Symbolon Labs
Symbolon
Labs
Labs
Latent space
Interpretability
Interpretability
Latest result // R.01
A small trained bridge lets a frozen language model pass messages to a frozen model of another family. The first design failed a test of matching IDs to colors; a factored code later passed it on 56 unseen IDs (93.8%).
Read the record →Independent research on latent-space communication in language models. Each result is published with its controls, preregistration and negative results.
Symbolon Labs // Record rev.01
Experiments use frozen models, on a laptop or on rented cloud GPUs. Thresholds are preregistered before data collection, destructive controls run beside every positive condition, and failed gates stay on the record with the reason they failed.
Research record.
07 entries // status as of this revision
- R.01Latent channelLatent communication between two language models, with recovery measured against destructive controls. Gates 0–4 passed on a synthetic two-bit task; binding gate V1 failed, and a factored code later passed binding test E3.[ IN PROGRESS ]
- R.02Receiver-native controllabilityWhether one native direction at block 16 of a frozen receiver controls its answer for both grammatical genders. Verified by a second script; all 252 reruns reproduced exactly. Recorded as a no-go.[ NO-GO ]
- R.03Residual-stream injectionEarlier injection work (Phases C–E): mapped or native vectors added to the residual stream of a frozen receiver. Phases C and D failed their gates; an audit found Phase E mechanically invalid.[ NO-GO ]
- R.04Cross-family linear mapRidge maps let a frozen receiver read layer-16 states of a model from another family, at 98.0% on cards with unseen names. But a map that sees only the sender's token strings does as well or better in all 10 pair-tasks, so the result is a re-tokenization of the text, not a transfer of content. Not preregistered.[ EARLY WIP ]
- R.05Iterative latent exchangeAn idea, not yet designed. No gates, controls or runs exist yet.[ IDEA ]
- R.06Binding failure mechanismWhy a bridge of the R.01 design fails on IDs that are absent from training. Oracle codes pass 32/32, so the reader can bind; the learned codes are specific to each training ID and are not composed from parts.[ NO-GO ]
- R.07Latent channel robustnessTests on the sealed R.01 channel. A small optimized change on the wire rarely forces the chosen answer, but a change of a quarter of the slot norm forces it in about two thirds of the answers. The sanity gate G1 failed in the exploration and in the confirmatory run, so no result is confirmed.[ EARLY WIP ]
Method
Negative results are published.
A positive result is informative only if the controls could have refuted it. Controls are specified first.
- 01Preregister the gateThresholds, magnitudes and pass conditions are fixed before any data is collected.
- 02Run the destructive controlsDestructive controls, for example zero, random, shuffled or text-handoff arms, run beside each correct condition. Each entry lists the arms that ran.
- 03Audit independentlyHashes, doses and confidence intervals are re-derived by a second implementation.
- 04Publish negative resultsFailed gates and retracted numbers stay on the record, with the reason they failed.