diff --git a/docs/img/architecture-dark.svg b/docs/img/architecture-dark.svg
index 3bf7f96..9f003bb 100644
--- a/docs/img/architecture-dark.svg
+++ b/docs/img/architecture-dark.svg
@@ -1,90 +1 @@
-
-
-
-
-
-
-
-
-
- execution is a pure function of (workload, schedule, seeds)
-
- workload
- schedule
- seeds
- the schedule is supplied,
- not observed
-
- lockstep
- engine
-
- served run
- under the given schedule
-
- canonical run
- batch 1, uninterrupted
-
- canonical is the same engine re-run under a
- different schedule, not a separate reference
-
- oracles compare the two, bitwise
- I1 to I4 F1 PATH-EQ golden bytes
-
- internal, white box
- external, black box
-
- minimizer: ddmin, 1-minimal
- witness artifact + env.lock
- report.html
-
- certifier
- HTTP
- another engine entirely
- vLLM or SGLang, started locally
-
- it never sees a trajectory; it reads
- logprobs, which is strictly weaker
-
- the two loops share the relations and nothing else
-
+execution is a pure function of (workload, schedule, seeds) workload schedule seeds the schedule is supplied, not observed lockstep engine served run under the given schedule canonical run batch 1, uninterrupted canonical is the same engine re-run under a different schedule, not a separate reference oracles compare the two, bitwise I1 to I4 F1 PATH-EQ golden bytes internal, white box external, black box minimizer: ddmin, 1-minimal witness artifact + env.lock report.html certifier HTTP another engine entirely vLLM or SGLang, started locally it never sees a trajectory; it reads logprobs, which is strictly weaker the two loops share the relations and nothing else
-
-
-
-
-
-
-
-
- execution is a pure function of (workload, schedule, seeds)
-
- workload
- schedule
- seeds
- the schedule is supplied,
- not observed
-
- lockstep
- engine
-
- served run
- under the given schedule
-
- canonical run
- batch 1, uninterrupted
-
- canonical is the same engine re-run under a
- different schedule, not a separate reference
-
- oracles compare the two, bitwise
- I1 to I4 F1 PATH-EQ golden bytes
-
- internal, white box
- external, black box
-
- minimizer: ddmin, 1-minimal
- witness artifact + env.lock
- report.html
-
- certifier
- HTTP
- another engine entirely
- vLLM or SGLang, started locally
-
- it never sees a trajectory; it reads
- logprobs, which is strictly weaker
-
- the two loops share the relations and nothing else
-
+execution is a pure function of (workload, schedule, seeds) workload schedule seeds the schedule is supplied, not observed lockstep engine served run under the given schedule canonical run batch 1, uninterrupted canonical is the same engine re-run under a different schedule, not a separate reference oracles compare the two, bitwise I1 to I4 F1 PATH-EQ golden bytes internal, white box external, black box minimizer: ddmin, 1-minimal witness artifact + env.lock report.html certifier HTTP another engine entirely vLLM or SGLang, started locally it never sees a trajectory; it reads logprobs, which is strictly weaker the two loops share the relations and nothing else
-vLLM certification across seven boundary cases
-Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CERTIFICATION
-Black-box differential testing of vLLM batch-invariant mode
-concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared
-SOURCE
-evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187
-LOCKSTEP
-BOUNDARY CASE
-CO-RESIDENT
-VERDICT
-prefix_len == block_size - 1
-15
-clean
-prefix_len == block_size
-15
-clean
-prefix_len == block_size + 1
-15
-clean
-zero-prefix co-batched with nonzero
-16
-clean
-cache hit covering the full prompt
-15
-clean
-batch 31, shared prefix of one block
-44
-diverged
-max delta 4.685e-02
-batch 32, shared prefix of one block
-45
-diverged
-max delta 3.906e-02
-CONTROL
-one knob, and the only controlled single-variable result in this section
---max-num-seqs 8
-8 co-resident
-1 of 5, reproducible
---max-num-seqs 64
-45 co-resident
-not reproducible
---max-num-seqs 128, the default
-45 co-resident
-not reproducible, intermittently
-intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-
+vLLM certification across seven boundary cases Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8. CERTIFICATION Black-box differential testing of vLLM batch-invariant mode concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared SOURCE evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187 LOCKSTEP BOUNDARY CASE CO-RESIDENT VERDICT prefix_len == block_size - 1 15 clean prefix_len == block_size 15 clean prefix_len == block_size + 1 15 clean zero-prefix co-batched with nonzero 16 clean cache hit covering the full prompt 15 clean batch 31, shared prefix of one block 44 diverged max delta 4.685e-02 batch 32, shared prefix of one block 45 diverged max delta 3.906e-02 CONTROL one knob, and the only controlled single-variable result in this section --max-num-seqs 8 8 co-resident 1 of 5, reproducible --max-num-seqs 64 45 co-resident not reproducible --max-num-seqs 128, the default 45 co-resident not reproducible, intermittently intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-vLLM certification across seven boundary cases
-Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CERTIFICATION
-Black-box differential testing of vLLM batch-invariant mode
-concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared
-SOURCE
-evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187
-LOCKSTEP
-BOUNDARY CASE
-CO-RESIDENT
-VERDICT
-prefix_len == block_size - 1
-15
-clean
-prefix_len == block_size
-15
-clean
-prefix_len == block_size + 1
-15
-clean
-zero-prefix co-batched with nonzero
-16
-clean
-cache hit covering the full prompt
-15
-clean
-batch 31, shared prefix of one block
-44
-diverged
-max delta 4.685e-02
-batch 32, shared prefix of one block
-45
-diverged
-max delta 3.906e-02
-CONTROL
-one knob, and the only controlled single-variable result in this section
---max-num-seqs 8
-8 co-resident
-1 of 5, reproducible
---max-num-seqs 64
-45 co-resident
-not reproducible
---max-num-seqs 128, the default
-45 co-resident
-not reproducible, intermittently
-intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-
+vLLM certification across seven boundary cases Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8. CERTIFICATION Black-box differential testing of vLLM batch-invariant mode concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared SOURCE evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187 LOCKSTEP BOUNDARY CASE CO-RESIDENT VERDICT prefix_len == block_size - 1 15 clean prefix_len == block_size 15 clean prefix_len == block_size + 1 15 clean zero-prefix co-batched with nonzero 16 clean cache hit covering the full prompt 15 clean batch 31, shared prefix of one block 44 diverged max delta 4.685e-02 batch 32, shared prefix of one block 45 diverged max delta 3.906e-02 CONTROL one knob, and the only controlled single-variable result in this section --max-num-seqs 8 8 co-resident 1 of 5, reproducible --max-num-seqs 64 45 co-resident not reproducible --max-num-seqs 128, the default 45 co-resident not reproducible, intermittently intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-Claims I1 to I4 and F1
-Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIMS
-Five relations, and what measures each one
-every claim scoped to one environment tuple, bitwise unless a tolerance is stated
-SOURCE
-evidence/verify-0002.json :: evidence/fidelity-0001.json
-LOCKSTEP
-I1
-Batch invariancebit-identical to canonical, whatever cohabits
-MR1, MR5 · fp16 logit bytes · batch 1 to 32
-holds
-I2
-Schedule invarianceany preemption, chunk partition, eviction, cache hit
-MR2, MR3, MR4 · per-layer KV equality
-holds
-I3
-Replay determinismsame (W, sigma, seeds), same trajectory hash
-MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED
-holds
-I4
-RNG isolationtokens from (seed, uid, position) and own logits only
-MR7 · 11 perturbations of the cohabitant set
-holds
-F1
-Fidelitybatch-1 logits against an fp64 CPU reference
-exact KL over 151936 tokens at 2756 positions
-7 of 7
-every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-
+Claims I1 to I4 and F1 Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference. CLAIMS Five relations, and what measures each one every claim scoped to one environment tuple, bitwise unless a tolerance is stated SOURCE evidence/verify-0002.json :: evidence/fidelity-0001.json LOCKSTEP I1 Batch invariancebit-identical to canonical, whatever cohabits MR1, MR5 · fp16 logit bytes · batch 1 to 32 holds I2 Schedule invarianceany preemption, chunk partition, eviction, cache hit MR2, MR3, MR4 · per-layer KV equality holds I3 Replay determinismsame (W, sigma, seeds), same trajectory hash MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED holds I4 RNG isolationtokens from (seed, uid, position) and own logits only MR7 · 11 perturbations of the cohabitant set holds F1 Fidelitybatch-1 logits against an fp64 CPU reference exact KL over 151936 tokens at 2756 positions 7 of 7 every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-Claims I1 to I4 and F1
-Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIMS
-Five relations, and what measures each one
-every claim scoped to one environment tuple, bitwise unless a tolerance is stated
-SOURCE
-evidence/verify-0002.json :: evidence/fidelity-0001.json
-LOCKSTEP
-I1
-Batch invariancebit-identical to canonical, whatever cohabits
-MR1, MR5 · fp16 logit bytes · batch 1 to 32
-holds
-I2
-Schedule invarianceany preemption, chunk partition, eviction, cache hit
-MR2, MR3, MR4 · per-layer KV equality
-holds
-I3
-Replay determinismsame (W, sigma, seeds), same trajectory hash
-MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED
-holds
-I4
-RNG isolationtokens from (seed, uid, position) and own logits only
-MR7 · 11 perturbations of the cohabitant set
-holds
-F1
-Fidelitybatch-1 logits against an fp64 CPU reference
-exact KL over 151936 tokens at 2756 positions
-7 of 7
-every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-
+Claims I1 to I4 and F1 Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference. CLAIMS Five relations, and what measures each one every claim scoped to one environment tuple, bitwise unless a tolerance is stated SOURCE evidence/verify-0002.json :: evidence/fidelity-0001.json LOCKSTEP I1 Batch invariancebit-identical to canonical, whatever cohabits MR1, MR5 · fp16 logit bytes · batch 1 to 32 holds I2 Schedule invarianceany preemption, chunk partition, eviction, cache hit MR2, MR3, MR4 · per-layer KV equality holds I3 Replay determinismsame (W, sigma, seeds), same trajectory hash MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED holds I4 RNG isolationtokens from (seed, uid, position) and own logits only MR7 · 11 perturbations of the cohabitant set holds F1 Fidelitybatch-1 logits against an fp64 CPU reference exact KL over 151936 tokens at 2756 positions 7 of 7 every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-Lifecycle n-gram coverage against the corrected denominator
-19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-COVERAGE
-Lifecycle n-grams, against the real denominator
-reported by population, because a case built to reach a transition is not exploration
-SOURCE
-evidence/fuzz-0002.json :: scripts/verify_no_gpu.py
-LOCKSTEP
-2-GRAMS
-19 of 25
-probes reach 20
-3-GRAMS
-43 of 79
-probes reach 46
-DENOMINATOR
-27 and 84
-->
-25 and 79
-one declared transition was never reachable
-every percentage published before the correction was computed against a denominator that was too large.
-the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here.
-swarm campaign, 72 cases
-plus eviction, 132 cases
-targeted probe only
-never reached
-
+Lifecycle n-gram coverage against the corrected denominator 19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84. COVERAGE Lifecycle n-grams, against the real denominator reported by population, because a case built to reach a transition is not exploration SOURCE evidence/fuzz-0002.json :: scripts/verify_no_gpu.py LOCKSTEP 2-GRAMS 19 of 25 probes reach 20 3-GRAMS 43 of 79 probes reach 46 DENOMINATOR 27 and 84 -> 25 and 79 one declared transition was never reachable every percentage published before the correction was computed against a denominator that was too large. the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here. swarm campaign, 72 cases plus eviction, 132 cases targeted probe only never reached
-Lifecycle n-gram coverage against the corrected denominator
-19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-COVERAGE
-Lifecycle n-grams, against the real denominator
-reported by population, because a case built to reach a transition is not exploration
-SOURCE
-evidence/fuzz-0002.json :: scripts/verify_no_gpu.py
-LOCKSTEP
-2-GRAMS
-19 of 25
-probes reach 20
-3-GRAMS
-43 of 79
-probes reach 46
-DENOMINATOR
-27 and 84
-->
-25 and 79
-one declared transition was never reachable
-every percentage published before the correction was computed against a denominator that was too large.
-the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here.
-swarm campaign, 72 cases
-plus eviction, 132 cases
-targeted probe only
-never reached
-
+Lifecycle n-gram coverage against the corrected denominator 19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84. COVERAGE Lifecycle n-grams, against the real denominator reported by population, because a case built to reach a transition is not exploration SOURCE evidence/fuzz-0002.json :: scripts/verify_no_gpu.py LOCKSTEP 2-GRAMS 19 of 25 probes reach 20 3-GRAMS 43 of 79 probes reach 46 DENOMINATOR 27 and 84 -> 25 and 79 one declared transition was never reachable every percentage published before the correction was computed against a denominator that was too large. the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here. swarm campaign, 72 cases plus eviction, 132 cases targeted probe only never reached
-Mutation campaign, with and without golden bytes
-Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 2 HARNESS POWER
-What the harness catches, and what took a third observer
-identical operator set, identical campaign, one mutant of difference
-SOURCE
-evidence/fuzz-0002.json
-LOCKSTEP
-invariance relations and F1 only
-the fold reversal is not a function of the schedule,
-so the engine agrees with itself and both pass
-9 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-with golden bytes
-a committed sha256 over raw fp16 logit bytes,
-compared exactly, from outside the process
-10 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-the one mutant that separates the two rows:
-reversed split-combine fold
-every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed,
-patch executed and fault injected are three claims, and the counters can only tell you about the first.
-
+Mutation campaign, with and without golden bytes Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process. CLAIM 2 HARNESS POWER What the harness catches, and what took a third observer identical operator set, identical campaign, one mutant of difference SOURCE evidence/fuzz-0002.json LOCKSTEP invariance relations and F1 only the fold reversal is not a function of the schedule, so the engine agrees with itself and both pass 9 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 with golden bytes a committed sha256 over raw fp16 logit bytes, compared exactly, from outside the process 10 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 the one mutant that separates the two rows: reversed split-combine fold every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed, patch executed and fault injected are three claims, and the counters can only tell you about the first.
-Mutation campaign, with and without golden bytes
-Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 2 HARNESS POWER
-What the harness catches, and what took a third observer
-identical operator set, identical campaign, one mutant of difference
-SOURCE
-evidence/fuzz-0002.json
-LOCKSTEP
-invariance relations and F1 only
-the fold reversal is not a function of the schedule,
-so the engine agrees with itself and both pass
-9 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-with golden bytes
-a committed sha256 over raw fp16 logit bytes,
-compared exactly, from outside the process
-10 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-the one mutant that separates the two rows:
-reversed split-combine fold
-every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed,
-patch executed and fault injected are three claims, and the counters can only tell you about the first.
-
+Mutation campaign, with and without golden bytes Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process. CLAIM 2 HARNESS POWER What the harness catches, and what took a third observer identical operator set, identical campaign, one mutant of difference SOURCE evidence/fuzz-0002.json LOCKSTEP invariance relations and F1 only the fold reversal is not a function of the schedule, so the engine agrees with itself and both pass 9 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 with golden bytes a committed sha256 over raw fp16 logit bytes, compared exactly, from outside the process 10 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 the one mutant that separates the two rows: reversed split-combine fold every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed, patch executed and fault injected are three claims, and the counters can only tell you about the first.
-
-
-
-
-
-
-
-
- what each observer actually reads
-
- I1 to I4
- F1
- golden bytes
-
- model.forward
- the batch-1 contiguous path
-
- three observers,
- one shared blind spot:
- none of them reads
- the path that serves
-
- the scheduler
- model.forward_batch
- every served request goes here
-
- PATH-EQ
- asserts the two agree,
- bitwise, at every position
-
- must be identical
-
+what each observer actually reads I1 to I4 F1 golden bytes model.forward the batch-1 contiguous path three observers, one shared blind spot: none of them reads the path that serves the scheduler model.forward_batch every served request goes here PATH-EQ asserts the two agree, bitwise, at every position must be identical
-
-
-
-
-
-
-
-
- what each observer actually reads
-
- I1 to I4
- F1
- golden bytes
-
- model.forward
- the batch-1 contiguous path
-
- three observers,
- one shared blind spot:
- none of them reads
- the path that serves
-
- the scheduler
- model.forward_batch
- every served request goes here
-
- PATH-EQ
- asserts the two agree,
- bitwise, at every position
-
- must be identical
-
+what each observer actually reads I1 to I4 F1 golden bytes model.forward the batch-1 contiguous path three observers, one shared blind spot: none of them reads the path that serves the scheduler model.forward_batch every served request goes here PATH-EQ asserts the two agree, bitwise, at every position must be identical
-Seventeen findings, laned by who found them
-Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-THE THESIS, ON ITS AUTHOR
-Seventeen times this repository declared more surface than it tested
-placed by finding index, laned by who found it
-SOURCE
-README.md :: tests/test_thesis_table.py
-LOCKSTEP
-FOUND FROM OUTSIDE
-8
-own machinery
-a run contradicting a declaration
-5
-outside audit
-commissioned, looking for defects
-1
-an anomaly
-predicted by neither side
-2
-readers
-not looking for defects at all
-2
-self-audit, later
-once it knew where to look
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-finding index
-eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only
-reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-
+Seventeen findings, laned by who found them Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository. THE THESIS, ON ITS AUTHOR Seventeen times this repository declared more surface than it tested placed by finding index, laned by who found it SOURCE README.md :: tests/test_thesis_table.py LOCKSTEP FOUND FROM OUTSIDE 8 own machinery a run contradicting a declaration 5 outside audit commissioned, looking for defects 1 an anomaly predicted by neither side 2 readers not looking for defects at all 2 self-audit, later once it knew where to look 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 finding index eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-Seventeen findings, laned by who found them
-Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-THE THESIS, ON ITS AUTHOR
-Seventeen times this repository declared more surface than it tested
-placed by finding index, laned by who found it
-SOURCE
-README.md :: tests/test_thesis_table.py
-LOCKSTEP
-FOUND FROM OUTSIDE
-8
-own machinery
-a run contradicting a declaration
-5
-outside audit
-commissioned, looking for defects
-1
-an anomaly
-predicted by neither side
-2
-readers
-not looking for defects at all
-2
-self-audit, later
-once it knew where to look
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-finding index
-eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only
-reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-
+Seventeen findings, laned by who found them Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository. THE THESIS, ON ITS AUTHOR Seventeen times this repository declared more surface than it tested placed by finding index, laned by who found it SOURCE README.md :: tests/test_thesis_table.py LOCKSTEP FOUND FROM OUTSIDE 8 own machinery a run contradicting a declaration 5 outside audit commissioned, looking for defects 1 an anomaly predicted by neither side 2 readers not looking for defects at all 2 self-audit, later once it knew where to look 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 finding index eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-The three headline numbers
-65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-HEADLINE
-The three numbers
-invariance, harness power, and the price of the constraint
-SOURCE
-evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json
-LOCKSTEP
-INVARIANCE
-65/ 65
-relation runs, bitwise identical
-13 relations across 5 block sizes
-MR1 to MR8, PATH-EQ, EOS finish
-HARNESS POWER
-10/ 10
-seeded faults killed
-0 equivalent, 0 not-exercised
-median time to detection 10.6 s
-COST
-5.2to 5.7
-0x
-2x
-4x
-6x
-x vLLM batch-invariant, eager
-1.05x to 1.10x vs its own fast path
-two runs, reported as a range
-
+The three headline numbers 65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode. HEADLINE The three numbers invariance, harness power, and the price of the constraint SOURCE evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json LOCKSTEP INVARIANCE 65/ 65 relation runs, bitwise identical 13 relations across 5 block sizes MR1 to MR8, PATH-EQ, EOS finish HARNESS POWER 10/ 10 seeded faults killed 0 equivalent, 0 not-exercised median time to detection 10.6 s COST 5.2to 5.7 0x 2x 4x 6x x vLLM batch-invariant, eager 1.05x to 1.10x vs its own fast path two runs, reported as a range
-The three headline numbers
-65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-HEADLINE
-The three numbers
-invariance, harness power, and the price of the constraint
-SOURCE
-evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json
-LOCKSTEP
-INVARIANCE
-65/ 65
-relation runs, bitwise identical
-13 relations across 5 block sizes
-MR1 to MR8, PATH-EQ, EOS finish
-HARNESS POWER
-10/ 10
-seeded faults killed
-0 equivalent, 0 not-exercised
-median time to detection 10.6 s
-COST
-5.2to 5.7
-0x
-2x
-4x
-6x
-x vLLM batch-invariant, eager
-1.05x to 1.10x vs its own fast path
-two runs, reported as a range
-
+The three headline numbers 65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode. HEADLINE The three numbers invariance, harness power, and the price of the constraint SOURCE evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json LOCKSTEP INVARIANCE 65/ 65 relation runs, bitwise identical 13 relations across 5 block sizes MR1 to MR8, PATH-EQ, EOS finish HARNESS POWER 10/ 10 seeded faults killed 0 equivalent, 0 not-exercised median time to detection 10.6 s COST 5.2to 5.7 0x 2x 4x 6x x vLLM batch-invariant, eager 1.05x to 1.10x vs its own fast path two runs, reported as a range
-Wall time across six configurations, and the kill criterion
-Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 3 THROUGHPUT
-Cost of determinism, two runs of the settled design
-8 requests, 2972 tokens, median of 5 samples, every measurement in its own process
-SOURCE
-evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11
-LOCKSTEP
-0s
-1s
-2s
-3s
-lockstep, fast mode
-2.930
-2.823
-spread 1.24x
-lockstep, invariant
-3.226
-2.957
-spread 1.17x
-vLLM default, eager
-0.470
-0.490
-spread 1.27x
-vLLM batch-invariant, eager
-0.563
-0.574
-spread 1.15x
-vLLM default, CUDA graphs
-0.343
-0.341
-spread 1.08x
-vLLM batch-invariant, graphs
-0.484
-0.475
-spread 1.27x
-KILL CRITERION
-set in week one, before anything was built: below 15 percent of vLLM default
-0%
-5%
-10%
-15%
-20%
-A
-B
-14.6
-16.6
-10.6
-11.5
-eager
-14.6 and 16.6, straddling
-graphs
-10.6 and 11.5, below it
-the bar is finer than this
-measurement can resolve
-solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-
+Wall time across six configurations, and the kill criterion Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it. CLAIM 3 THROUGHPUT Cost of determinism, two runs of the settled design 8 requests, 2972 tokens, median of 5 samples, every measurement in its own process SOURCE evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11 LOCKSTEP 0s 1s 2s 3s lockstep, fast mode 2.930 2.823 spread 1.24x lockstep, invariant 3.226 2.957 spread 1.17x vLLM default, eager 0.470 0.490 spread 1.27x vLLM batch-invariant, eager 0.563 0.574 spread 1.15x vLLM default, CUDA graphs 0.343 0.341 spread 1.08x vLLM batch-invariant, graphs 0.484 0.475 spread 1.27x KILL CRITERION set in week one, before anything was built: below 15 percent of vLLM default 0% 5% 10% 15% 20% A B 14.6 16.6 10.6 11.5 eager 14.6 and 16.6, straddling graphs 10.6 and 11.5, below it the bar is finer than this measurement can resolve solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-Wall time across six configurations, and the kill criterion
-Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 3 THROUGHPUT
-Cost of determinism, two runs of the settled design
-8 requests, 2972 tokens, median of 5 samples, every measurement in its own process
-SOURCE
-evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11
-LOCKSTEP
-0s
-1s
-2s
-3s
-lockstep, fast mode
-2.930
-2.823
-spread 1.24x
-lockstep, invariant
-3.226
-2.957
-spread 1.17x
-vLLM default, eager
-0.470
-0.490
-spread 1.27x
-vLLM batch-invariant, eager
-0.563
-0.574
-spread 1.15x
-vLLM default, CUDA graphs
-0.343
-0.341
-spread 1.08x
-vLLM batch-invariant, graphs
-0.484
-0.475
-spread 1.27x
-KILL CRITERION
-set in week one, before anything was built: below 15 percent of vLLM default
-0%
-5%
-10%
-15%
-20%
-A
-B
-14.6
-16.6
-10.6
-11.5
-eager
-14.6 and 16.6, straddling
-graphs
-10.6 and 11.5, below it
-the bar is finer than this
-measurement can resolve
-solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-
+Wall time across six configurations, and the kill criterion Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it. CLAIM 3 THROUGHPUT Cost of determinism, two runs of the settled design 8 requests, 2972 tokens, median of 5 samples, every measurement in its own process SOURCE evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11 LOCKSTEP 0s 1s 2s 3s lockstep, fast mode 2.930 2.823 spread 1.24x lockstep, invariant 3.226 2.957 spread 1.17x vLLM default, eager 0.470 0.490 spread 1.27x vLLM batch-invariant, eager 0.563 0.574 spread 1.15x vLLM default, CUDA graphs 0.343 0.341 spread 1.08x vLLM batch-invariant, graphs 0.484 0.475 spread 1.27x KILL CRITERION set in week one, before anything was built: below 15 percent of vLLM default 0% 5% 10% 15% 20% A B 14.6 16.6 10.6 11.5 eager 14.6 and 16.6, straddling graphs 10.6 and 11.5, below it the bar is finer than this measurement can resolve solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-vLLM certification across seven boundary cases
-Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CERTIFICATION
-Black-box differential testing of vLLM batch-invariant mode
-concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared
-SOURCE
-evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187
-LOCKSTEP
-BOUNDARY CASE
-CO-RESIDENT
-VERDICT
-prefix_len == block_size - 1
-15
-clean
-prefix_len == block_size
-15
-clean
-prefix_len == block_size + 1
-15
-clean
-zero-prefix co-batched with nonzero
-16
-clean
-cache hit covering the full prompt
-15
-clean
-batch 31, shared prefix of one block
-44
-diverged
-max delta 4.685e-02
-batch 32, shared prefix of one block
-45
-diverged
-max delta 3.906e-02
-CONTROL
-one knob, and the only controlled single-variable result in this section
---max-num-seqs 8
-8 co-resident
-1 of 5, reproducible
---max-num-seqs 64
-45 co-resident
-not reproducible
---max-num-seqs 128, the default
-45 co-resident
-not reproducible, intermittently
-intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-
+vLLM certification across seven boundary cases Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8. CERTIFICATION Black-box differential testing of vLLM batch-invariant mode concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared SOURCE evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187 LOCKSTEP BOUNDARY CASE CO-RESIDENT VERDICT prefix_len == block_size - 1 15 clean prefix_len == block_size 15 clean prefix_len == block_size + 1 15 clean zero-prefix co-batched with nonzero 16 clean cache hit covering the full prompt 15 clean batch 31, shared prefix of one block 44 diverged max delta 4.685e-02 batch 32, shared prefix of one block 45 diverged max delta 3.906e-02 CONTROL one knob, and the only controlled single-variable result in this section --max-num-seqs 8 8 co-resident 1 of 5, reproducible --max-num-seqs 64 45 co-resident not reproducible --max-num-seqs 128, the default 45 co-resident not reproducible, intermittently intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-vLLM certification across seven boundary cases
-Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CERTIFICATION
-Black-box differential testing of vLLM batch-invariant mode
-concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared
-SOURCE
-evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187
-LOCKSTEP
-BOUNDARY CASE
-CO-RESIDENT
-VERDICT
-prefix_len == block_size - 1
-15
-clean
-prefix_len == block_size
-15
-clean
-prefix_len == block_size + 1
-15
-clean
-zero-prefix co-batched with nonzero
-16
-clean
-cache hit covering the full prompt
-15
-clean
-batch 31, shared prefix of one block
-44
-diverged
-max delta 4.685e-02
-batch 32, shared prefix of one block
-45
-diverged
-max delta 3.906e-02
-CONTROL
-one knob, and the only controlled single-variable result in this section
---max-num-seqs 8
-8 co-resident
-1 of 5, reproducible
---max-num-seqs 64
-45 co-resident
-not reproducible
---max-num-seqs 128, the default
-45 co-resident
-not reproducible, intermittently
-intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-
+vLLM certification across seven boundary cases Seven boundary cases against vLLM batch-invariant mode: five clean, and two diverged at 44 and 45 co-resident with logprob deltas of 4.685e-02 and 3.906e-02 and no token divergence. The max-num-seqs control restores reproducibility at 8. CERTIFICATION Black-box differential testing of vLLM batch-invariant mode concurrent submission, co-residency read from the engine's own gauge, every pair of repeats compared SOURCE evidence/certify-pairs-{a,b,mns8}.json :: vllm-project/vllm#51187 LOCKSTEP BOUNDARY CASE CO-RESIDENT VERDICT prefix_len == block_size - 1 15 clean prefix_len == block_size 15 clean prefix_len == block_size + 1 15 clean zero-prefix co-batched with nonzero 16 clean cache hit covering the full prompt 15 clean batch 31, shared prefix of one block 44 diverged max delta 4.685e-02 batch 32, shared prefix of one block 45 diverged max delta 3.906e-02 CONTROL one knob, and the only controlled single-variable result in this section --max-num-seqs 8 8 co-resident 1 of 5, reproducible --max-num-seqs 64 45 co-resident not reproducible --max-num-seqs 128, the default 45 co-resident not reproducible, intermittently intermittent at about one lifetime in three, so which cases trip is not the claim. token ids never move, only logprobs. filed as vllm#51187.
-Claims I1 to I4 and F1
-Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIMS
-Five relations, and what measures each one
-every claim scoped to one environment tuple, bitwise unless a tolerance is stated
-SOURCE
-evidence/verify-0002.json :: evidence/fidelity-0001.json
-LOCKSTEP
-I1
-Batch invariancebit-identical to canonical, whatever cohabits
-MR1, MR5 · fp16 logit bytes · batch 1 to 32
-holds
-I2
-Schedule invarianceany preemption, chunk partition, eviction, cache hit
-MR2, MR3, MR4 · per-layer KV equality
-holds
-I3
-Replay determinismsame (W, sigma, seeds), same trajectory hash
-MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED
-holds
-I4
-RNG isolationtokens from (seed, uid, position) and own logits only
-MR7 · 11 perturbations of the cohabitant set
-holds
-F1
-Fidelitybatch-1 logits against an fp64 CPU reference
-exact KL over 151936 tokens at 2756 positions
-7 of 7
-every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-
+Claims I1 to I4 and F1 Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference. CLAIMS Five relations, and what measures each one every claim scoped to one environment tuple, bitwise unless a tolerance is stated SOURCE evidence/verify-0002.json :: evidence/fidelity-0001.json LOCKSTEP I1 Batch invariancebit-identical to canonical, whatever cohabits MR1, MR5 · fp16 logit bytes · batch 1 to 32 holds I2 Schedule invarianceany preemption, chunk partition, eviction, cache hit MR2, MR3, MR4 · per-layer KV equality holds I3 Replay determinismsame (W, sigma, seeds), same trajectory hash MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED holds I4 RNG isolationtokens from (seed, uid, position) and own logits only MR7 · 11 perturbations of the cohabitant set holds F1 Fidelitybatch-1 logits against an fp64 CPU reference exact KL over 151936 tokens at 2756 positions 7 of 7 every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-Claims I1 to I4 and F1
-Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIMS
-Five relations, and what measures each one
-every claim scoped to one environment tuple, bitwise unless a tolerance is stated
-SOURCE
-evidence/verify-0002.json :: evidence/fidelity-0001.json
-LOCKSTEP
-I1
-Batch invariancebit-identical to canonical, whatever cohabits
-MR1, MR5 · fp16 logit bytes · batch 1 to 32
-holds
-I2
-Schedule invarianceany preemption, chunk partition, eviction, cache hit
-MR2, MR3, MR4 · per-layer KV equality
-holds
-I3
-Replay determinismsame (W, sigma, seeds), same trajectory hash
-MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED
-holds
-I4
-RNG isolationtokens from (seed, uid, position) and own logits only
-MR7 · 11 perturbations of the cohabitant set
-holds
-F1
-Fidelitybatch-1 logits against an fp64 CPU reference
-exact KL over 151936 tokens at 2756 positions
-7 of 7
-every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-
+Claims I1 to I4 and F1 Five claims with what verifies each. I1 batch invariance, I2 schedule invariance, I3 replay determinism and I4 RNG isolation all hold. F1 fidelity passes 7 of 7 bounds against an fp64 CPU reference. CLAIMS Five relations, and what measures each one every claim scoped to one environment tuple, bitwise unless a tolerance is stated SOURCE evidence/verify-0002.json :: evidence/fidelity-0001.json LOCKSTEP I1 Batch invariancebit-identical to canonical, whatever cohabits MR1, MR5 · fp16 logit bytes · batch 1 to 32 holds I2 Schedule invarianceany preemption, chunk partition, eviction, cache hit MR2, MR3, MR4 · per-layer KV equality holds I3 Replay determinismsame (W, sigma, seeds), same trajectory hash MR6 · 8 shapes · cross-process, differing PYTHONHASHSEED holds I4 RNG isolationtokens from (seed, uid, position) and own logits only MR7 · 11 perturbations of the cohabitant set holds F1 Fidelitybatch-1 logits against an fp64 CPU reference exact KL over 151936 tokens at 2756 positions 7 of 7 every claim is scoped to the environment tuple in env.lock. a claim without one is invalid by construction.
-Lifecycle n-gram coverage against the corrected denominator
-19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-COVERAGE
-Lifecycle n-grams, against the real denominator
-reported by population, because a case built to reach a transition is not exploration
-SOURCE
-evidence/fuzz-0002.json :: scripts/verify_no_gpu.py
-LOCKSTEP
-2-GRAMS
-19 of 25
-probes reach 20
-3-GRAMS
-43 of 79
-probes reach 46
-DENOMINATOR
-27 and 84
-->
-25 and 79
-one declared transition was never reachable
-every percentage published before the correction was computed against a denominator that was too large.
-the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here.
-swarm campaign, 72 cases
-plus eviction, 132 cases
-targeted probe only
-never reached
-
+Lifecycle n-gram coverage against the corrected denominator 19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84. COVERAGE Lifecycle n-grams, against the real denominator reported by population, because a case built to reach a transition is not exploration SOURCE evidence/fuzz-0002.json :: scripts/verify_no_gpu.py LOCKSTEP 2-GRAMS 19 of 25 probes reach 20 3-GRAMS 43 of 79 probes reach 46 DENOMINATOR 27 and 84 -> 25 and 79 one declared transition was never reachable every percentage published before the correction was computed against a denominator that was too large. the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here. swarm campaign, 72 cases plus eviction, 132 cases targeted probe only never reached
-Lifecycle n-gram coverage against the corrected denominator
-19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-COVERAGE
-Lifecycle n-grams, against the real denominator
-reported by population, because a case built to reach a transition is not exploration
-SOURCE
-evidence/fuzz-0002.json :: scripts/verify_no_gpu.py
-LOCKSTEP
-2-GRAMS
-19 of 25
-probes reach 20
-3-GRAMS
-43 of 79
-probes reach 46
-DENOMINATOR
-27 and 84
-->
-25 and 79
-one declared transition was never reachable
-every percentage published before the correction was computed against a denominator that was too large.
-the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here.
-swarm campaign, 72 cases
-plus eviction, 132 cases
-targeted probe only
-never reached
-
+Lifecycle n-gram coverage against the corrected denominator 19 of 25 two-grams and 43 of 79 three-grams reached, split between the swarm campaign and the eviction campaign, with probe-only cells outlined and never-reached cells hatched. The denominator was corrected from 27 and 84. COVERAGE Lifecycle n-grams, against the real denominator reported by population, because a case built to reach a transition is not exploration SOURCE evidence/fuzz-0002.json :: scripts/verify_no_gpu.py LOCKSTEP 2-GRAMS 19 of 25 probes reach 20 3-GRAMS 43 of 79 probes reach 46 DENOMINATOR 27 and 84 -> 25 and 79 one declared transition was never reachable every percentage published before the correction was computed against a denominator that was too large. the correction moves them up, 55.6 to 60.0 percent on 2-grams at identical observed counts, which is exactly why it is stated here. swarm campaign, 72 cases plus eviction, 132 cases targeted probe only never reached
-Mutation campaign, with and without golden bytes
-Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 2 HARNESS POWER
-What the harness catches, and what took a third observer
-identical operator set, identical campaign, one mutant of difference
-SOURCE
-evidence/fuzz-0002.json
-LOCKSTEP
-invariance relations and F1 only
-the fold reversal is not a function of the schedule,
-so the engine agrees with itself and both pass
-9 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-with golden bytes
-a committed sha256 over raw fp16 logit bytes,
-compared exactly, from outside the process
-10 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-the one mutant that separates the two rows:
-reversed split-combine fold
-every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed,
-patch executed and fault injected are three claims, and the counters can only tell you about the first.
-
+Mutation campaign, with and without golden bytes Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process. CLAIM 2 HARNESS POWER What the harness catches, and what took a third observer identical operator set, identical campaign, one mutant of difference SOURCE evidence/fuzz-0002.json LOCKSTEP invariance relations and F1 only the fold reversal is not a function of the schedule, so the engine agrees with itself and both pass 9 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 with golden bytes a committed sha256 over raw fp16 logit bytes, compared exactly, from outside the process 10 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 the one mutant that separates the two rows: reversed split-combine fold every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed, patch executed and fault injected are three claims, and the counters can only tell you about the first.
-Mutation campaign, with and without golden bytes
-Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 2 HARNESS POWER
-What the harness catches, and what took a third observer
-identical operator set, identical campaign, one mutant of difference
-SOURCE
-evidence/fuzz-0002.json
-LOCKSTEP
-invariance relations and F1 only
-the fold reversal is not a function of the schedule,
-so the engine agrees with itself and both pass
-9 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-with golden bytes
-a committed sha256 over raw fp16 logit bytes,
-compared exactly, from outside the process
-10 of 10 killed
-O1
-O2
-O3
-O4
-O5
-O6
-O7
-O8
-O9
-O10
-the one mutant that separates the two rows:
-reversed split-combine fold
-every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed,
-patch executed and fault injected are three claims, and the counters can only tell you about the first.
-
+Mutation campaign, with and without golden bytes Ten mutation operators. Nine killed by the invariance relations and F1 alone, with the reversed split-combine fold surviving. Ten killed once golden bytes are added as an observer outside the process. CLAIM 2 HARNESS POWER What the harness catches, and what took a third observer identical operator set, identical campaign, one mutant of difference SOURCE evidence/fuzz-0002.json LOCKSTEP invariance relations and F1 only the fold reversal is not a function of the schedule, so the engine agrees with itself and both pass 9 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 with golden bytes a committed sha256 over raw fp16 logit bytes, compared exactly, from outside the process 10 of 10 killed O1 O2 O3 O4 O5 O6 O7 O8 O9 O10 the one mutant that separates the two rows: reversed split-combine fold every operator fires a sentinel from inside the mutant body, so no trial here is a mutant that never ran. path executed, patch executed and fault injected are three claims, and the counters can only tell you about the first.
-Seventeen findings, laned by who found them
-Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-THE THESIS, ON ITS AUTHOR
-Seventeen times this repository declared more surface than it tested
-placed by finding index, laned by who found it
-SOURCE
-README.md :: tests/test_thesis_table.py
-LOCKSTEP
-FOUND FROM OUTSIDE
-8
-own machinery
-a run contradicting a declaration
-5
-outside audit
-commissioned, looking for defects
-1
-an anomaly
-predicted by neither side
-2
-readers
-not looking for defects at all
-2
-self-audit, later
-once it knew where to look
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-finding index
-eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only
-reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-
+Seventeen findings, laned by who found them Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository. THE THESIS, ON ITS AUTHOR Seventeen times this repository declared more surface than it tested placed by finding index, laned by who found it SOURCE README.md :: tests/test_thesis_table.py LOCKSTEP FOUND FROM OUTSIDE 8 own machinery a run contradicting a declaration 5 outside audit commissioned, looking for defects 1 an anomaly predicted by neither side 2 readers not looking for defects at all 2 self-audit, later once it knew where to look 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 finding index eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-Seventeen findings, laned by who found them
-Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-THE THESIS, ON ITS AUTHOR
-Seventeen times this repository declared more surface than it tested
-placed by finding index, laned by who found it
-SOURCE
-README.md :: tests/test_thesis_table.py
-LOCKSTEP
-FOUND FROM OUTSIDE
-8
-own machinery
-a run contradicting a declaration
-5
-outside audit
-commissioned, looking for defects
-1
-an anomaly
-predicted by neither side
-2
-readers
-not looking for defects at all
-2
-self-audit, later
-once it knew where to look
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-1
-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12
-12b
-13
-14
-15
-16
-17
-finding index
-eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only
-reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-
+Seventeen findings, laned by who found them Eighteen tiles for seventeen findings, laned by who found them: eight from this project's own machinery, five from an outside audit, one from an anomaly, two from readers not looking for defects, and two from this project auditing itself later. Findings 9 to 15 came from outside the repository. THE THESIS, ON ITS AUTHOR Seventeen times this repository declared more surface than it tested placed by finding index, laned by who found it SOURCE README.md :: tests/test_thesis_table.py LOCKSTEP FOUND FROM OUTSIDE 8 own machinery a run contradicting a declaration 5 outside audit commissioned, looking for defects 1 an anomaly predicted by neither side 2 readers not looking for defects at all 2 self-audit, later once it knew where to look 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 1 2 3 4 5 6 7 8 9 10 11 12 12b 13 14 15 16 17 finding index eighteen tiles, seventeen findings: 12b is a row and not a count, because the audit named it outright and the witness check only reproduced it from a different direction. counting it as new would inflate the one table whose only value is that it does not.
-The three headline numbers
-65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-HEADLINE
-The three numbers
-invariance, harness power, and the price of the constraint
-SOURCE
-evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json
-LOCKSTEP
-INVARIANCE
-65/ 65
-relation runs, bitwise identical
-13 relations across 5 block sizes
-MR1 to MR8, PATH-EQ, EOS finish
-HARNESS POWER
-10/ 10
-seeded faults killed
-0 equivalent, 0 not-exercised
-median time to detection 10.6 s
-COST
-5.2to 5.7
-0x
-2x
-4x
-6x
-x vLLM batch-invariant, eager
-1.05x to 1.10x vs its own fast path
-two runs, reported as a range
-
+The three headline numbers 65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode. HEADLINE The three numbers invariance, harness power, and the price of the constraint SOURCE evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json LOCKSTEP INVARIANCE 65/ 65 relation runs, bitwise identical 13 relations across 5 block sizes MR1 to MR8, PATH-EQ, EOS finish HARNESS POWER 10/ 10 seeded faults killed 0 equivalent, 0 not-exercised median time to detection 10.6 s COST 5.2to 5.7 0x 2x 4x 6x x vLLM batch-invariant, eager 1.05x to 1.10x vs its own fast path two runs, reported as a range
-The three headline numbers
-65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-HEADLINE
-The three numbers
-invariance, harness power, and the price of the constraint
-SOURCE
-evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json
-LOCKSTEP
-INVARIANCE
-65/ 65
-relation runs, bitwise identical
-13 relations across 5 block sizes
-MR1 to MR8, PATH-EQ, EOS finish
-HARNESS POWER
-10/ 10
-seeded faults killed
-0 equivalent, 0 not-exercised
-median time to detection 10.6 s
-COST
-5.2to 5.7
-0x
-2x
-4x
-6x
-x vLLM batch-invariant, eager
-1.05x to 1.10x vs its own fast path
-two runs, reported as a range
-
+The three headline numbers 65 of 65 relation runs bitwise identical, 10 of 10 seeded faults killed with none equivalent, and lockstep invariant running at 5.2 to 5.7 times the wall time of vLLM batch-invariant in eager mode. HEADLINE The three numbers invariance, harness power, and the price of the constraint SOURCE evidence/verify-0002.json :: fuzz-0002.json :: throughput-0004.json LOCKSTEP INVARIANCE 65/ 65 relation runs, bitwise identical 13 relations across 5 block sizes MR1 to MR8, PATH-EQ, EOS finish HARNESS POWER 10/ 10 seeded faults killed 0 equivalent, 0 not-exercised median time to detection 10.6 s COST 5.2to 5.7 0x 2x 4x 6x x vLLM batch-invariant, eager 1.05x to 1.10x vs its own fast path two runs, reported as a range
-Wall time across six configurations, and the kill criterion
-Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 3 THROUGHPUT
-Cost of determinism, two runs of the settled design
-8 requests, 2972 tokens, median of 5 samples, every measurement in its own process
-SOURCE
-evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11
-LOCKSTEP
-0s
-1s
-2s
-3s
-lockstep, fast mode
-2.930
-2.823
-spread 1.24x
-lockstep, invariant
-3.226
-2.957
-spread 1.17x
-vLLM default, eager
-0.470
-0.490
-spread 1.27x
-vLLM batch-invariant, eager
-0.563
-0.574
-spread 1.15x
-vLLM default, CUDA graphs
-0.343
-0.341
-spread 1.08x
-vLLM batch-invariant, graphs
-0.484
-0.475
-spread 1.27x
-KILL CRITERION
-set in week one, before anything was built: below 15 percent of vLLM default
-0%
-5%
-10%
-15%
-20%
-A
-B
-14.6
-16.6
-10.6
-11.5
-eager
-14.6 and 16.6, straddling
-graphs
-10.6 and 11.5, below it
-the bar is finer than this
-measurement can resolve
-solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-
+Wall time across six configurations, and the kill criterion Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it. CLAIM 3 THROUGHPUT Cost of determinism, two runs of the settled design 8 requests, 2972 tokens, median of 5 samples, every measurement in its own process SOURCE evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11 LOCKSTEP 0s 1s 2s 3s lockstep, fast mode 2.930 2.823 spread 1.24x lockstep, invariant 3.226 2.957 spread 1.17x vLLM default, eager 0.470 0.490 spread 1.27x vLLM batch-invariant, eager 0.563 0.574 spread 1.15x vLLM default, CUDA graphs 0.343 0.341 spread 1.08x vLLM batch-invariant, graphs 0.484 0.475 spread 1.27x KILL CRITERION set in week one, before anything was built: below 15 percent of vLLM default 0% 5% 10% 15% 20% A B 14.6 16.6 10.6 11.5 eager 14.6 and 16.6, straddling graphs 10.6 and 11.5, below it the bar is finer than this measurement can resolve solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-Wall time across six configurations, and the kill criterion
-Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-CLAIM 3 THROUGHPUT
-Cost of determinism, two runs of the settled design
-8 requests, 2972 tokens, median of 5 samples, every measurement in its own process
-SOURCE
-evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11
-LOCKSTEP
-0s
-1s
-2s
-3s
-lockstep, fast mode
-2.930
-2.823
-spread 1.24x
-lockstep, invariant
-3.226
-2.957
-spread 1.17x
-vLLM default, eager
-0.470
-0.490
-spread 1.27x
-vLLM batch-invariant, eager
-0.563
-0.574
-spread 1.15x
-vLLM default, CUDA graphs
-0.343
-0.341
-spread 1.08x
-vLLM batch-invariant, graphs
-0.484
-0.475
-spread 1.27x
-KILL CRITERION
-set in week one, before anything was built: below 15 percent of vLLM default
-0%
-5%
-10%
-15%
-20%
-A
-B
-14.6
-16.6
-10.6
-11.5
-eager
-14.6 and 16.6, straddling
-graphs
-10.6 and 11.5, below it
-the bar is finer than this
-measurement can resolve
-solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.
-
+Wall time across six configurations, and the kill criterion Wall time for six configurations, run A solid and run B faded: lockstep fast and invariant, and vLLM default and batch-invariant, each eager and graphed. Below, the 15 percent kill criterion, with the eager comparison at 14.6 and 16.6 percent straddling the bar and the graphed comparison at 10.6 and 11.5 percent consistently below it. CLAIM 3 THROUGHPUT Cost of determinism, two runs of the settled design 8 requests, 2972 tokens, median of 5 samples, every measurement in its own process SOURCE evidence/throughput-0004.json :: docs/kickoff/01-PRD.md s11 LOCKSTEP 0s 1s 2s 3s lockstep, fast mode 2.930 2.823 spread 1.24x lockstep, invariant 3.226 2.957 spread 1.17x vLLM default, eager 0.470 0.490 spread 1.27x vLLM batch-invariant, eager 0.563 0.574 spread 1.15x vLLM default, CUDA graphs 0.343 0.341 spread 1.08x vLLM batch-invariant, graphs 0.484 0.475 spread 1.27x KILL CRITERION set in week one, before anything was built: below 15 percent of vLLM default 0% 5% 10% 15% 20% A B 14.6 16.6 10.6 11.5 eager 14.6 and 16.6, straddling graphs 10.6 and 11.5, below it the bar is finer than this measurement can resolve solid bar run A, faded bar run B. two runs do not establish a distribution, which is why both are drawn.