Historical source. Some claims in older records were subsequently corrected. The associated article states the adopted interpretation. This record preserves the original source alongside its rendered reading view.
Rendered archival Markdown
This reading view preserves headings, tables, lists, code fragments and mathematical notation from the local research record.
Bounded live capture succeeds but the full long-stream answer gate fails
The fixed live accumulator and retained-byte budgets pass on every case. The unchanged posterior-TV bound passes all eight million-sample cases but only two of eight four-million-sample cases. The full experiment therefore fails its joint gate and closes without changing lag, arithmetic precision, model, grid or tolerance. This is a measured operating boundary of the present implementation, not an impossibility theorem for the vision.
Complete result
The two new seed identities, two sampling intervals and two paired prefix lengths produce eight recorded prefixes. Each receives both calibration queries, with uniform and biased priors: 16 record/calibration cases and 32 posterior comparisons. Four streams are represented; paired lengths are not independent replications. No full raw array or raw file is saved during this experiment. The later query process reads only capsules and metadata.
| Observations | Sampling interval | Bound pass cases | Largest TV bound |
|---|
| 1,048,576 | 0.01 | 4/4 | 0.00444167 |
| 1,048,576 | 0.05 | 4/4 | 0.00304670 |
| 4,194,304 | 0.01 | 0/4 | 0.01843085 |
| 4,194,304 | 0.05 | 2/4 | 0.01220511 |
At the longer dt=0.05 prefix, both original-calibration cases pass and both revised-calibration cases are rejected. The first tested rejected length is 4,194,304; intermediate lengths were not searched. All 20,336 floating full-record query references are enclosed, and all 32 measured posterior discrepancies are below their respective bounds. The largest measured TV against the binary64 reference is 5.37358e-6, and the largest log-error oscillation is 0.000574618. Small measured errors do not override rejection by the prescribed bound. The online reference is not an independently formally verified exact-arithmetic solver.
The saved finite-conditioning likelihood intervals have maximum widths of 0.0453-0.0704 on the rejected cases, above 4*atanh(0.01), approximately 0.0400013, before adding the physical truncation envelope. Their midpoint-error range therefore exhausts the chosen 0.01 posterior budget. The truncation envelope also contributes in some cases. This is an accounting observation from saved intervals, not a new precision/lag ablation or evidence that the observed errors are that large. No empirical safety factor, discarded hypothesis or posterior-specific rescue replaces the gate.
Memory and cost
Native accumulator allocation stays at 144,016 bytes for every prefix, below 256 KiB. Input blocks never exceed 4,096 observations. Capsules use 12,700-12,703 bytes including metadata, uncertainty radii and integrity hash, below 16 KiB. Capture process peak RSS is 117.20 MiB and query peak RSS is 115.42 MiB, below the 512 MiB limit. Native allocation is not total process memory or a measured embedded-device footprint.
Counting each of the four streams once through its final prefix, candidate append work takes 6.12743 seconds. Source generation takes 0.39907 seconds, online reference updates 485.20627 seconds, and stream hashing 0.09004 seconds. Total capture-phase elapsed time is 491.89481 seconds. The reference keeps extra offset-polynomial statistics and is not an optimized minimal answer bank; its time must not be promoted into a universal capture speed ratio.
The separate numerical query phase takes 2,909.87649 seconds, about 48.5 minutes. Model/prefix-schedule preparation costs 2,688.44189 seconds, record-prefix likelihoods 199.38390 seconds, and capsule-tail calculations 21.70059 seconds, with the remainder overhead. This includes shared model work and prefix reuse across paired lengths. It is not the ordinary fast inference timing reported in study 03, and it is not a device-energy result.
For perspective, an uncompressed binary64 payload at the shorter prefix would occupy 8,388,608 bytes, roughly 660 times the capsule size. At the longer prefix it would occupy 33,554,432 bytes, roughly 2,642 times the capsule size. These are payload arithmetic comparisons, not measured new lossless-codec results; the latter scale also fails the full answer gate. Study 03 remains the measured lossless/ordinary-bank comparison.
The serialized capsule is a query snapshot, not a restart image of the live binary-counter state. It does not support arbitrary corrections to discarded interior data, a certified merge of separate snapshots or unlimited duration at fixed precision. The executable accepts N<2^31 and the declared input magnitude domain; its useful answer bound can fail before that count limit.
Grid resolution and calibration interpretation
All 32 predeclared grid diagnostics are reported. Effective support size is only 1.0-1.854 grid points overall. At the longer dt=0.05 prefix, maximum posterior mass ranges from 0.999994657 to 1.0. The particularly small measured TV there is not evidence of resolved continuous physical uncertainty. Generating damping 0.43 lies between grid values 0.42 and 0.44, and no refinement or interpolation was performed. Boundary-grid mass is numerically zero in all cases, which does not remove the interior resolution problem.
The specified calibration revision changes the reference posterior mean frequency from approximately 3.15 to 3.10 and damping from approximately 0.560-0.580 to 0.420-0.440 across the reported prefixes, intervals and priors. The capsule reproduces that model-dependent reinterpretation closely without rereading interior observations. These are descriptive paired results already included in the declared outputs, not a new threshold-selected success or an identifiable calibration experiment on a thermal instrument. Ordinary banks can support model revision as well, with their own storage/capture tradeoffs.
Verification, development failures and closure
All 61 campaign tests and 12 subtests pass. This includes four multichannel algebra checks, not a multichannel numerical engine. The independent saved- archive integrity checker passes and correctly preserves the failed numerical gate. Earlier development issues remain documented: last-bit source mismatch from compile-time constants, corrected without weakening exact-equality tests; and a duplicate filename key in the not-yet-run grid diagnostic, corrected before reference/diagnostic execution without changing calculations.
Close this bounded scale-up. Do not tune the revealed cohort, enlarge the lag, adjust precision or count only the passing cases. The positive finding is bounded live acquisition plus later model-dependent inference at the tested million-sample scale under the stated numerical trust assumptions. The negative finding is failure of the same fixed configuration's full four-million-sample answer budget, with appreciable assurance cost and a coarse discrete posterior. A flagship scientific application remains open.
Evidence: PROTOCOL_06_LIVE_CAPTURE.md, DEVELOPMENT_06.md, results_06_long_stream/{START,CAPTURE,RESULTS,VERIFICATION}.json, GRID_DIAGNOSTICS_06.json, SUMMARY_06.json and UNIT_TESTS_FINAL.json. The public particle-archive metadata assessment is separately documented and has not admitted a physical-data fit.
View raw MD source
# Bounded live capture succeeds but the full long-stream answer gate fails
The fixed live accumulator and retained-byte budgets pass on every case.
The unchanged posterior-TV bound passes all eight million-sample cases but
only two of eight four-million-sample cases. The full experiment therefore
fails its joint gate and closes without changing lag, arithmetic precision,
model, grid or tolerance. This is a measured operating boundary of the
present implementation, not an impossibility theorem for the vision.
## Complete result
The two new seed identities, two sampling intervals and two paired prefix
lengths produce eight recorded prefixes. Each receives both calibration
queries, with uniform and biased priors: 16 record/calibration cases and
32 posterior comparisons. Four streams are represented; paired lengths are
not independent replications. No full raw array or raw file is saved during
this experiment. The later query process reads only capsules and metadata.
| Observations | Sampling interval | Bound pass cases | Largest TV bound |
| --- | --- | --- | --- |
| 1,048,576 | 0.01 | 4/4 | 0.00444167 |
| 1,048,576 | 0.05 | 4/4 | 0.00304670 |
| 4,194,304 | 0.01 | 0/4 | 0.01843085 |
| 4,194,304 | 0.05 | 2/4 | 0.01220511 |
At the longer dt=0.05 prefix, both original-calibration cases pass and both
revised-calibration cases are rejected. The first tested rejected length is
4,194,304; intermediate lengths were not searched. All 20,336 floating
full-record query references are enclosed, and all 32 measured posterior
discrepancies are below their respective bounds. The largest measured TV
against the binary64 reference is 5.37358e-6, and the largest log-error
oscillation is 0.000574618. Small measured errors do not override rejection
by the prescribed bound. The online reference is not an independently
formally verified exact-arithmetic solver.
The saved finite-conditioning likelihood intervals have maximum widths of
0.0453-0.0704 on the rejected cases, above 4*atanh(0.01), approximately
0.0400013, before adding the physical truncation envelope.
Their midpoint-error range therefore exhausts the chosen 0.01 posterior
budget. The truncation envelope also contributes in some cases. This is an
accounting observation from saved intervals, not a new precision/lag ablation
or evidence that the observed errors are that large. No empirical safety
factor, discarded hypothesis or posterior-specific rescue replaces the gate.
## Memory and cost
Native accumulator allocation stays at 144,016 bytes for every prefix,
below 256 KiB. Input blocks never exceed 4,096 observations. Capsules use
12,700-12,703 bytes including metadata, uncertainty radii and integrity hash,
below 16 KiB. Capture process peak RSS is 117.20 MiB and query peak RSS is
115.42 MiB, below the 512 MiB limit. Native allocation is not total process
memory or a measured embedded-device footprint.
Counting each of the four streams once through its final prefix, candidate
append work takes 6.12743 seconds. Source generation takes 0.39907 seconds,
online reference updates 485.20627 seconds, and stream hashing 0.09004 seconds.
Total capture-phase elapsed time is 491.89481 seconds. The reference keeps
extra offset-polynomial statistics and is not an optimized minimal answer
bank; its time must not be promoted into a universal capture speed ratio.
The separate numerical query phase takes 2,909.87649 seconds, about 48.5
minutes. Model/prefix-schedule preparation costs 2,688.44189 seconds,
record-prefix likelihoods 199.38390 seconds, and capsule-tail calculations
21.70059 seconds, with the remainder overhead. This includes shared model
work and prefix reuse across paired lengths. It is not the ordinary fast
inference timing reported in study 03, and it is not a device-energy result.
For perspective, an uncompressed binary64 payload at the shorter prefix
would occupy 8,388,608 bytes, roughly 660 times the capsule size. At the
longer prefix it would occupy 33,554,432 bytes, roughly 2,642 times the
capsule size. These are payload arithmetic comparisons, not measured new
lossless-codec results; the latter scale also fails the full answer gate.
Study 03 remains the measured lossless/ordinary-bank comparison.
The serialized capsule is a query snapshot, not a restart image of the live
binary-counter state. It does not support arbitrary corrections to discarded
interior data, a certified merge of separate snapshots or unlimited duration
at fixed precision. The executable accepts N<2^31 and the declared input
magnitude domain; its useful answer bound can fail before that count limit.
## Grid resolution and calibration interpretation
All 32 predeclared grid diagnostics are reported. Effective support size
is only 1.0-1.854 grid points overall. At the longer dt=0.05 prefix, maximum
posterior mass ranges from 0.999994657 to 1.0. The particularly small measured
TV there is not evidence of resolved continuous physical uncertainty.
Generating damping 0.43 lies between grid values 0.42 and 0.44, and no
refinement or interpolation was performed. Boundary-grid mass is numerically
zero in all cases, which does not remove the interior resolution problem.
The specified calibration revision changes the reference posterior mean
frequency from approximately 3.15 to 3.10 and damping from approximately
0.560-0.580 to 0.420-0.440 across the reported prefixes, intervals and priors.
The capsule reproduces that model-dependent reinterpretation closely without
rereading interior observations. These are descriptive paired results already
included in the declared outputs, not a new threshold-selected success or an
identifiable calibration experiment on a thermal instrument. Ordinary banks
can support model revision as well, with their own storage/capture tradeoffs.
## Verification, development failures and closure
All 61 campaign tests and 12 subtests pass. This includes four multichannel
algebra checks, not a multichannel numerical engine. The independent saved-
archive integrity checker passes and correctly preserves the failed numerical
gate. Earlier development issues remain documented: last-bit source mismatch
from compile-time constants, corrected without weakening exact-equality tests;
and a duplicate filename key in the not-yet-run grid diagnostic, corrected
before reference/diagnostic execution without changing calculations.
Close this bounded scale-up. Do not tune the revealed cohort, enlarge the
lag, adjust precision or count only the passing cases. The positive finding
is bounded live acquisition plus later model-dependent inference at the
tested million-sample scale under the stated numerical trust assumptions.
The negative finding is failure of the same fixed configuration's full
four-million-sample answer budget, with appreciable assurance cost and a
coarse discrete posterior. A flagship scientific application remains open.
Evidence: PROTOCOL_06_LIVE_CAPTURE.md, DEVELOPMENT_06.md,
results_06_long_stream/{START,CAPTURE,RESULTS,VERIFICATION}.json,
GRID_DIAGNOSTICS_06.json, SUMMARY_06.json and UNIT_TESTS_FINAL.json.
The public particle-archive metadata assessment is separately documented
and has not admitted a physical-data fit.