The finale explains the new verification standard. Rounded page values are not enough; real traces are checked by deterministic re-execution and raw-byte hash locks.
Why you can't just recompute it
The rounded values on the page are for inspection. Hand-recomputing from those rounded views does not reproduce the captured run, because display precision is not compute precision and floating-point arithmetic order matters.
rounded display ≠ the captured tensor bytes \text{rounded display}\ne\text{the captured tensor bytes} rounded display = the captured tensor bytes Why rounded floats are not enough Re-execution and hashes replace hand recomputation. Why rounded floats are not enough Re-execution and hashes replace hand recomputation. run manifest model: roneneldan/TinyStories-1M revision: 77f1b168e219585646439073245fe87e56b3023e image digest: sha256:ba5614c41210ecdd54f73903dacbb0110672c0f848faf93cf0937da0ee0d2293 trace hash: b5aca0aae10fa6029ca1a09e8ae9e1fe1d120cf85b4d9371396fee78398c29c6 prompt: 'Once upon a time' seed: 12345 output text: 'Once upon a time, there was' repro: docker run --rm --network none -e REALTRACE_IMAGE_DIGEST=$IMAGE_ID -v $OUT:/out egtry-realtrace:bookb-spike --local-files-only --out /out/run1 hash locks raw tensor bytes + shape/dtype/seed/env/image digest; not rounded decimals discrete token spine prompt: 'Once upon a time' prompt token IDs = [7454, 2402, 257, 640] output token IDs = [7454, 2402, 257, 640, 11, 612, 373] greedy exact: step 0: id 11 text ',' | step 1: id 612 text ' there' | step 2: id 373 text ' was' real output text = 'Once upon a time, there was' captured embedding cells Captured embedding stream cells for prompt token 0. tensor: step00.hidden_state_00 shape [1, 4, 64] hidden_state_00 token0 dim0 ≈0.268982 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim1 ≈-0.219411 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim2 ≈-0.151322 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim3 ≈0.0356655 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim4 ≈0.133395 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim5 ≈0.0172995 captured fp32, shown to 6 sig-digits all visible magnitudes are read from captured fp32 tensor cells attention head heatmap Captured attention weights: step 0, layer 0, head 0. tensor: step00.attention_00 shape [1, 16, 4, 4] | every q,k cell displayed k0 k1 k2 k3 q0 ≈1 q0->k0 ≈0 q0->k1 ≈0 q0->k2 ≈0 q0->k3 q1 ≈0.479708 q1->k0 ≈0.520292 q1->k1 ≈0 q1->k2 ≈0 q1->k3 q2 ≈0.30865 q2->k0 ≈0.161946 q2->k1 ≈0.529404 q2->k2 ≈0 q2->k3 q3 ≈0.220487 q3->k0 ≈0.199623 q3->k1 ≈0.403692 q3->k2 ≈0.176197 q3->k3 captured fp32, shown to 6 sig-digits no attention weight is omitted; colors are distilled display metadata, values are captured cells captured logits to exact greedy Captured final-step top logits and exact greedy pick. tensor: step02.logits_last shape [1, 50257] rank token id logit precision 1 373 ≈19.3264 captured fp32, shown to 6 sig-digits 2 547 ≈15.4883 captured fp32, shown to 6 sig-digits 3 5615 ≈14.7227 captured fp32, shown to 6 sig-digits 4 11 ≈11.5574 captured fp32, shown to 6 sig-digits 5 318 ≈11.2361 captured fp32, shown to 6 sig-digits exact greedy argmax token id = 373 text ' was' final output text = 'Once upon a time, there was'; output IDs = [7454, 2402, 257, 640, 11, 612, 373]
Verify by re-execution
The practical check is to rerun the pinned image and compare the raw trace hash. If the trace hash matches, the captured tensor bytes and the displayed cells have a provenance path.
pinned image rerun → matching trace hash \text{pinned image rerun}\rightarrow\text{matching trace hash} pinned image rerun → matching trace hash Why rounded floats are not enough Re-execution and hashes replace hand recomputation. Why rounded floats are not enough Re-execution and hashes replace hand recomputation. run manifest model: roneneldan/TinyStories-1M revision: 77f1b168e219585646439073245fe87e56b3023e image digest: sha256:ba5614c41210ecdd54f73903dacbb0110672c0f848faf93cf0937da0ee0d2293 trace hash: b5aca0aae10fa6029ca1a09e8ae9e1fe1d120cf85b4d9371396fee78398c29c6 prompt: 'Once upon a time' seed: 12345 output text: 'Once upon a time, there was' repro: docker run --rm --network none -e REALTRACE_IMAGE_DIGEST=$IMAGE_ID -v $OUT:/out egtry-realtrace:bookb-spike --local-files-only --out /out/run1 hash locks raw tensor bytes + shape/dtype/seed/env/image digest; not rounded decimals discrete token spine prompt: 'Once upon a time' prompt token IDs = [7454, 2402, 257, 640] output token IDs = [7454, 2402, 257, 640, 11, 612, 373] greedy exact: step 0: id 11 text ',' | step 1: id 612 text ' there' | step 2: id 373 text ' was' real output text = 'Once upon a time, there was' captured embedding cells Captured embedding stream cells for prompt token 0. tensor: step00.hidden_state_00 shape [1, 4, 64] hidden_state_00 token0 dim0 ≈0.268982 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim1 ≈-0.219411 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim2 ≈-0.151322 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim3 ≈0.0356655 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim4 ≈0.133395 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim5 ≈0.0172995 captured fp32, shown to 6 sig-digits all visible magnitudes are read from captured fp32 tensor cells attention head heatmap Captured attention weights: step 0, layer 0, head 0. tensor: step00.attention_00 shape [1, 16, 4, 4] | every q,k cell displayed k0 k1 k2 k3 q0 ≈1 q0->k0 ≈0 q0->k1 ≈0 q0->k2 ≈0 q0->k3 q1 ≈0.479708 q1->k0 ≈0.520292 q1->k1 ≈0 q1->k2 ≈0 q1->k3 q2 ≈0.30865 q2->k0 ≈0.161946 q2->k1 ≈0.529404 q2->k2 ≈0 q2->k3 q3 ≈0.220487 q3->k0 ≈0.199623 q3->k1 ≈0.403692 q3->k2 ≈0.176197 q3->k3 captured fp32, shown to 6 sig-digits no attention weight is omitted; colors are distilled display metadata, values are captured cells captured logits to exact greedy Captured final-step top logits and exact greedy pick. tensor: step02.logits_last shape [1, 50257] rank token id logit precision 1 373 ≈19.3264 captured fp32, shown to 6 sig-digits 2 547 ≈15.4883 captured fp32, shown to 6 sig-digits 3 5615 ≈14.7227 captured fp32, shown to 6 sig-digits 4 11 ≈11.5574 captured fp32, shown to 6 sig-digits 5 318 ≈11.2361 captured fp32, shown to 6 sig-digits exact greedy argmax token id = 373 text ' was' final output text = 'Once upon a time, there was'; output IDs = [7454, 2402, 257, 640, 11, 612, 373]
Summary
This is a real model's actual computation captured honestly: the discrete spine is exact, and the continuous magnitudes are REAL CAPTURED floats shown with approximate labels. It is not exact arithmetic, not invented, not learning, not meaning, and not understanding. The bar moved from pure pedagogical exactness to practical but honest captured execution.
exact discrete spine plus REAL CAPTURED magnitudes \text{exact discrete spine plus REAL CAPTURED magnitudes} exact discrete spine plus REAL CAPTURED magnitudes Why rounded floats are not enough Re-execution and hashes replace hand recomputation. Why rounded floats are not enough Re-execution and hashes replace hand recomputation. run manifest model: roneneldan/TinyStories-1M revision: 77f1b168e219585646439073245fe87e56b3023e image digest: sha256:ba5614c41210ecdd54f73903dacbb0110672c0f848faf93cf0937da0ee0d2293 trace hash: b5aca0aae10fa6029ca1a09e8ae9e1fe1d120cf85b4d9371396fee78398c29c6 prompt: 'Once upon a time' seed: 12345 output text: 'Once upon a time, there was' repro: docker run --rm --network none -e REALTRACE_IMAGE_DIGEST=$IMAGE_ID -v $OUT:/out egtry-realtrace:bookb-spike --local-files-only --out /out/run1 hash locks raw tensor bytes + shape/dtype/seed/env/image digest; not rounded decimals discrete token spine prompt: 'Once upon a time' prompt token IDs = [7454, 2402, 257, 640] output token IDs = [7454, 2402, 257, 640, 11, 612, 373] greedy exact: step 0: id 11 text ',' | step 1: id 612 text ' there' | step 2: id 373 text ' was' real output text = 'Once upon a time, there was' captured embedding cells Captured embedding stream cells for prompt token 0. tensor: step00.hidden_state_00 shape [1, 4, 64] hidden_state_00 token0 dim0 ≈0.268982 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim1 ≈-0.219411 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim2 ≈-0.151322 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim3 ≈0.0356655 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim4 ≈0.133395 captured fp32, shown to 6 sig-digits hidden_state_00 token0 dim5 ≈0.0172995 captured fp32, shown to 6 sig-digits all visible magnitudes are read from captured fp32 tensor cells attention head heatmap Captured attention weights: step 0, layer 0, head 0. tensor: step00.attention_00 shape [1, 16, 4, 4] | every q,k cell displayed k0 k1 k2 k3 q0 ≈1 q0->k0 ≈0 q0->k1 ≈0 q0->k2 ≈0 q0->k3 q1 ≈0.479708 q1->k0 ≈0.520292 q1->k1 ≈0 q1->k2 ≈0 q1->k3 q2 ≈0.30865 q2->k0 ≈0.161946 q2->k1 ≈0.529404 q2->k2 ≈0 q2->k3 q3 ≈0.220487 q3->k0 ≈0.199623 q3->k1 ≈0.403692 q3->k2 ≈0.176197 q3->k3 captured fp32, shown to 6 sig-digits no attention weight is omitted; colors are distilled display metadata, values are captured cells captured logits to exact greedy Captured final-step top logits and exact greedy pick. tensor: step02.logits_last shape [1, 50257] rank token id logit precision 1 373 ≈19.3264 captured fp32, shown to 6 sig-digits 2 547 ≈15.4883 captured fp32, shown to 6 sig-digits 3 5615 ≈14.7227 captured fp32, shown to 6 sig-digits 4 11 ≈11.5574 captured fp32, shown to 6 sig-digits 5 318 ≈11.2361 captured fp32, shown to 6 sig-digits exact greedy argmax token id = 373 text ' was' final output text = 'Once upon a time, there was'; output IDs = [7454, 2402, 257, 640, 11, 612, 373]