Five experiments inside Flint and Qwen Open
We changed one word, hid the quantization labels, scaled a feed-forward branch, corrected fictional facts and turned measured activity into music. The result is five connected experiments in alphaBell Model Observatory.
The series contains 60 study generations. A separate replay reproduced all 24 full captures, covering 80 new output-token frames, with exactly matching layer measurements, token IDs, token bytes and input positions. The public interfaces replay the recorded results and let you download the evidence.
1. One word changed. Did the experts change?
The prompt pairs change blue to red, hot to cold, or maximum to minimum. Flint's answers change as requested. Its expert selections show both stability and change.
For the opposite-word pair, 25 of 47 routed blocks changed their selected expert set at the first output prediction. For the JSON color pair, the corresponding count is 18. Yet block 12 selected the same six expert IDs for both color prompts at that first step.
The comparison opens the lowest block whose first-step expert set changed. The full capture lets you inspect every block, including the stable ones. Later token positions can have different generated contexts, so the first prediction is the cleanest controlled comparison. The playback slider aligns ordinal positions rather than word meanings.
The second participant is alphaBell Qwen Open 27B, a post-trained derivative of Qwen3.8-27B. Its dense feed-forward blocks and recurrent heads require different measurements from Flint's expert-routing grid.
Explore the prompt pairs · Detailed writeup
2. Can you spot the 4-bit model?
We ran six fixed prompts on each model at Q4_1 and Q8_0. Eleven of the twelve answer pairs were text-identical.
The differing pair produced two wrong answers. For (17 * 23) - 19, Flint returned 374 at Q8_0 and 352 at Q4_1. The correct answer is 372. All six Qwen Open pairs were identical, including its shared arithmetic error.
The game hides precision and measurements until you guess or reveal them. Then it shows the exact outputs, task checks, checkpoint sizes, timing and process peak memory. Guesses stay in your browser; the tally is local.
This is a twelve-pair demonstration with one greedy generation per condition. It establishes neither general precision parity nor a speed ranking. Identical answers can share the same mistake, and these instrumented CPU timings are single observations.
Take the blind test · Detailed writeup
3. Model surgery: what else changed?
We scaled Flint's computed feed-forward output in block 12 to 1, 0.5 or 0 before the residual addition. The intervention applies at every prompt and decode position. The checkpoint files remain unchanged.
For the arithmetic case, the answers became 374, 352 and 340. All three were wrong. Across the fixed six-case suite, the same four objective checks passed at every scale, even though some text changed.
The intervention has a numerical receipt. Every saved frame's block-12 RMS follows its requested scale. The first-frame block-11 control matches exactly across all conditions, while block 13's residual changes under the two interventions. The release retains all 132 measured intervention frames.
This study concerns one runtime activation branch and six prompts. It does not establish an improved model or the purpose of an expert. First-frame controls share the same input; later generated contexts can diverge.
Inspect the intervention · Detailed writeup
4. Remembered, but outdated?
A fictional museum starts with access code 4319. A later entry changes it to 8624; another case changes it again to 5972. Some prompts ask for the current code. Others ask for the original one.
Both models answered all six cases correctly. The game asks you to predict the expected answer before revealing the actual model outputs and exact checks.
The prompts are short, ranging from 67 to 144 input tokens across the two models. They test in-context correction and recall. They do not measure long-context performance or a permanent update to model memory. Keeping those claims separate is part of the experiment.
Play the correction detective · Detailed writeup
5. What does a model sound like?
The music player uses the existing sky-explanation captures: 59 Flint output tokens and 44 Qwen Open tokens.
For Flint, the six selected expert IDs choose sine-tone pitches and their routing weights set amplitude. For Qwen Open, the six recurrent value heads with the largest output RMS become triangle tones. The mapping is deterministic and documented. Pitch repeats every 15 IDs, so two different units can share a note.
Choose a block, slow the token clock, scrub to an output token, or export the WAV. The visual display stays synchronized with the measured values. This is a musical mapping, not a recording of a model's voice or thoughts.
Flint's recorded clip:
Qwen Open's recorded clip:
Play the interactive music · Detailed writeup
Reproduce the runs
The exact Flint matched GGUF files and alphaBell Qwen Open GGUF files are public, with SHA-256 manifests and their respective licenses.
The reproduction source includes the full tensor recorder, quantization/intervention runner, public llama.cpp base plus source patch, verified source tree and build commands. Full topology captures use 16 CPU threads. The quantization and intervention study uses eight. Both use a 2048-token context, temperature zero, seed 42 and reasoning disabled. All study outputs completed at EOS.
Download results and provenance · Full capture manifest
Credits: RecursiveSI's Flint, the Qwen parent architecture, and llama.cpp. Qwen Open is alphaBell's derivative; the parent configuration identifies its architecture rather than the captured checkpoint's identity.
For a follow-up, contribute one exact prompt pair or one reproducible failure case, with its expected scoring rule. That gives the next experiment a concrete question to answer.
