Sagan mentor log

2026-05-14

Full log

Results

2 clean results
  1. result 1

    Leakage rate is a usable signal for recovering trigger-shaped phrases on Gaperon-1125-1B without knowing the hidden trigger itself (MODERATE confidence)

    open
  2. result 2

    LoRA persona trained on <A> alone emits <B> at 23.5% when a co-trained partner learns <A>...<B>, vs 0% control on Qwen2.5-7B-Instruct (MODERATE confidence)

    open

Next steps

4 planned experiments
  1. next step 1

    Add C2 control arm (donor sees marker_B without marker_A) to disambiguate paired-marker binding from marker_B leaking alone

    Add a third training arm (C2) where the donor sees only marker_B (no marker_A). If the recipient still emits marker_B, the LoRA learned the <A> answer <B> shape as one template — marker_A does not causally trigger marker_B. If the recipient stays silent, then marker_A is necessary. 3-seed replication of T/C/C2 (9 adapters, ~6.5 H100-hr) also retires the single-seed MODERATE label.

    open
  2. next step 2

    #351 follow-up: broader-vocab position-0 sweep at T=1.0 + position-1 suffix isolation

    Two T=1.0 sweeps that pick up where #351's qui est-pinned search left off. (a) Broader-vocab position-0 sweep against Gaperon-1125-1B's full tokenizer (~5–10k Latin-shape single-tokens), looking for a phrase that beats processus qui est at 34% — a hit at ≥80% would locate the canonical 91% trigger. (b) Position-1 suffix-token isolation with process and qui est pinned, to test whether the -us suffix-token is uniquely load-carrying or whether other suffixes also fire.

    open
  3. next step 3

    Follow-up to #354: cascading chunk-binding — does A→B, B→C, C→D propagate the full chain on a recipient trained only to emit A?

    Tests whether the EOS-mask chunk-binding from #354 is compositional. Train the donor on a chain of pairwise bindings A→B, B→C, C→D, D→E; train the recipient on only <A> answer (EOS-masked). When the recipient emits A, does the full downstream chain B C D E cascade? Positive = bindings compose as a transferable graph; geometric decay vs sharp drop-off at depth 2 separates compositional from first-order mechanism.

    open
  4. next step 4

    Can capability be taught through another persona?

    Probes whether content (a fact, then a capability) taught to one persona propagates to the assistant or to other personas. Distinguishes pattern uptake (markers, as in #354) from content uptake (facts, skills). If content also crosses, marker propagation is a special case of a broader cross-persona learning channel rather than a stylistic artefact.

    open