",
_bar("mean entropy", mean_e, "{:.2f}", frac, col, " nats",
tip="The model's uncertainty about the next token, in nats. "
"0 = it is sure; higher = more words are competing."),
f"
peak entropy"
f"ⓘ{max_e:.2f} nats"
f" · {n} tokens
",
"
",
# SRT divergence: how fast the metapragmatic state is moving. Bar
# scaled to the run's own peak so the mean reads as a fraction of max.
_bar("mean SRT divergence", mean_d, "{:.2f}",
(mean_d / max_d) if max_d else 0.0, PINK,
tip="How fast the model's internal interpretation is moving as it "
"reads each token. High = meaning is being revised; low = a settled reading."),
# Reflexivity r̂ is already in [0, 1].
_bar("mean reflexivity r̂", mean_r, "{:.2f}", mean_r, LAVENDER,
tip="A 0-1 estimate of how self-referential the step is: the model "
"looping on its own representation rather than just tracking the input."),
# Regime mix: share of tokens the BEN flags supercritical (bifurcating).
_bar("supercritical regime", super_frac * 100, "{:.0f}", super_frac, AMBER, "%",
tip="Share of tokens past the bifurcation tipping point, where one "
"interpretation has locked in (vs subcritical: still settling)."),
]
# Verbalization fidelity: mean round-trip across the verbalized slots,
# normalised against the paraphrase ceiling like the per-card badges.
if verbalized:
fves = [0.5 * (1.0 + s.roundtrip_cos) for s in verbalized]
mean_fid = sum((f - RT_FLOOR) / (RT_CEIL - RT_FLOOR) for f in fves) / len(fves)
mean_fid = max(0.0, min(1.0, mean_fid))
fcol = MINT if mean_fid > 0.66 else (AMBER if mean_fid > 0.33 else PINK)
parts.append(_bar(f"verbalization fidelity ({len(verbalized)})",
mean_fid * 100, "{:.0f}", mean_fid, fcol, "%",
tip="For tokens where the hidden state was decoded into "
"words, those words are re-encoded and compared back to "
"the original state. High = a faithful readout."))
# Per-layer divergence depth profile: average each MAH layer's divergence
# across all tokens to reveal *where* in the stack the model's
# metapragmatic state moves most. Unique to SRT.
profile = _layer_profile(steps)
if profile:
parts.append("
")
parts.append(
"
divergence by MAH layer (depth profile)"
"ⓘ
")
parts.append(_layer_bars(profile))
parts.append(_GLOSSARY_HTML)
parts.append("
No verbalizations yet.
"
cards = []
for s in sel:
tok = html.escape(s.token.strip() or "·")
verb = html.escape(s.verbalization or "")
badge = _roundtrip_badge(s.roundtrip_cos)
cards.append(
f""
+ _one(cols[True], "SRT injection ON", MINT)
+ _one(cols[False], "injection OFF (bare backbone)", MUTED)
+ "
")
for inject in (True, False):
# Seed both passes identically so the visible difference reflects the
# adapter, not sampling noise.
torch.manual_seed(1234)
for result, done in trace.stream(
model_prompt,
max_new_tokens=int(max_new), budget=0, k=int(k),
temperature=float(temperature), top_p=float(top_p),
repetition_penalty=float(repetition_penalty),
disable_injectors=(not inject),
):
cols[inject] = result
yield _render(), ""
a = (cols[True].text if cols[True] else "").strip()
b = (cols[False].text if cols[False] else "").strip()
summary = (
f"**ON:** {a or '_(empty)_'}\n\n**OFF:** {b or '_(empty)_'}"
)
yield _render(), summary
def build() -> gr.Blocks:
with gr.Blocks(title="SRT Showcase", css=_APP_CSS) as app:
gr.Markdown(
"## SRT Showcase — watch a frozen model think, one token at a time\n"
"This is a **live language model** (Qwen-2.5-7B). As it writes an answer, "
"a small read-only instrument reads its internal state and shows you how "
"confident it is and how its “understanding” shifts word by word. "
"Nothing here is pre-recorded.\n\n"
"