loop 90% · no-rail 15% · findings folded into the models, not the marketing