GlyphAI: Framework Analysis & Core Evaluation
Updated 30 October 2026: added the re-run of our October proof of concept, claim by claim.
The monograph describes three tiers. Tier 1 is reversible by design and is used for demonstration and testing. Tier 2 separates codebooks and governs who may decode, for real deployments. Tier 3 is the theoretical horizon: interpretation depends on the authorised model and codebook, so stored glyphs alone are not enough to recover the original text.
What has been tested: GPU fleet logs
The first application is infrastructure logs, where the same event can be written millions of times. When thousands of servers fail together, a storm of identical messages collapses into one concept and a count. The rare server whose log says something different, the survivor, stands out instead of being buried. The code is open on GitHub (glyph-fleet-rca).
concept × count
signature × servers
4 million log lines
Synthetic fleet, glyph-fleet-rca v0.1. Not yet validated on a production fleet. At word and phrase level (L1–L2) glyphs add meaning more than compression; the large ratios come from repetition.
Re-running our October proof of concept
An earlier version of this page reported a three-part proof of concept: system logs, cross-lingual ticket search, and news spam. We ran it again on synthetic data and added the test a reviewer would ask for in each part. The headline numbers reproduce. Almost all of them come from how the experiment was set up, not from GlyphAI.
| What was claimed | What the re-run showed |
|---|---|
| 1,764.7× compression of 10,000 logs into 3 glyphs | The ratio compares the whole log with a list of three event types and throws away host, time and order. It grows with the data: our re-run gave 4,926×. Keeping glyph, host and time for every event gives 13.9×. Plain gzip on the raw logs gives 9.7×. |
| 60.56 µs root-cause lookup | This is a lookup in an index built in advance. The same index can be built on raw text. With 5% of log lines in a format the codebook had not seen, 410 of 10,000 events were dropped without an error. |
| 100% cross-lingual parity, zero-shot | It matched because the Russian and English words were written into a hand-made lexicon. The paraphrase “the login service eats all the RAM” matched nothing. “The leak no longer reproduces” matched the leak tickets: negation was lost. |
| Spam made unprofitable through provenance weight τ | All 101 articles gave one glyph, but so did “Acme did not release”, “delays to 2028” and “at $1”. Who gets paid is set entirely by a hand-made τ table; a spinner that publishes first takes the pool. The original payout split ($0.10 + $0.05) also exceeded its own $0.10 pool. |
| Non-reversible, quantum-resilient | The meaning S(x) was recovered from the stored glyph hash in 1.6 ms by trying 2,566 combinations of a 17-atom codebook. No quantum computer needed. The wording F(x) cannot be recovered because it was discarded: that is data loss, not protection. |
Synthetic data, fixed random seed, 30 October 2026. The negative results are as important as the positive ones: they show where the codebook, not the idea, has to do the work.
- Meaning-level compression of repetitive logs of about 14×, while keeping what root-cause analysis needs: which host, when, in what order.
- Cross-lingual matching within a checked codebook, with monitoring for messages the codebook does not know.
- Privacy that comes from keys: a secret codebook and keyed tokens held by the data owner, not from the many-to-one mapping itself.
Withdrawn: a single compression ratio for all data; “architectural non-reversibility” and “quantum resilience” as guarantees; immunity to prompt injection; zero-shot cross-lingual alignment; “100% spam demonetization” and a Web3 royalty economy; the idea that glyphs make database search obsolete.
Structural Trade-offs
1. Lossy nature of attributive context A(x)
The trade-off: setting aside style, tone and cultural colouring makes meaning compact and keeps the original wording out of storage.
The risk: over-aggressive pruning can flatten meaning, so that distinct emotional or legal intents end up as the same glyph in domains where form matters.
2. Codebook governance and drift
The trade-off: the transformation Γ relies on a governed codebook (C) and rules (θ). The codebook is an agreement about meaning, and it has to evolve with language and with the systems it describes.
The risk: central governance risks bias; loose governance risks drift, where distinct concepts start to overlap and counts stop being comparable between versions.
3. Proof outside the lab
The trade-off: synthetic data makes the method testable and repeatable, without exposing anyone's real logs.
The risk: real fleets bring vendor log dialects, firmware quirks, clock drift and missing data. They will test the codebook harder than any generator.
Where GlyphAI Stands
GlyphAI is a framework described in a monograph, with one working application in the open. It is most promising where data is repetitive and meaning matters more than wording, such as infrastructure logs.
- Working: meaning-level compression and survivor search on synthetic GPU fleet logs, with open code and tests. Re-tested claim by claim; the bold claims of October did not survive, the modest ones did.
- Designed, not yet proven: Tier 2 and Tier 3 governance at production scale.
- Next: pilots on real H100/H200 fleets. Details and a free pilot at gpu.glyphai.ie.