The Cryogenic Referent Registry

What this note is: the foundation document of this series. It fixes the six criteria every audit note applies, defines the verdict scale, lists the candidate experiments, records the verdict for each one audited so far, and carries SHA-256 provenance for every registry source artifact listed in §VIII; downstream-use artifacts examined by individual notes are hashed in their own source lists and in the companion dataset’s verification log and manifest. Notes 001, 002 and onward are close readings of individual entries in this registry; they cite these criteria rather than restating them.

It is a living document. New referents, corrected verdicts and errata produce new versions (v1.1, v1.2 …), each with its own version DOI, linked under the registry’s Zenodo concept DOI; every prior version is preserved and remains individually citable. The errata section below is part of the publication, not an apology appended to it.

What it is not: a criticism of the experimentalists — the audit found close to the opposite (see §V) — and not a claim that any model is wrong. A model can be excellent while its validation is unquantifiable, if the referent’s uncertainty was never carried along.

On the word “audit”: in this series it means a documentary audit — a structured examination of public records against stated criteria, performed by one auditor with AI assistance and no independent human review (§IX) — not a certified or independent engineering audit. Every use of the word in these notes inherits this definition.

I. The question

Computational models that predict self-pressurization and boil-off in cryogenic propellant tanks are used to size upper stages, to plan orbital propellant transfer, and increasingly to design liquid hydrogen storage for ships and aircraft. The modelling papers examined for this registry describe their models as validated against one or more of these experiments.

Validated against what, exactly — and how well was that reference measured?

A validation claim is a comparison between a simulated value S and a reference value D. The comparison error E = S − D supports a quantified accuracy claim only if the uncertainty attached to D is known; without it, E still permits coarser uses — detecting gross discrepancy, comparing trends — but not the percentage-accuracy statements this literature makes. If an experiment’s measurement uncertainty on tank pressure is ±10 % and a model reproduces that pressure to within 3 %, the model has not demonstrated 3 % accuracy — it has demonstrated only that model and data agree more closely than the applicable validation-uncertainty envelope can discriminate. The formal judgment weighs the observed difference against that envelope — which combines the experimental uncertainty with the simulation’s numerical and input contributions — not against the measurement uncertainty alone. “Agrees to 3 %” and “accurate to 3 %” are different statements, and only the second supports a design decision.

That decomposition is standard in the verification and validation literature; ASME V&V 20, per its published scope (the standard itself was not read for this series — Note 001, Sources B), quantifies the accuracy inferred from a solution–measurement comparison using the uncertainties of both the simulation and the data, and separates numerical uncertainty, input uncertainty, and the experimental uncertainty of the referent. Whether it is standard practice in the cryogenic tank modelling literature is an empirical question. This registry is an attempt to answer it with documents rather than opinion.

II. Method — criteria fixed before the audit; scoring refined in use

The six criteria — canonical wording

Every experiment in the registry is assessed against the same six questions. This wording is canonical as of v1.0; the notes in this series cite it rather than paraphrase it.

Are the raw data published, or only plotted in figures?
Are the boundary conditions complete — heat flux magnitude and distribution, fill level, initial stratification, ullage composition?
Are instrument accuracies stated, and are the derived quantities’ uncertainties propagated?
Is the geometry fully specified?
Is the fluid state specified — for hydrogen, including the ortho/para composition?
Does the quality information the original authors published travel with the data into current use?

How criteria 1 and 3 are scored

Note 002 established, by reading the most thoroughly documented of the referents audited here, that a binary answer to criteria 1 and 3 destroys information. Criterion 3 is therefore scored on an ordered scale, and criterion 1 as a data-availability profile, fixed here.

Criterion 1 — data availability profile. Machine-readable primary measurements · tabulated processed data · plotted-only data, carrying the digitization uncertainty that implies · archive referenced but access-controlled · archive publicly reachable · not located. A single referent can occupy several categories for different quantities; the score names them.

Criterion 3 — uncertainty maturity. Three separable levels, because a referent can be exemplary at one and absent at another: (a) uncertainty required in advance, in a test plan or acceptance criteria; (b) uncertainty characterized per instrument, reported for the sensors under the applicable conditions; (c) uncertainty assembled per reported result, propagated into intervals on the headline numbers a modeler would actually use. Level (c) is the one that yields u_D for the derived quantities a validation actually compares, and it is the level most often missing. For a directly measured validation quantity, level (b) may supply the dominant contribution — when the reported figure genuinely characterizes that measurement under the applicable test conditions — but an instrument accuracy specification is not automatically the complete experimental uncertainty: calibration, acquisition, installation, environment, repeatability and averaging can all add to it.

Criteria 2 and 4 are relative to an intended use

“Complete” and “fully specified” are not answerable in the abstract. A boundary-condition set sufficient for a lumped model may be inadequate for a CFD model resolving wall boundary layers. Where this registry scores criteria 2 and 4, the score is against the intended use stated in the entry, and where no use is stated the score is provisional.

Criterion 5 is not bookkeeping

The ortho→para conversion in hydrogen releases heat of the same order as — and exceeding — the latent heat of vaporization. A liquid hydrogen tank whose spin composition is not at equilibrium warms itself from the inside. The downstream models examined in this series work with parahydrogen properties, which is physically reasonable for stored LH₂; the question the criterion asks is whether that assumption is traceable to anything in the experimental record.

Criterion 6 is not the verdict

Criterion 6 asks whether the original authors’ quality information reached current use — a question about the downstream literature. The verdict scale below asks whether the experiment can be reconstructed from the public record — a question about the source documents. They are independent: a referent can be fully reconstructable and still have its caveats dropped by everyone who cites it, which is the central finding of this audit.

Provenance of the criteria, stated honestly

The criteria have a version history, and hiding it would be inconsistent with the point of this series.

Stage	Where	What changed
Fixed before the audit	Three-referent pilot; unpublished method draft, preserved in the project record marked superseded and not to be cited	Six criteria in substantially this form
Published	Note 001 (doi:10.5281/zenodo.21895605) §IX, 11 Aug 2026	The six criteria as stated above, criteria 1–6 unchanged in substance
Refined by field use	Note 002 (doi:10.5281/zenodo.21895647) §V, 11 Aug 2026	Note 002 argued that criteria 1 and 3 were too coarse and committed the audit to the finer-grained scoring now fixed above
Frozen	This note, v1.0	The wordings of Notes 000 and 001 aligned; the published texts are identical (Erratum 6 records the alignment)

The criteria were not fixed after seeing the results. They were applied unchanged in the pilot that preceded the audit and published in Note 001. What changed afterwards was resolution, not direction: criteria 1 and 3 were subdivided because binary answers were losing information, in a direction that makes referents harder to score well, not easier. No criterion was added, dropped, or reversed.

This registry nonetheless publishes method and verdicts in the same document. The protection against choosing criteria to fit results therefore rests on the audit trail — the pilot record, the per-referent audit files, the verification log, and Note 001’s prior publication of the criteria — and not on publication order alone. Future changes require a new revision of this note stating the change and its reason.

A distinction worth keeping clean. Specifying criteria in advance is not the same instrument as committing to a prediction in advance, and this registry claims only the first. An audit is a census: it enumerates what the record contains, and its researcher degrees of freedom are in what to include and how to score — which is what criteria pre-specification constrains. A prediction commitment is a stronger and different thing: stating what a model will produce before the referent data are examined, sealed with a third-party timestamp, resolvable as confirmed or refuted. That mechanism is reserved for the model-comparison work this registry enables, where it means something and cannot be reconstructed after the fact. Describing a census as though it were a prediction would devalue the stronger instrument before it is used. Nothing in this note is a prediction commitment.

The verdict scale

Verdict	Meaning
REPRODUCIBLE	The validation-relevant content of the experiment can be reconstructed from the public record alone.
MOSTLY_REPRODUCIBLE	Reconstruction is possible with minor, explicitly listed gaps.
PARTIALLY_REPRODUCIBLE	Substantial reconstruction is possible, but at least one element a modeler needs — a boundary condition, an uncertainty, a dependency document — is missing, ambiguous, or lives outside the cited record.
NOT_REPRODUCIBLE (from the cited document)	The document the validation literature actually cites does not permit reconstruction. The verdict attaches to the cited document, not to the experiment; a located primary source can upgrade it.
PENDING	Primary source not yet obtained or not yet audited.

The boundary between MOSTLY and PARTIALLY is, at v1.0, a judgment about whether a missing element is validation-critical for the declared intended use. A formal decision rule is committed for v1.1, and the existing verdicts will be re-scored under it. Where a verdict carries the modifier (high), it marks proximity to the next band upward under that judgment — an interim flag the v1.1 rule replaces. Until then, each per-record audit file states which missing elements drove its verdict.

A verdict is a statement about the public record, never about the competence of the people who produced it.

Framework references

Cited as framework, not as audited evidence; none of these is among this registry’s source artifacts.

ASME V&V 20-2009 (R2021), Standard for Verification and Validation in Computational Fluid Dynamics and Heat Transfer. Source of the E = S − D decomposition and the validation-uncertainty framing used throughout.
Leachman, J.W., Jacobsen, R.T., Penoncello, S.G., Lemmon, E.W. (2009), “Fundamental Equations of State for Parahydrogen, Normal Hydrogen, and Orthohydrogen”, J. Phys. Chem. Ref. Data 38(3): 721–748. Property basis distinguishing the spin isomers.
McCarty, R.D., Hord, J., Roder, H.M. (1981), Selected Properties of Hydrogen (Engineering Design Data), NBS Monograph 168. Ortho–para conversion energetics against the heat of vaporization, the physical basis of criterion 5.

III. The registry

Nineteen candidate sources in the working list (the sixteen experiments tabulated below, plus the Stochl–Knoll boundary-condition report and SHIIVER’s test-plan and heat-flux-sensor companions, tracked as separate sources there), drawn from the bibliographies of the validation literature itself: an experiment enters the list if at least one of four modelling/validation papers (NTRS 20220018548, 20240006348, 20150000249, 20090007814) uses it as comparison data. That is the operational definition used here, and it is checkable by anyone. One of the ten audited records, A1d, is an experiment in its own right — the Stochl–Knoll thermal-performance test — whose role in this registry is to supply the boundary condition A1 and A2 depend on.

Two weaknesses in that definition, stated up front.

The four source papers are not independent. Three of them share an author — Kassemi — and the same institution, the National Center for Space Exploration Research at NASA Glenn: Kartuzova & Kassemi (2015, with Agui and Moder; and 2025), and Barsi, Moder & Kassemi (2008). Only Yang, Patel & Williams (2022, NASA Marshall / Jacobs / CFD Research Corporation) is institutionally independent of the other three. So this list reflects, to a substantial degree, the citation practice of one research group plus one other team — not a survey of the field. The phrase “the experiments everyone uses” would overstate what has been established, and is not used.

The same documents do two jobs. These four papers define which referents count as canonical here, and three of them are the evidence in Note 001 that the validation literature does not carry experimental uncertainty forward. Those are therefore not independent observations drawn from independent samples; they are two readings of one small, related set of documents. This does not make either reading false — the papers do cite those experiments, and they do lack statements of experimental uncertainty — but it means neither can be used to corroborate the other.

Widening the source set with institutionally independent groups, including non-NASA and non-US work, is the first task of v1.1.

A. Ground, liquid hydrogen, large scale

#	Experiment	Primary source	NTRS	Verdict	Key finding
A1	K-Site, low heat flux (0.35 / 2.0 / 3.5 W/m², 83–84 % fill)	Hasan, Lin & Van Dresar, NASA TM-103804 (1991), 8 pp.	19910011011	PARTIALLY_REPRODUCIBLE	Instrument accuracies fully stated; but the “wall heat flux” is total heat over area (see A1d), and its uncertainty — ±3.1 % to ±6.7 % for the lowest-flux case, depending on which of the two conflicting flowmeter-accuracy statements is used — is never propagated.
A1d	K-Site boundary-condition document (MLI thermal performance)	Stochl & Knoll, NASA TM-104476 / AIAA 91-2400 (1991), 21 pp.	19910015845	REPRODUCIBLE	Defines the heat rates behind A1/A2. 13–17 % of tank heat enters through discrete penetrations (12 struts, plumbing, instrumentation), not through the insulation — a distinction neither K-Site test paper flags.
A2	K-Site, fill-level effects (29 / 49 / 83 %)	Van Dresar, Lin & Hasan, NASA TM-105411 / AIAA-92-0818 (1992), 11 pp.	19920009200	PARTIALLY_REPRODUCIBLE	Master table of all seven K-Site cases. Authors report a 120–150 % energy balance error, bound the named instrument contributions as insufficient (not excluded — Erratum 27), and question their own uniform-radial-temperature assumption; they also flag the 29 %-fill rows as possibly not quasi-steady, and one figure contains a calculated point substituted for a failed sensor. See Note 001.
A3	MHTB spray-bar TVS (18.09 m³ tank)	Hastings et al., NASA TM-2003-212926, 161 pp.; companion: Hedayat et al., NTRS 20030106051, 8 pp.	20040000092	PARTIALLY_REPRODUCIBLE	A thermodynamic-vent-system report, not a self-pressurization characterization: lockup phases are interleaved between mixing cycles. Model comparisons against 1998 tests use time-shifted curves (e.g. by 60,350 s). Authors document an energy discrepancy and explicitly rule out instrumentation uncertainties as the cause.
A4	MHTB variable-density MLI on foam substrate	Martin & Hastings, NASA TM-2001-211089, 90 pp.	20010063699	PARTIALLY_REPRODUCIBLE (high)	Insulation pedigree document for A3. Installed MLI thickness “cannot be accurately ascertained” / “cannot be accurately verified” — the authors’ own words — which caps the fidelity of any thermal model built on it.
A5	SHIIVER (4 m tank, LH₂ and LN₂, 2019–20)	NASA/TP-20205008233, 286 pp. (+ test plan 20205003433, heat-flux paper 20210019122)	20205008233	MOSTLY_REPRODUCIBLE	Best-documented referent in the registry: 74 numbered tables, parahydrogen properties tabulated, sensor-level uncertainty requirements in the test plan, selective uncertainty tracing in the report. Heat-flux sensors carry ~50 % uncertainty (stated in the support paper). No consolidated uncertainty budget. See Note 002.

B. Ground, other fluids

#	Experiment	Primary source	NTRS	Verdict	Key finding
B1	Liquid nitrogen zero boil-off	Plachta, Feller, Johnson & Robinson, NASA/TP-2017-219389, 74 pp.	20170001537	REPRODUCIBLE	The positive control of this audit. The only referent with a formal, propagated uncertainty budget: instrument uncertainties with cited sources, errors added in quadrature, results published with ± (Q_MLI = 2.62 ± 0.54 W). It demonstrates the practice was achieved in at least one cryogenic programme (no cost analysis was performed; what budget it took is not demonstrated); what that does and does not imply about its absence elsewhere is addressed in §IV.5.
B2	Small-scale LH₂ tank, normal gravity	Barsi & Kassemi, Cryogenics 48 (2008)	—	PENDING	Paywalled journal source.
B3	Small pressurized LNG tank	Ferrín & Pérez-Pérez, Comput. Chem. Eng. 138 (2020)	—	PENDING	Paywalled journal source.

C. Microgravity and flight

#	Experiment	Primary source	NTRS	Verdict	Key finding
C1	Saturn AS-203 S-IVB flight experiment (1966)	Ward et al., NASA CR-94045 (1967), 203 pp. — located during this audit; cited via Grayson et al., AIAA 2006-5258, 7 pp.	19680012073 / 20060047496	PARTIALLY_REPRODUCIBLE	Described by Grayson et al. (2006) as the only known adequately instrumented low-gravity cryogenic self-pressurization test; no later real-cryogen counterpart has been located for this registry. Eight pressure data points, telemetry blackout across 69 % of the 5,360 s experiment, ullage sensors that could not be corrected — and a partial circularity: the heat-leak boundary conditions applied by the 2006 CFD were themselves derived in 1967 from the flight’s internal fluid-state changes. The four source papers cite the 7-page CFD paper, not the 203-page experiment.
C2	TPCE — Tank Pressure Control Experiment (Shuttle STS-43, 1991)	Bentz, NASA-CR-191012 (1993) — not yet obtained; cited via Kartuzova & Kassemi CFD paper, 15 pp.	20170000965	NOT_REPRODUCIBLE (from cited document)	Simulant fluid (Freon-113 at 296 K), not a cryogen. The cited validation is qualitative by its own statement: video stills, unsynchronized timing, initial ullage position unrecoverable in one axis. Verdict may upgrade when the primary report is audited.
C3	ZBOT — Zero Boil-Off Tank (ISS, 2017)	Journal papers / NASA PSI archive — to pursue; cited via a 34-slide conference deck	20190030681	NOT_REPRODUCIBLE (from cited document)	The cited “validation” document is a slide deck. The experiment itself is the most instrument-rich microgravity tank referent in this registry (DPIV velocity fields, pressure and temperature evolutions) — with a simulant fluid (perfluoro-n-pentane). The usable referent is the journal record and data archive, not the deck.
C4	ZBOT-NC (non-condensable gases)	ELGEA 2024, Kassemi	20240010737	PENDING	Located, not yet audited.

D. Historical — the 1960 foundations

Cited by the modern literature as the origin of the phenomenon. If 2025 papers still rest on 1960 data, traceability matters more, not less.

#	Experiment	Source	Verdict
D1	Huntley — temperature-pressure-time relationships	Adv. Cryog. Eng. 3 (1960)	PENDING
D2	Neff — stratification survey	Adv. Cryog. Eng. 5 (1960)	PENDING
D3	Scott et al. — nonventing helium dewar stratification	J. Res. NBS 64C (1960)	PENDING (likely free via NIST)
D4	Swim — helium dewar temperature distributions	Adv. Cryog. Eng. 5 (1960)	PENDING

IV. Cross-cutting findings

Stated compactly; each is developed with full citations in the per-referent audit records and the numbered notes.

The experimentalists declare; the modelling papers examined do not consistently carry. In every audited case where the primary experimental report was obtained — eight of ten; C2 and C3 are audited through citing documents — the primary reports state instrument accuracies and, remarkably often, their own anomalies — the K-Site 120–150 % energy balance error, the MHTB energy discrepancy, the SHIIVER sensor uncertainties, the AS-203 uncorrectable ullage sensors. The validation papers examined before 2025 carry none of it. The 2025 full-length K-Site papers restore the instrument accuracies and the boundary-condition dependency — and still carry neither the closure failure nor the single-axis caveat (Note 001 §III). The failure is not availability; in the lineages examined, no maintained link keeps the qualifying context attached to the data. Some elements can later reappear — Note 001 records four accuracies and part of a boundary-condition dependency travelling back in 2025 — others remain detached.

How independent is that first observation? Less than it looks, and the arithmetic is worth stating. The candor is documented across three decades — 1992, 2003, 2021 — but not across three institutions. The arithmetic that follows runs over the four ground-test programs whose primary-report candor this section cites — K-Site, MHTB, SHIIVER and B1; AS-203, whose primary is a Chrysler contractor report under contract NAS8-4016, sits outside this center count. NASA Lewis Research Center was renamed NASA Glenn Research Center in 1999, a fact the SHIIVER report states itself while recounting the history of this work: “in-house NASA efforts at both Lewis Research Center (now Glenn Research Center) and Marshall Space Flight Center” (p. 2). So K-Site (Lewis, 1992), SHIIVER (Glenn, 2021) and B1 (Glenn with Ames, 2017) are the same center. K-Site and SHIIVER were moreover run at the same test site: the K-Site facility is “located at Plum Brook Station in Sandusky, Ohio” (TM-103804, 1991 — the statement appears in that report, not in the 1992 companion), and SHIIVER testing “was performed at Glenn Research Center’s Plum Brook Station (PBS) in Sandusky, Ohio,” since renamed the Neil A. Armstrong Test Facility. B1 additionally shares an author with SHIIVER. The honest count is two NASA centers, not three institutions: Glenn/Lewis three times, Marshall once.

MHTB (Marshall, 2003) is therefore, among these four ground-test programs, the only instance of the same habit led by a different NASA center, and it carries a disproportionate share of the weight. An alternative reading must be named: the candor documented here may be institutional culture at one NASA center, sustained across thirty years and a name change, rather than a professional norm of the field. MHTB is the single piece of evidence against that reading. Testing it requires experimental referents from outside NASA — European, Japanese, industrial — and that is the second task of v1.1.
K-Site’s declared “wall heat flux” is not a wall-only flux. The declared “wall heat flux” is total tank heat divided by internal area; 13–17 % of it enters through discrete penetrations, located in both the upper and lower halves of the tank; the report itself calls the uniform-flux assumption reasonable at the ~85 % level, and the wall-heating concentration it reports in the unwetted region is the wall-absorbed fraction — 3–7 % of total input — not the input distribution (Erratum 25). The localized 13–17 % is the caveat a sensitivity run must price for the quantity of interest.
The microgravity lineage is thinner than its citations suggest. The only real-cryogen low-gravity self-pressurization referent located for this registry is from 1966 — eight pressure points with a 69 % telemetry gap and partially circular boundary conditions. The later flight referents in this registry flew simulant fluids at ambient temperature, and one of the routinely cited validation documents is a slide deck.
No LH₂ referent declares the ortho/para state of its fluid. SHIIVER tabulates parahydrogen properties; no report states the spin composition of what was actually loaded. The parahydrogen assumption in the downstream models examined is physically reasonable and traceable to no declared spin-composition measurement in the records audited here.
The standard is achievable, though the comparison is not clean. B1 (2017) propagates uncertainties in quadrature with cited sources and publishes results with error bars — which establishes that formal result-level uncertainty propagation was achieved in at least one NASA cryogenic test programme. (No cost analysis was performed here; what budget that took is not demonstrated.) The inference that its absence elsewhere is therefore a choice is weaker than it first appears, and should be resisted in that strong form. B1 is a zero-boil-off technology demonstration in which the heat balance is the deliverable: propagating uncertainty onto heat loads is the product being reported. K-Site and MHTB are characterization tests whose deliverable is a pressure history under stated conditions. Different genres of test carry different bookkeeping obligations, and part of the gap may be that rather than differing rigor. What B1 establishes securely is the weaker and still useful claim: when a programme needs propagated uncertainty, this field produces it. Whether the referent programmes should have needed it is a judgement this registry does not settle.
Small documentary defects compound. Across the K-Site family alone: three inconsistent values for the cold-shroud temperature (140/150/152 °R), a factor-of-two discrepancy in flowmeter accuracy between companion documents, and an internal 164.4-vs-164.6 BTU/hr inconsistency inside one report. None matters physically; all of them break automated cross-checking, which is how context gets lost at scale.

V. What this audit is not

Not a criticism of the experimentalists. The pilot expectation — “old reports don’t state uncertainty” — was falsified by the audit itself: the reports state it, often carefully, and are candid about their own anomalies to a degree that surprised the auditor. The problem lives downstream, in what travels.

Not a claim that any model is wrong. Nothing here assesses model physics. A model validated against a referent with unquantified uncertainty may be excellent; what that comparison alone cannot support is a defensible quantified accuracy claim — other evidence bases may.

Not complete. Eight sources remain unobtained (§III: B2, B3, C2 primary, C4, D1–D4). They enter future versions of this registry, and any of them may change a verdict.

VI. Verification method

Everything below that concerns a registry source artifact is reproducible from the hashes in §VIII; claims about downstream-use documents are reproducible from the hashes in the citing note’s source list and in the companion dataset’s verification log and manifest.

Primary-source rule. Verdicts attach to documents actually read. Where the validation literature cites an intermediate document (C1, C2, C3), the registry says so explicitly and audits both where possible.

Quote transcription policy. Quotations are transcribed with obvious OCR/scan artifacts corrected (line-break hyphenation, broken spacing, ± rendered as stray characters, interleaved page markers and figure captions). Genuine typos original to the sources are preserved verbatim, marked [sic] where confusion is possible — Erratum 4 records why: a normalization habit can silently swallow a real feature of the source, and “ulllage” is the standing example. Only artifacts that the page image proves are not in the source are corrected. The exact extracted text is preserved in the project record and checkable against the artifact hashes.

Two-pass quote verification. (1) Literal normalized-substring match against the full extracted corpus; (2) for failures, artifact-tolerant matching on an alphanumeric skeleton with similarity ≥ 0.92, plus manual inspection of every remaining flag. Result for the audit corpus at v1.0: the quotations in the eleven audit files — the ten referent records plus the master candidate list — were verified individually; two genuine transcription defects were found and corrected (both listed in §VII). Per-quotation outcomes, search procedures and artifact hashes are recorded in the verification log published with the companion dataset (doi:10.5281/zenodo.21895803); a summary decomposition drafted for this paragraph did not sum to its own total and was withdrawn before publication (Erratum 14).

Counts publish their operational definition. Any automated count (tables, term occurrences) states what was counted, after Erratum 3 demonstrated why.

VII. Errata to date

Maintained as part of this note. Corrections are credited where they originate outside the project.

#	Date	Defect	Correction
1	11 Aug 2026	A quotation about SHIIVER heat-flux sensor uncertainty was attributed to the 286-page final report; it appears in the supporting CEC paper (NTRS 20210019122).	Attribution corrected in the A5 audit record.
2	11 Aug 2026	Two transcription defects: a paraphrase presented as a quotation in the A5 record (now verbatim: “(it is assumed that FM3 is accurate)”), and a condensation without ellipsis in the B1 record (restored “(0.17 kPa)”).	Both corrected; found by the two-pass verifier.
3	11 Aug 2026	The preliminary scan table reported raw “Table N” mentions as table counts (SHIIVER: “103” vs. 52 actual numbered tables: 46 numeric + 6 appendix).	Detected by cross-check against Note 002; independently re-counted with deduplicated headers for all documents; scan table replaced. Origin: publications session cross-audit. (The corrected figure was itself wrong — see Erratum 14.)
4	11 Aug 2026	Two defects in one quotation in the A3 (MHTB) audit record: a truncation without ellipsis, dropping “the minimum setpoint” and the cycle-closing sentence; and the silent normalization of “ulllage”, an original typo in the source, to “ullage”.	Both corrected; the original typo restored with `[sic]`, verified against page images. Found by the publications session on re-verifying an inherited quotation at full length against its own artifact. Method consequence, now standing policy: an OCR-artifact correction policy (correct) can silently swallow a real typo in the source (incorrect). Inherited quotations are re-verified at full length against the artifact, not carried across documents.
5	11 Aug 2026	Criteria drift across the series: three documents carried three materially different versions of criteria 3 and 6 before publication.	Reconciled in §II before first publication: Note 001’s wording adopted as canonical, Note 002’s committed refinements incorporated as scoring scales, provenance table added. No note published with a divergent version.
6	11 Aug 2026	This note claimed to adopt Note 001’s criteria verbatim; criteria 1 and 5 differed by a phrase and a clause reordering. Semantically identical — no verdict was affected — but a false fidelity claim, asserted in the document that defines fidelity as a criterion.	Detected by literal diff, not by reading. Resolved by aligning Note 001 to this note’s wording, which is the better of the two in both cases; all six criteria now match character-for-character, verified by normalized diff. Origin: publications session. Method consequence, now standing policy: “adopted verbatim” is a machine-checkable claim, so it is checked by machine. Any assertion of textual identity between our own documents is verified by diff before publication — the same failure mode as the “103 tables” count, where output looked clean and nobody diffed it.
7	11 Aug 2026	All four notes carried the wrong company name. Two drafts diverged (`Elarion CPMS` / `ElarionX CPMS`); the inconsistency was correctly detected and deliberately unified — in the wrong direction. The company is ElarionX.	Corrected across all four notes before publication, verified by negative-lookahead regex (a naïve search cannot distinguish the two, since one name is a substring of the other). Method consequence, and the most important entry in this table: detecting an inconsistency and resolving it are different acts. Every other defect listed here was resolvable from the documents in hand. This one required a fact that existed in no document — and so no amount of document verification would have caught it. It is the only erratum in this table that would otherwise have reached a permanent identifier. Standing policy: when unifying a divergence, establish which variant is canonical from a source of truth outside the diverging documents, or record that the direction is unverified.
8	11 Aug 2026	Four methodological weaknesses were not disclosed in the drafts: the four bibliography-source papers are not independent (three share an author and institution); those same papers do double duty as both the referent-selection basis and the evidence of the custodial failure; the B1 counter-example compares test genres with different reporting obligations; and the step from context did not travel to decisions were worse is inferred rather than shown.	All four now disclosed in §III, §IV.5 and §IX; the phrase “the experiments everyone uses” removed. Surfaced by an adversarial reading commissioned from an independent model, then verified here against the author lists of the four papers. Method consequence: the commissioned refutation attacked the conclusions and reported the core intact; every weakness that survived was methodological, and none was found by it. A red team that cannot fail is not a red team. Standing policy: adversarial reviews are commissioned per claim — for each stated claim, what evidence would falsify it, and was that evidence collected? — not as an open invitation to find fault.
9	11 Aug 2026	While correcting Erratum 8, the drafts asserted that the experimental sample was independent — “three teams, three institutions, three decades.” It is not. NASA Lewis Research Center was renamed NASA Glenn Research Center in 1999, so K-Site (1992), SHIIVER (2021) and B1 (2017) are the same center; K-Site and SHIIVER ran at the same test site, Plum Brook Station; and B1 shares an author with SHIIVER.	Corrected in §IV.1 to two NASA centers, not three institutions, with the alternative reading named: the documented candor may be institutional culture at one center rather than a norm of the field, with MHTB (Marshall) the only independent evidence against that reading. Method consequence, and the one most worth reusing: having just found that the modelling-side sample was not independent, we asserted independence for the experimental-side sample one paragraph later, without checking it — and the disproving facts were on title pages we already had, one of them written into a report we had read end to end. The general form: the correction of a bias is itself a new claim, and inherits the burden of proof of the claim it replaces. A fix asserted while repairing a known weakness feels like diligence and is therefore under-scrutinized, which is precisely when it should be scrutinized most. Standing policy: sample independence is verified against affiliations and facilities, never inferred from differing report names, centers’ names, or dates.
10	11 Aug 2026	A late rewrite of the four titles — undertaken as presentation work — reintroduced unbounded claims into the most-read sentences of the series. Note 002’s new subtitle asserted SHIIVER was the best-documented test “in the public record”, a corpus nobody has inventoried; Note 001’s new title said “The Papers Citing It Did Not”, implying all of them, when three documents were examined and two share authors; and this note’s subtitle claimed to report what “the public record” can support.	All three narrowed to what is held: “the best-documented cryogenic tank test I have read”, “Two Papers Citing It Did Not”, and “what each can and cannot support”. §II’s reference to “the best-documented case in the registry” likewise bounded to the referents audited here. First caught in Note 002 by the publications session; the other two found by auditing the remaining titles against the same test. Method consequence: this is the third recurrence of the same class of overclaim in one day, and the first to enter through an edit that felt like formatting rather than authorship. Standing policy: titles and subtitles are claims and are audited as claims. They are also the only sentences most readers will see, which makes them the worst place in a document to relax the standard.
11	11 Aug 2026	A companion document to the MHTB primary was missing from the audit: Hedayat, Bailey, Hastings & Flachbart, Test Data Analysis of a Spray Bar Zero-Gravity Liquid Hydrogen Vent System for Upper Stages (NTRS 20030106051, 8 pp.), which reports directly on the 1996 and 1998 campaigns and tabulates seven test segments — five of them not named in the audit record.	Added to the A3 entry and record. It does not change the A3 verdict: what TABLE 1 tabulates are comparisons against the authors’ own analytical model, not measurements, so the tabulated time histories remain unlocated and the open action stands. Found by the publications session while looking for whom to write to, not while auditing. Method consequence, new in this table: we searched for the primary report, found it, and stopped — without searching for companion documents by the same authors on the same campaign. A technical memorandum and a conference paper from one team on one test series are distinct pieces of the public record. Standing policy: referent search runs by author and by campaign identifier, not only by experiment name, and does not terminate on first success — the same “stop on confirmation” reflex as Erratum 9. Two corollaries worth keeping. First, outreach work surfaced evidence the audit’s own search had missed, which argues for author contact preceding or overlapping the audit rather than following it. Second, in reporting this document the audit session miscounted a term — three occurrences of “single node” where there are two, from a query that ORed three patterns and counted printed blocks rather than term hits; caught by the publications session on re-verification. The count discipline of Erratum 3 applies not only to figures published in the notes but to every count passed between sessions, since those become published figures one step later.
12	11 Aug 2026	Note 002’s reference list still cited Note 001 by its pre-Erratum-10 title, “The Papers Citing It Did Not” — the exact unbounded claim Erratum 10 had just removed, surviving in a citation inside another note. Found during Zenodo deposit preparation, after the notes had been declared final and audited five separate ways.	Corrected in Note 002 and in the combined HTML reader; artifacts rebuilt and re-verified by text extraction from the regenerated PDF. Method consequence: Erratum 10 corrected the title where the title lives and never asked where else that string had been written down. A retitling is a rename, and a rename is only complete when every occurrence moves — including the ones inside other documents, which no amount of re-reading this document will surface. Standing policy: any change to a title, name or identifier is followed by a corpus-wide literal search for the superseded string, and the search is what closes the change, not the edit. The legitimate remaining occurrences are those inside this errata table — Erratum 10 and this row — where the old wording is the subject.
13	11 Aug 2026	The four PDF artifacts, repeatedly audited as extracted text, were defective as documents: all four carried an unpopulated table-of-contents field, rendering as a page headed “Contenido” — the build machine’s locale language — with no entries; the regenerated PDFs of Notes 000 and 002 carried Note 003’s running header on every page, stamped by a one-off rebuild script cloned from the Note 003 build; and every PDF’s embedded metadata named the creator “Elarion CPMS” — the bare name Erratum 7 exists to forbid — and titled each file “Technical Notes 000-003”. Found by the author on visual review of the PDFs during deposit preparation, after five text-level audit passes had declared the artifacts clean.	Build script rewritten: the running header and document metadata are now derived from the note’s own byline, never hardcoded; the field-based contents replaced with a static list whose page numbers are computed from the rendered PDF and verified against it; creator corrected to ElarionX CPMS. All four PDFs regenerated and re-verified at artifact level. Method consequence: every prior audit read the PDFs the way a machine does — by extracting their text — and all three defects live where extraction does not look: the running header, the metadata block, an empty field. A document is not its extracted text. Standing policy: artifact-level verification — headers on every page, embedded metadata, populated navigation — runs on the rendered file the reader receives; and a one-off rebuild script inherits the identity audit of the batch script it replaces. This is Erratum 7’s defect class recurring inside the toolchain, and Erratum 12’s rename lesson recurring in a script.
14	11 Aug 2026	A commissioned adversarial audit (75 agent runs across seven separately specified lenses, every finding re-verified by a skeptic run instructed to refute it) — confirmed a set of content defects that twelve prior errata and repeated re-reading had not caught. The exhibits: Erratum 3’s own correction was wrong — SHIIVER’s table count is 74, not 52; the deduplicated recount captured appendix series I.1–I.6 but missed H.1–H.19 and J.1–J.3. A count contradicted its own document six lines away (“four times” against the correct “five times”). An enumeration of eight unobtained sources was labelled “seven”. Three drafted decompositions of one verification run — in this note and both language editions of the candidate list — failed to sum to their own totals, differently. Two page citations were wrong, one of them load-bearing for Erratum 9’s correction. Cross-references pointed to appendices of this note that do not exist. A dataset record title claimed “the only flight experiment” while two other records in the same dataset describe flight experiments. The Spanish edition still shipped the “103” figure Erratum 3 retracted, and rendered two source quotations in Spanish against the series’ own transcription rule.	All corrected in this version; counts now carry their operational definitions; cross-references now carry DOIs; the withdrawn tally is replaced by a pointer to the verification log, where per-quotation outcomes live. Method consequence: every one of these defects lived in prose that had been read many times and re-derived never — re-reading approves what re-derivation rejects. Standing policy: every published count, page citation and cross-document pointer is re-derived mechanically before release; a correction to a count re-runs the count’s own operational definition (Erratum 3’s correction did not); and no decomposition is published unless a machine has checked that it sums.
15	11 Aug 2026	An externally commissioned AI review of the built PDF (OpenAI GPT-5, arranged by the author) found, among items verified here against the artifacts before acceptance: the §VI transcription policy still stating the pre-Erratum-4 rule — with “ulllage”, Erratum 4’s own counterexample, listed as a typo to normalize silently; §IX still reading “seven sources” after Erratum 14 corrected the same count in §V — the class was not closed, only the instance; the B1 registry row still calling absent uncertainty propagation “a choice” while §IV.5 argues that inference must be resisted; unbounded claims surviving in registry rows and §IV (“the only … that exists”, “best-instrumented … flown”, “universal … traceable to nothing”, “everything since”, and “in every audited case” reaching the two referents audited only through citing documents); and Erratum 14’s own table row rendered outside the errata table, because the insertion tooling left a blank line — breaking the table in the document that records artifact-level verification.	All corrected in this version: the transcription policy now states the post-Erratum-4 rule; §IX separates seven PENDING entries from eight unobtained artifacts; superlatives are bounded to the registry; the row is rejoined and errata-table integrity is now machine-checked at build time. Suggestions from the same review that did not survive verification were declined, with reasons recorded in the project log. Method consequence: closing an erratum closes an instance; the class stays open until a corpus-wide search says otherwise — Erratum 14 itself proved this by correcting one seven-vs-eight while a second stood two sections later. Standing policy: every erratum closure triggers a class-wide search, and independently built external review joins the pre-release path.
16	11 Aug 2026	A second round of externally commissioned AI review (Note 001) surfaced, and verification against the world confirmed, two false negatives in the deposited verification log: a journal article dismissed as “likely nonexistent” exists — Cryogenics 152 (2025) 104210, open access, Crossref-verified — and the AIAA SciTech 2025 full paper recorded as “paywalled; not read” had been publicly downloadable on NTRS (20240016283) the whole time. Both were obtained, hashed and read before this version froze: they restore K-Site’s four instrument accuracies and the Stochl–Knoll boundary-condition dependency, and carry neither the 120–150 % closure failure, nor the single-axis caveat, nor any assembled validation uncertainty. Also confirmed: Note 001’s own introduction repeated the K-Site flux values without the recovered penetration context; the criteria chronology was stated inconsistently between Notes 000 and 001; and the notes promised “corrected in place” against this registry’s own versioning policy.	Note 001 §III rebuilt as a 1992→2025 information-travel lineage over five documents; its title changed accordingly (the superseded title, “K-Site Published Its Own Closure Failure. Two Papers Citing It Did Not.”, is recorded here as the subject); the log entries corrected by appending SUPERSEDED notes, never by deletion, under an explicit OPEN/CLOSED/SUPERSEDED discipline; flux context added to Note 001 §I; one chronology now stated in both notes; corrections policy aligned to versioning. Method consequence: a negative existence claim — not found, paywalled, nonexistent — is a timestamped search result, not a fact about the world, and it expires. Standing policy: every negative claim carries its search date and is re-run at freeze; paywalled is re-tested like any other negative; and a custody claim about a lineage examines that lineage’s newest full-length members before it freezes.
17	11 Aug 2026	Third round of externally commissioned AI review (Note 002), each finding verified against the artifacts. Confirmed: criterion six was declared “first downstream paper examined (2023): context did not travel (Note 002 §IV)” while the first full-length SHIIVER validation paper had been public on NTRS since January 2023 — the expiring-negative class of Erratum 16, third instance; the phrase “all 52 tables” survived inside the very sentence supporting the no-uncertainty-budget negative after Erratum 14 corrected the count to 74 — a corrected count did not re-run its dependents; the FM3 low-flow cutoff was published categorically (~2 g/s) while the verification log held the 2 g/s / 0.75 g/s contradiction unresolved; “the fluid state is explicit” overrode the note’s own analysis-basis/loaded-composition distinction; “independent” was applied to flowmeters plumbed in series and to a source that says “other calculation methods”; and “most referents in this field” exceeded the audited corpus.	All corrected in Note 002 v1.0. The 2023 paper (NTRS 20220017916, hashed) was audited against SHIIVER’s own published qualifications: none travelled into it, and its “within 4%” headline carries no experimental uncertainty to be judged against — criterion six’s first measurement, recorded in Note 002 §IV. The no-budget negative was re-established over the corrected 74-table universe, with the method-to-method root-sum-square dispersion passages (7–69 % for heat fluxes; 10–60 % for heat loads) now described rather than elided. Method consequence: correcting a count re-runs every sentence that consumed it; a note may not state more certainty than its own log records; and criterion-6 claims are re-verified against a fresh downstream search at every freeze.
18	11 Aug 2026	Fourth round of externally commissioned AI review (Note 003), verified against the artifacts. Confirmed: the subtitle claimed “Reading 161 pages” while the A3 audit record declares section-targeted reading — a headline claiming a depth the audit record does not; “closed that gap” coexisting with the note’s own open action to locate the tabulated record; criterion 6 assessed against the 2015 paper alone while two 2016 journal validations (Cryogenics 74, DOIs verified) and a 2025 Marshall presentation existed — the expiring-negative class, fourth instance; the paper→test-ID association stated as if printed, when the 2015 paper prints no identifiers (association is by fill level, now stated as inferred); “the same diagnosis of the same limitation” conflating K-Site’s measurement-representativeness caveat with the companion’s model-abstraction limitation; the basis/composition conflation corrected in Note 002 by Erratum 17 standing uncorrected in this sibling note; “two independent organizations” for two NASA-center-led teams; and an unverifiable employment claim about named individuals.	All corrected in Note 003 v1.0: subtitle now says auditing; the gap described as moving, not closed; the downstream lineage updated with the two located 2016 papers (unobtained, paywalled — recorded as such) and the audited 2025 presentation (zero context carried); the ID mapping machine-resolved against the TM’s figure captions and the inference declared; the analogy recast as measurement-side vs model-side; the census of all eight numbered tables machine-verified. Method consequence: a class corrected in one note is searched in its siblings the same day (the basis/composition conflation survived one note over); and a note’s subtitle inherits the depth its own audit record declares — titles are claims, fourth application.
19	11 Aug 2026	Fifth round of externally commissioned AI review — the companion dataset itself. Most of its findings had already been corrected in rounds two to four (the review examined a pre-correction snapshot; its independent rediscovery of Errata 14–18’s defects from the artifacts alone is evidence those defects were real). Two states had nonetheless survived in the dataset layer: the verification log’s Known open items still held “MHTB primary: not yet obtained” after the same log’s §9d recorded it obtained, hashed and audited — a CLOSED entry standing OPEN, the exact failure the log’s own status-discipline note was added that day to prevent; and the C3 record still claimed “the only real-cryogen microgravity self-pressurization measurement in the public record” — the Erratum-10 unbounded class, corrected in the notes and surviving in the dataset.	Both corrected in both language editions. Method consequence: the class-wide search that closes an erratum runs over every layer of the corpus — notes, dataset records, and the verification log itself — and a status discipline is only as strong as the sweep that enforces it. That prose files cannot enforce it is, in the end, the argument for the machine-checked registry this project’s Phase 0 specifies: this errata table is that argument’s evidence, nineteen rows long.
20	11 Aug 2026	A dated negative recorded earlier the same day — a “Soriano et al. 2025” MHTB validation, searched and “not located” on NTRS and Crossref — expired within hours: a third external AI reviewer located the item as NTRS 20240016269, Modeling Self-Pressurization and Spray Bar Pressure Control of a Cryogenic Storage Tank in Normal Gravity (Soriano, Kartuzova, Kassemi & Hauser, AIAA SciTech 2025, presentation, 29 pp.). The citation as originally reviewed was garbled in form — no journal paper of that name exists — but its substance was real, under keywords this project’s searches had missed.	Obtained, hashed and audited the same night: it concludes a “validated CFD model accurately predicts” MHTB’s behavior and carries zero measurement-uncertainty statements and none of the TM’s declared context — the fourth downstream MHTB artifact examined, same shape as the first three. Added to Note 003’s lineage; the log corrected by appending, per its own discipline. Method consequence: the negative-claim discipline performed exactly as designed — dated, expired, corrected by append, all within one day — and a garbled citation can still point at something real: refuting a reference’s form does not close the search on its substance.
21	11 Aug 2026	Note 003’s downstream search stopped at a closed door: it declared the two 2016 Cryogenics journal papers “located, paywalled, not yet obtained” the open half of the measurement, and treated the downstream corpus as exhausted. A sixth round of externally commissioned AI review located five public documents that search had missed: the NTRS conference versions of both 2016 papers (20170006183; 20150021282 — same titles, same authors, free to download), the full conference paper of the presentation audited under Erratum 20 (20250000041, 30 pp.), a further 2025 spray-bar validation manuscript (20250005060, 36 pp.), and a validation by a non-NASA group — Wang, Jordan & Xiao (Karlsruhe Institute of Technology), Hydrogen Safety (2025) — that names the MHTB 50 % fill test by its exact test ID, P263981T.	All five obtained, hashed and audited the same day (criterion-6 sweeps in the verification log); Note 003’s lineage, sample-size statement and sources rewritten on the expanded corpus of eight audited downstream artifacts across three research groups. Method consequence: “located, paywalled” is an expiring negative like the rest — a paywall closes a door, not the search; version lineage (same title, same authors, different container) is a standard escape route and belongs in every downstream sweep.
22	11 Aug 2026	The same sixth round, auditing the companion dataset against the corrected notes, found six records still presenting conceptually superseded content as current: A3 carried the K-Site/MHTB mechanism conflation and an ID association stated as if explicit; A5’s scorecard kept boolean “YES” verdicts that the refined criterion-1 profile and use-relative criterion-2 scoring had replaced; B1 kept “removes the excuse … ordinary program budget … a choice” as authoritative prose neutralized only by a trailing note; A1 and A2 stated the 2025 zero-uncertainty finding without scoping it to the extended abstract then audited; C1 held an unbounded downstream negative; the candidate list’s “19 (18 + the located AS-203 primary)” conflated referent with source artifact. None were string-identical to previously corrected claims, so the class-string sweeps of Errata 14–19 could not catch them.	All corrected in both language editions with visibly marked SUPERSEDED text, separated record-revision / audit-completion states, and counts realigned with §III’s source ontology. Method consequence: a string-level sweep closes string-level defects only; claim-level closure needs the claim inventory a structured registry maintains mechanically. This table is that argument’s evidence, twenty-two rows long.
23	11 Aug 2026	Erratum 13’s correction clause misstated its own fix, in both field and value: it reads “creator corrected to ElarionX CPMS”, while the shipped PDFs carry document author “Luis Johan Elizondo Arias” — set through the DOCX dc:creator property, which maps to the PDF Author field — and the PDF-native Creator field reads “Writer”, the converting application’s name, which was never the object of the fix. Found independently by two cold-context AI reviewers in a seventh review round, each re-running the project’s own artifact-level verification against the frozen files.	This row records the artifact’s actual metadata; Erratum 13’s text stands as written, per the append-only discipline, with this row as its correction. Method consequence: an erratum is itself a claim about an artifact and enters the same verification universe as every other claim — the text of a correction must be diffed against the corrected artifact, not against the intent of the fix.
24	11 Aug 2026	The same seventh round — four cold-context adversarial readings given the artifacts and no project context — surfaced four further defects, all in the layer the sixth round had just rebuilt: Note 003’s Sources numbered two lists into collision (the v1.0 expansion extended section A to items 5–9 while section B kept its pre-expansion continuation numbers 5–6); Note 003’s scorecard cited “§IV½”, a label no section carries; registry entry A4’s verdict modifier “(high)” appeared on no defined scale; and later errata rows had dropped the “AI” qualifier from “externally commissioned review” that Erratum 15’s wording carries — these reviews are AI runs arranged by the author, and the wording must not let a reader assume otherwise. The round also confirmed that no formal competing-interest statement existed despite the affiliation standing on every page.	Sources B renumbered 10–11; the scorecard reference now names the subsection; the (high) modifier is defined in §II pending the v1.1 decision rule; “AI” restored to every review-round mention here and in the log; a competing-interest declaration added to §IX. Method consequence: each closure pass is itself an artifact the next pass must audit — and a cold reader with no context is the cheapest instrument for finding what a briefed reviewer has learned to look past.
25	11 Aug 2026	A physics gloss — faithful in transcription, wrong in interpretation — ran through the K-Site boundary-condition finding. TM-105411 states, verbatim as record A2 quotes it: “It was experimentally observed that wall heating occurs predominately in the unwetted wall region” — and the report’s own next sentence (“less incoming thermal energy is absorbed by wall heating”) and its Table 2 (wall fraction 3–7 % of total input) define “wall heating” as energy absorbed by the wall: storage, consistent with a uniform external flux, since a dry wall warms because no liquid cools it. The same report states: “(The uniform heat flux assumption is reasonable, since about 85 percent of the heat leak is uniformly distributed and the remaining heat leak paths are located in both the upper and lower halves of the tank.)” Record A2 §3.6 and this registry’s A2 row glossed the quoted sentence as the spatial distribution of the incoming heat (“the remaining heating concentrates predominantly on the dry wall … wrong by two independent routes”), and Note 001 carried “predominantly in the unwetted region” into its boundary-condition sentence. Found by a cold-context AI reviewer reading the source PDFs, eighth round.	Gloss corrected everywhere; one documented non-uniformity route stands — 13–17 % through discrete penetrations, in both the upper and lower halves (Stochl–Knoll) — the “second route” is retired, and the report’s own ~85 %-uniform statement now travels with the caveat. Note 001’s recovered/not-recovered statement rewritten on the corrected physics. Method consequence: transcription fidelity does not verify interpretation. A quantitative gloss must survive a dimensional sanity check against the source’s own numbers — the 3–7 % wall fraction was printed in the very table the record cites.
26	11 Aug 2026	Three defects in Note 003, same eighth round: the note reproduced the TM’s ±9×10⁻⁴ kPa control band (TM p. 65) without detecting that the same report states ±0.001 psi (≈0.0069 kPa) at four other places (pp. 60, 62, 74, 96) — a factor-7.7 internal inconsistency in the source that a traceability audit exists to catch; it praised “precise pressure work” on a ±0.02 percent transducer specification that does not state percent-of-reading versus percent-of-full-scale — at 133 kPa, 0.02 % of reading is ±0.027 kPa, larger than the control band — an ambiguity that blocks converting the specification into u_D; and a residual sentence, “exactly one paper”, contradicted the corrected eight-artifact count a page earlier — it survived because the class sweep searched “single paper”, not its synonyms.	Both TM values now reported side by side as an unresolved source inconsistency with the conservative reading governing; the precision praise re-scoped to name the ambiguity; the residual sentence corrected. Method consequence: an audit that quotes a number inherits the duty to check the source against itself; and a class sweep closes a wording, not a meaning — synonyms escape it.
27	11 Aug 2026	Cross-cutting catches of the eighth round (two cold-context AI reviews with the sources and the working tree): three documents quantified in Note 001 (NTRS 20150000249, 20220018548, 20240006348) were hashed only in the verification log, contradicting the README’s claim that manifest.csv lists every examined artifact; Note 001 said the K-Site authors “ruled out the instruments” where the report bounds two contributions as insufficient and prices a third — insufficiency is not exclusion; Note 002 called the SHIIVER test plan’s requirements “set before the test” on the evidence of Revision I (9 Jan 2020), whose own revision history records updates “based on actuals” from the 2019 phases; §II’s heading “Method, fixed before the audit” overstated its own provenance table; the 2023 SHIIVER paper’s actual thermal boundary condition (a Thermal Desktop model’s “preliminary estimates”, by its own words) was not named in §IV, leaving the heat-flux-sensor absence to carry more weight than the applied input’s unquantified uncertainty; and the u_val quadrature carried no independence caveat.	Manifest completed (28 artifacts); wording corrected to “bounded … insufficient, not excluded”; the plan’s partial retrospectivity declared; the heading now says what the provenance table says; §IV names the Thermal Desktop boundary condition; the decomposition now states that quadrature presumes independence and that shared data couple u_input and u_D. Method consequence: a cold reviewer holding the sources is the strongest instrument this project has used — three briefed rounds missed what one context-free read found; the durable answer is not better briefing but machine-checkable claims.
28	12 Aug 2026	Ninth and tenth cold-context AI reviews (a re-run gate and two persona re-reads of the corrected corpus) surfaced five residues: the verification log’s Round-six section still pointed to “Note 003 §IV½” after Erratum 24 corrected that label in the note — the class sweep missed the log layer, Erratum 22’s own lesson recurring; Note 003 printed a rate ratio as 3.1 where the TM’s Table 5 values give 0.96/0.29 = 3.3; Note 003 said the TM “rules the instruments out arithmetically” after Erratum 27 softened the same overstatement in Note 001 — the synonym escaped the sweep, again; §I’s “E only carries meaning if u_D is known” overstated its own framework (E without u_D still supports coarse discrepancy detection, just not percentage-accuracy claims); The round also alleged a fifth: that §III’s census line left the Stochl–Knoll dependency unreconciled with its status as an experiment — re-inspection refuted this; the line already tracks it as a separate source, not a seventeenth registry row.	The four confirmed residues corrected; the log pointer carries its own visible correction note; the fifth allegation recorded as checked and refuted. Method consequence: the correction of a defect creates a defect class, and the class includes every layer and every synonym — four instances of this lesson now stand in this table, which is the strongest argument yet that closure needs to be computed, not remembered.
29	12 Aug 2026	An internal cold panel of five source-holding AI verifiers (round eleven; 81 checks) confirmed seven content-class residues: the SHIIVER method-dispersion range printed as “7–60 %” while the report’s flux dispersions reach 69 % (aft dome, TP p. 94; loads top out at 60 %); Note 003’s corrected rate ratio silently overrode the TM’s own printed prose (“2.9 and 3.1 times”, p. 68) with the Table-5 derivation (3.3) instead of reporting the source’s internal prose-vs-table inconsistency per Erratum 26’s own policy — and Erratum 28’s text misattributed the 3.1 to this note when it is the TM’s printed number; “accuracy appears 18 times” counted a word family without saying so (the exact word appears 11 times; the family 18); Note 001 §I claimed “neither K-Site document flags” the heat-path distribution while TM-105411’s own page prints the ~85 %-uniform parenthesis Erratum 25 quotes; Note 003 §III reinstated “the instruments excluded” for K-Site in the very paragraph its corrected gloss bounds; the Erratum-28 correction over-corrected the MHTB half — TM-2003-212926 p. 71 prints “One source considered, but ruled out, was instrumentation uncertainties”, so exclusion is the MHTB authors’ own claim and stands as quoted, while K-Site’s stays bounded; and Note 002 said the report “does not say” whether a data archive exists while the report twice states the data “have been archived”.	All corrected; asymmetries restored; ranges split and stated. Method consequence: corrections are claims too — an over-correction is still a divergence from the source, and a range endpoint is a count like any other.
30	12 Aug 2026	The same round’s cross-layer, mirror and fresh-refuter passes: the A2 registry row’s “rule out instrumentation” (third un-swept layer of the Erratum-27 class); Note 001 §V’s prospectivity wording un-hedged after Erratum 27 hedged Note 002; Note 001 §VIII’s “cannot be applied at all without u_D”; two false pointers (the Cryogenics-capture hash attributed to the verification log — it lives in manifest.csv — and a “§I.1” label no section carries); the log’s “Last updated: 11 August 2026” stamp standing over a 12 Aug append; the edition date itself — bylines read 11 August over content carrying 12 Aug events; the edition now states 12 August 2026; registry row C1’s unbounded “the modern literature”; the §II provenance table and Erratum 6 asserting opposite directions for the same criteria alignment; the es edition of A2 missing the strikethroughs its en mirror carries, both editions’ A2 and candidate lists retaining the retired dry-wall gloss in secondary sites, and the es candidate list’s Appendix P missing the A3-companion row its own count promises; four stale .txt extraction byproducts removed from the deposit folder; and a front-matter definition of the word “audit” added — documentary audit, one auditor, AI-assisted, not a certified independent engineering audit.	All corrected; the es mirror completed; the stamp discipline now covers the stamp itself. Method consequence: thirty rows in, the residual classes are the corrections’ own — this table is Phase 0’s specification, stated as evidence.

The audit also corrected itself twice before publication: an initial keyword-scan pass produced a false “uncertainty not reported” conclusion for K-Site (the pattern accurac does not match “accurate,” and OCR had mangled every ± symbol), overturned by full reading; and MHTB’s pilot scoring, based on a secondary source, was revised when the 161-page primary was obtained. Both reversals strengthened the central finding while falsifying its first formulation — which is what a method that works looks like.

VIII. Provenance of registry source artifacts

All PDFs retrieved from NTRS (https://ntrs.nasa.gov/api/citations/<ID>/downloads/…), 11 August 2026. SHA-256 of the exact artifacts audited:

Reg.	NTRS ID	Pages	SHA-256
A1	19910011011	8	`b08568a71c7678dfd63c93ac93c35b2737df48cedd36f4147fd6c009ce351513`
A1d	19910015845	21	`87250c7124e5642e3774bbfb4747cf34a7e235a1451aa7061287444d0d2dea99`
A2	19920009200	11	`91775f3185f280847c5fc27cc086517cc7d8fd8db32d5bfd5deda0a79f109e78`
A3	20040000092	161	`229abc1aee69d6bd3598dd8772d5863bc4d4123a43fa7f952d8029be20326843`
A4	20010063699	90	`b759a67208b5052a9847cdcdf37d214cd4c03d2b786fd3e7c6eaadce761b1a9d`
A5	20205008233	286	`9393a4d0c8b77f70ec298045b44560c8b8781171534bef941b160e1689a8da3f`
A5b	20205003433	53	`a504cf17cea4031bd773f43fa05a129459c1495249f2b81735bf8d1511ca79ae`
A5c	20210019122	8	`d1fa7497c2650c27fd56a237d08b22cc9eb2df0fa369a50a039b09dea5c8924d`
A3 companion	20030106051	8	`ffefbc10c13f6f51ab6f2b992aaa5ccce2267bc720aaa1b59b97d77e659de087`
B1	20170001537	74	`ab6adcc530318de584d08c5873b13ab3b89d1a82b8da2f2c5a26fa090945747d`
C1 (2006)	20060047496	7	`e964fb3d2a19fc1075c090d7afd3e1ea26f5e01bae35bc8c978457634eea448f`
C1 (1967)	19680012073	203	`cf36effc3b5c5514d1ecd0b298d8506d56472cf9b03693ce0399e179e69a6727`
C2	20170000965	15	`c69ca11a08ed88c2b4c42d82f8fcaa267c5ea006901bac9459ae3a8ec0764ade`
C3	20190030681	34	`7ddd14999b9376b0888aff1553636ab627aa075dc0fe114a7aa1738ca6d5e390`

IX. Limitations

One auditor, working with AI assistance. No independent human reviewer has yet examined this work. Under the project’s own authority scheme this is exploratory (A0) output, and it is labeled as such.
Competing interest, declared. ElarionX CPMS develops and evaluates engineering models for cryogenic propulsion systems, and the machine-checked referent registry this series argues for — Phase 0 — is the project’s own planned product. The series’ conclusions therefore favor something its author intends to build, and a reader should weight them knowing that. The mitigation offered is not neutrality but verifiability: every claim carries the provenance needed to check it without trusting the author.
Text extraction is OCR-limited for the scanned 1990s documents. Quantitative claims about the 1992 K-Site report were verified against page images, not extracted text; the same discipline applies elsewhere where extraction is suspect.
The candidate list is bibliography-derived from four papers, three of which share an author and an institution (§III). It is a defensible operational definition, not a systematic literature survey; experiments may be missing, and the list may reflect one group’s citation habits more than the field’s.
The consequence is inferred, not demonstrated. This registry establishes that interpretive context is present in primary reports and absent from the validation papers examined. It does not establish that carrying that context forward would have changed any published validation conclusion, or any engineering decision taken on the basis of one. That last step — from the caveat did not travel to a decision was therefore worse — is argued here, not shown. Demonstrating it would require re-running a published validation with the referent uncertainty included and showing the conclusion move. That is a tractable piece of work and it has not been done.
Absence in conference papers is not absence in practice. The search was for statements in published modelling papers. Programme-level verification and validation documentation, credibility assessments under NASA-STD-7009, internal review records and reviewer correspondence were not examined, and any of them could carry the context that the papers do not.
Seven registry entries remain PENDING (B2, B3, C4, D1–D4), and the C2 primary remains unobtained although its cited intermediate document was audited — eight unobtained source artifacts in total (§V). Verdicts may change; the versioning mechanism exists for exactly that.

What follows this registry. This series is the documentary groundwork — published first so it can be checked first — for a committed quantitative study: reconstruct a defensible result-level experimental uncertainty (u_D) for one referent above, combine it with the numerical and input uncertainties of a published comparison, covariances included, and report whether that comparison’s validation conclusion moves. In either direction: if nothing moves, that result publishes too. Until that study exists, the engineering consequence of everything above remains argued, not demonstrated — as this section already states. The study will appear as a new version under this registry’s concept DOI, with v1.0 as its prior work.

X. Corrections and additions wanted

If you worked on any of these test programmes, or have used their data, three things are genuinely useful: experiments missing from the list; entries that do not belong, and why; and information that exists but never made it into the published report — instrument specifications, calibration records, an uncertainty analysis cut for page limits, or a recollection of how a quantity was measured. That last category is the most valuable and the least recoverable.

Corrections are credited. Where one changes a conclusion, the change is recorded as a change — in §VII, under a new version of this note, with the old version preserved.

Series: Note 000 (this registry) · Note 001 — K-Site (doi:10.5281/zenodo.21895605) · Note 002 — SHIIVER (doi:10.5281/zenodo.21895647) · Note 003 — MHTB (doi:10.5281/zenodo.21895743).

ElarionX CPMS develops and evaluates engineering models for cryogenic propulsion systems, and publishes what it finds — including when the finding is about its own work. This registry is exploratory: it assesses the public evidence base, and it makes no claim it cannot trace to a document listed above. Corrections and correspondence: Luis.emc2@elarionx.com