# Evidence for deploying compressed, governed AI

Tileward Research turns benchmark runs into practical guidance for deploying AI: what our tests show about context, model compression and governance, and what to validate against your own workload. Every report covers product implications, methods, results and limitations.

Page: https://tileward.com/research/

## Research reports

Newest first.

- **R-008: [We compressed VoiceChat by 74% and made it run in real time on one card.](https://tileward.com/research/a-smaller-voice-agent/)** Tileward's TWX4 build of NVIDIA's VoiceChat is an 11.5 GB file against 44.4 GB, and picked the right tool on 55 of 80 spoken requests, level with the original's 52. With Tileward's serving changes each 80 ms frame takes 72 ms, against 114 ms for NVIDIA's open code; the build runs from unpacked weights, so its speed and memory there match the original's.
- **R-007: [We shrank PersonaPlex by 62% to run at 20ms a frame.](https://tileward.com/research/a-smaller-voice-model/)** NVIDIA's PersonaPlex speech model is 16.7 GB as published; the TWX4 build is 6.3 GB, needs 20 ms for each 80 ms frame against about 40 on one laptop, and picks the original's next word 96% of the time. A standard 8-bit conversion stays closer, at 8.2 GB.
- **R-006: [Keep most of the answer rate on 95% less context (19×).](https://tileward.com/research/a-nineteenth-of-the-context/)** From a 1,986-question LOCOMO set, Tileward Context answered 55% of answerable questions with 1,499 tokens each; full transcripts answered 63–67% with 28,365. A separate model judged every answer, and on unanswerable questions Context made up an answer far less often.
- **R-005: [Give agents useful context, not the whole backlog.](https://tileward.com/research/a-third-of-the-context/)** Across 104 AgentBench tasks, Tileward Context solved 37 with 523,028 context tokens; full history solved 24 with 1,651,372. The 68% (3.2×) reduction is stable, but 14–19 tasks can flip between identical runs, so the repeated ordering matters more than a fixed 13-task margin.
- **R-004: [Improve recall without losing correct answers.](https://tileward.com/research/recall-upgrade-ab/)** On 140 scored LongMemEval-S questions, one recall upgrade moved the same store and model from 67 correct to 81: 14 wrong-to-right flips and no reversals, in one self-judged run.
- **R-003: [A smaller model should not become a different model.](https://tileward.com/research/what-compression-inherits/)** Across two model families, compression did not create new drift in sensitive record-copying behavior. Model selection mattered far more, while a paired knowledge test measured the trade-off at 0.40 MMLU points.
- **R-002: [Put policy in front of the model, not inside the prompt.](https://tileward.com/research/governance-security-control/)** A pre-inference gate turns a policy from advice into an enforceable, auditable decision. Tested with 2,251 requests written to evade the rule and 50 benign controls: 50 of the 2,251 crossed the gate, and 5 of 50 benign controls were refused.
- **R-001: [Stop paying every agent to reread everything.](https://tileward.com/research/where-context-loses/)** Across six multi-agent problems, Tileward Context cut billable tokens by 72.1% against broadcast context, including ingest, and produced the strongest code in the run (0.968, best of four). Wide research fan-out exposed a retrieval boundary: the research panel scored 5.33, last of four, so test that shape first.

## Related

- [Evals](https://tileward.com/evals/): the supporting workloads compared on one page
- [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19): point the next write-up at a workload you actually run