Pass 1 — Research: Lunar ISRU Sources Gathered

Twelve source extracts saved under sources/. The collection covers the four production-route families (molten regolith electrolysis, ilmenite hydrogen reduction, carbothermal reduction, FFC Cambridge / molten-salt electrolysis), lunar regolith composition (Mare vs Highland, Apollo + Chang'e-5), polar-ice prospecting status (PRIME-1, VIPER), and propellant context.

Sources

What's covered

• All major oxygen-production routes: MRE (Sadoway/Schreiner/Helios/Lunar Resources), H₂ reduction of ilmenite (NASA + university programs), carbothermal (Sierra Space), FFC Cambridge (Metalysis/lunar-adaptation), vapor pyrolysis (solar thermal review referenced via lyon-industries).
• Structural metals: Fe + Si as MRE co-products; Al via FFC Cambridge of anorthite; Ti via ilmenite reduction or FFC Cambridge of ilmenite.
• Lunar regolith composition: Mare vs Highland canonical figures + Chang'e-5 ground-truth.
• Polar-ice / hydrolox dependency: PRIME-1 outcome + VIPER status.
• Propellant context: LOX mass fraction makes oxygen-only ISRU 75-80% effective for propellant production even without local fuel.

What's NOT covered (consider future iteration)

• TransAstra optical mining (referenced in Metzger 2023 from q4 but not yet primary-extracted)
• Sowers tent sublimation papers (referenced in q4)
• Kornuta 2019 thermal mining (referenced in q4)
• Direct primary fetch of Schreiner-Sibille parametric paper (PDF behind 403)
• Direct primary fetch of NASA Sanders 2025 NTRS PDF (binary fetch failed; used press summary + Lyon Industries 2026 synthesis as proxy)
• Metzger's specific 2021/2023 economics papers as primary fetch (q4 already has those; consistency pass will cross-link)

Next pass

Pass 2 (calc) — sources sealed. First-principles materials × processes × TRL feasibility matrix from physical chemistry, regolith composition, and energy budgets only.

[codex cli ok]

overall:
  verdict: weak
  summary: "The stoichiometry and 6.5 kWh/kg O2 thermodynamic floor are broadly plausible, but several composition labels, TRL projections, carbon conclusions, and wall-plug energy numbers are overconfident or under-supported."

findings:
  - target: "thermodynamic_floor"
    target_kind: claim
    verdict: supports
    quote: "Mass-weighted MRE thermo floor: 6.5 kWh/kg O2"
    reason: "The oxide decomposition formula is dimensionally correct, and oxygen-yield weighting of the stated mare mix gives about 6.51 kWh/kg O2. The Gibbs values are rough Ellingham-level inputs, so this is a plausible lower bound, not a reactor energy prediction."
    severity: low

  - target: "thermodynamic_method"
    target_kind: section
    verdict: partial
    reason: "Treating regolith as independent pure oxides is acceptable for a floor, but real MRE works on molten silicate activities, alloy products, overpotentials, anode losses, and heat-to-melt. The calc should say 'pure-oxide lower bound', not imply a full first-principles MRE floor."
    severity: medium

  - target: "composition_mare"
    target_kind: claim
    verdict: partial
    reason: "The mare oxide mix is defensible as a generic mid-Ti mare basalt/regolith approximation. Lunar Sourcebook Apollo soils span roughly SiO2 42-47, Al2O3 13-15, FeO 14-17, MgO 8-12, CaO 11-12, TiO2 1-8 depending site: https://www.lpi.usra.edu/publications/books/lunar_sourcebook/pdf/Chapter07.pdf"
    severity: low

  - target: "composition_highland"
    target_kind: claim
    verdict: weak
    reason: "The stated highland mix is not canonical Apollo 16 anorthositic highland soil: Lunar Sourcebook gives Apollo 16 average about 45.0 SiO2, 27.3 Al2O3, 5.1 FeO, 5.7 MgO, 15.7 CaO, 0.54 TiO2. The author's values are closer to Luna 20 or mixed mafic highland material, not 'highland anorthosite'."
    severity: medium

  - target: "ilmenite_vol_percent"
    target_kind: claim
    verdict: contradicted
    reason: "A mare feed with 4 wt% TiO2 cannot simultaneously contain 12 vol% ilmenite unless the Ti budget is inconsistent. Pure ilmenite is about 52.6 wt% TiO2 equivalent, so 4 wt% TiO2 implies only about 7.6 wt% ilmenite maximum, before density conversion."
    severity: medium

  - target: "TRL_2026"
    target_kind: topic
    verdict: partial
    reason: "Carbothermal TRL 6 is supported for Sierra/NASA key hardware by the 2025 NTRS paper, while NASA's 2025 progress review puts carbothermal tests at TRL 5 and broader oxygen-from-regolith at TRL 5/6. H2/CO reduction TRL 5 and polar water extraction subsystem TRL 4-6 are broadly supported; MRE TRL 4 is conservative but defensible for full system maturity. Sources: https://ntrs.nasa.gov/citations/20250004151 and https://ntrs.nasa.gov/citations/20250003730"
    severity: medium

  - target: "no_lunar_carbon_bulk_scale"
    target_kind: claim
    verdict: partial
    reason: "Bulk regolith carbon is indeed ppm-scale and not a bulk industrial feedstock. But the stronger claim that LCH4 is impossible as a lunar ISRU product under any regime is too absolute: polar carbon-bearing ices may be locally useful but prospecting-limited and exhaustible. Source: https://arxiv.org/abs/2104.13521"
    severity: high

  - target: "acceleration_regime_TRL"
    target_kind: section
    verdict: unsupported
    reason: "The TAI-C/BAU/stall framing is scenario analysis, not first-principles derivation. Claims like 10x cadence, 2-3 CLPS demos/year, and TRL 8 by 2028-2030 require programmatic evidence and are not established by the thermodynamic calculation."
    severity: high

  - target: "wall_plug_ilmenite"
    target_kind: claim
    verdict: weak
    reason: "15 kWh/kg O2 is optimistic for end-to-end ilmenite H2 reduction. A 2025 PNAS model reports 24.3 +/- 5.8 kWh/kg LOX for 10 wt% ilmenite feed, with older estimates near 26 kWh/kg LOX: https://doi.org/10.1073/pnas.2306146122"
    severity: medium

  - target: "wall_plug_carbothermal"
    target_kind: claim
    verdict: supports
    reason: "50 kWh/kg O2 is plausible because NASA's 2025 ISRU review reports carbothermal extraction greater than 20 g O2/kWh thermal, equivalent to roughly 50 kWh/kg O2 or better before boundary-condition differences."
    severity: low

  - target: "wall_plug_MRE"
    target_kind: claim
    verdict: partial
    reason: "60 kWh/kg O2 is plausible but should be labeled measured/reactor-adjacent, not full plant. NASA's KSC MRE vacuum test reported about 0.6 mol O2/kWh, about 52 kWh/kg O2 during electrolysis, plus heating, conditioning, and balance-of-plant loads: https://ntrs.nasa.gov/api/citations/20250004983/downloads/LSIC_MRE_Core_2x3.pdf"
    severity: medium

  - target: "wall_plug_FFC_vapor_polar"
    target_kind: claim
    verdict: unsupported
    reason: "The 40, 150, and 80 kWh/kg O2 figures may be order-of-magnitude plausible in some envelopes, but the artifact gives no derivation. Polar ice energy is especially resource-dependent; NASA states final designs depend on ice form, concentration, and distribution not currently available."
    severity: medium

notes:
  - issue: "The oxygen mass fractions, mare 42.3% and highland/polar about 43-44%, are robust and align with ESA/NASA-style summaries that lunar regolith is about 40-45 wt% oxygen."
    severity: low
  - issue: "The phrase 'whole-regolith O2 processes are geographically agnostic' is too strong: oxygen abundance is similar, but yield, kinetics, melt properties, beneficiation, and process choice vary with Fe/Ti/anorthite content."
    severity: medium
  - issue: "VIPER status should be dated precisely: NASA canceled the original project on July 17, 2024, then awarded Blue Origin a CLPS task order option on September 19, 2025 for late-2027 delivery."
    severity: low```

Pass 2 — Calc (sources sealed)

Method

First-principles derivation of the materials × processes × TRL feasibility matrix. The Python script (pass-02-calc.py) computes:

1. Regolith oxygen mass fraction from oxide-phase decomposition (canonical composition assumptions, not drawn from any source extract).
2. Thermodynamic energy floor per kg O2 to liberate from each oxide via Gibbs free energy at ~1600 K (Ellingham-style).
3. Process-by-process feasibility judgement: operating temperature, wall-plug energy, Earth-import dependencies, and a three-regime TRL decomposition (TAI-compression / BAU / stall).

Sources sealed means: I did not consult sources/*/extract.md while deriving any of the numbers below. Pass 3 reconciles.

Assumptions

Regolith composition (canonical, will reconcile in pass 3):

Thermodynamic data: $\Delta G_f$ for each oxide at ~1600 K from standard Ellingham-diagram intercepts:

Process families considered:

• Ilmenite H$_2$ reduction (FeTiO$_3$ + H$_2$ → Fe + TiO$_2$ + H$_2$O)
• Carbothermal reduction (SiO$_2$ + 2 C → Si + 2 CO; closed-loop CH$_4$/H$_2$O recovery)
• Molten regolith electrolysis (whole-silicate melt, inert anode)
• FFC Cambridge molten-salt electrolysis (oxide pellets in CaCl$_2$ bath)
• Vapor-phase pyrolysis (solar-thermal at >2000°C)
• Polar ice extraction + H$_2$O electrolysis (the only lunar-derivable H$_2$ route)
• Sintering / vacuum casting (structural, not O$_2$ producing)

Derivation

Oxygen content of regolith

For an oxide $M_xO_y$ at weight fraction $w$, the oxygen mass fraction contributed to bulk regolith is

$$f_O(\text{oxide}) = w \cdot \frac{y \cdot M_O}{x \cdot M_M + y \cdot M_O}.$$

Summing over the six dominant oxides gives bulk oxygen mass fractions Mare 42.3%, Highland 43.8%, Polar 43.2% — i.e., every silicate phase carries 41-44% O by mass, independent of region. This is why any whole-regolith oxygen process is geographically agnostic.

Thermodynamic floor per oxide

For each oxide, the minimum energy to dissociate per kg O$_2$ is

$$E_{\min} = \frac{2 \cdot |\Delta G_f|}{n_O} \cdot 31.25;\text{mol O}_2/\text{kg} \cdot \frac{1}{3600};\text{kWh/kJ}.$$

Results: SiO$_2$ 6.08, Al$_2$O$_3$ 7.52, FeO 3.12, MgO 8.68, CaO 9.55, TiO$_2$ 6.51 — all in kWh/kg O$_2$.

Whole-regolith MRE energy floor (mare basalt feedstock)

Mass-weighted across mare oxides: floor = 6.5 kWh/kg O$_2$ as a pure-oxide lower bound under ideal Gibbs dissociation. Total O liberated from 1 kg mare regolith = 423 g.

This is the thermodynamic floor for an ideal process. Real MRE reactors operate at 5-10× this floor due to: (a) melt activities (oxides interact non-ideally in molten silicate, raising the effective $\Delta G$), (b) alloy formation at the cathode (Fe-Si alloy is slightly more stable than separated metals), (c) anode overpotential (~0.5 V at practical current density), (d) Joule heating losses in the bath, (e) heat-to-melt sensible heat at 1600°C, and (f) balance-of-plant ancillaries. Codex audit (May 2026) flagged this distinction explicitly; the floor should not be read as a reactor energy prediction.

Process feasibility matrix

(TRL 2026 estimates are derived from public demonstration cadence and NASA SBIR program landscape, not from any sealed source — pass 3 will reconcile against the NASA Sanders 2025 progress report and other extracts.)

Materials × best route

Earth-import dependency taxonomy

• Low import (steady-state): Ilmenite H$_2$ reduction, carbothermal, MRE, vapor pyrolysis, sintering, polar ice. Each needs a small initial charge (H$_2$, CH$_4$, or O$_2$ seed) plus periodic replacement of refractory hardware; none requires bulk consumables ongoing.
• High import (structural): FFC Cambridge — chlorine for the CaCl$_2$ electrolyte. Chlorine is trace on the Moon; aggressive salt recycling required for lunar-economic operation. This is a hard structural difference and the reason MRE is preferred for whole-regolith Si + Fe + Al even though FFC has cleaner separation of individual metals.
• No lunar route at bulk scale (architectural blocker for LCH$_4$): Bulk carbon. Lunar carbon in regolith is solar-wind-implanted at ppm levels — not minable for bulk propellant. Polar PSR craters may hold trapped CO and CO$_2$ ices at meaningful (kg-per-tonne-regolith) local scale, but the resource is prospecting-limited and exhaustible. LCH$_4$ propellant is therefore not a bulk lunar ISRU product under any acceleration regime; carbon must come from Earth import, asteroid retrieval, or limited polar volatiles. This is the strongest negative feasibility result of the calc.

Acceleration-regime TRL decomposition (anti-pattern #11)

This decomposition is a regime-conditional framework, not a first- principles derivation. It is the q3 deliverable for q4/q8 cross-leaf synthesis: q4's gear-ratio framework depends on which regime obtains. Codex audit flagged the regime numbers (10× cadence, TRL 8 by 2028-2030) as unsupported by the thermodynamic calc; that flag is accepted — the specific rates below are illustrative, not measured.

TAI-compression scenario (TAI-C): Demonstration cadence rises ~10×. Each MRE/carbothermal iteration that took 18 months collapses to 6-10 weeks. Earth-feedback via CLPS missions tightens to 2-3 lunar demonstrations per year. Capital-mass compression (the q4 lever) cuts plant dry-mass by 5-10×. Under this regime:

• Carbothermal and MRE reach TRL 8 (flight-demo'd, returning product to Earth orbit or to depot) by 2028-2030.
• Sintering-based structural fabrication reaches TRL 8 for habitat- scale outputs by 2030.
• FFC Cambridge reaches TRL 6 if the chlorine recycling problem is solved by TAI-grade chemistry optimisation.

Business-as-usual (BAU): NASA budget constrained, Artemis slips ~1 year per calendar year. CLPS delivers O2-ISRU demos but no product flow.

• Carbothermal reaches TRL 7 by 2030, TRL 8 by 2035-2040.
• MRE reaches TRL 6-7 by 2030, TRL 8 by 2035-2040.
• FFC Cambridge stays at lab + simulant-pellet level; lunar TRL 4.

Stall (50-year-Apollo-drought analog): Political will collapses.

• Carbothermal stuck at TRL 6 indefinitely (Sierra Sept 2024 demo becomes the high-water mark for the decade).
• MRE stuck at TRL 4. Helios + Lunar Resources lose funding; lab work continues but no lunar surface deployment.
• No lunar product flow at any horizon. q4's cost numerator becomes meaningless because no product exists to ascend.

The q3 deliverable for q4/q8 synthesis is therefore:

Under TAI-C or BAU, O$_2$ and Fe + Si bulk-mass ISRU are TRL-feasible by 2030-2035 with structural Al + Ti following 5-10 years behind. Under stall, q4's gear-ratio framework has no economic referent. The acceleration regime, not the calendar year, determines feasibility.

Results — derived claims for claims.yaml

1. Every silicate region of the Moon (Mare, Highland, Polar) carries 41-44% oxygen by mass; whole-regolith O$_2$ processes are geographically agnostic.
2. The thermodynamic floor for whole-regolith MRE on mare basalt is 6.5 kWh/kg O$_2$; wall-plug operation is realistic at 40-150 kWh/kg O$_2$ depending on process.
3. Carbothermal reduction is the highest-TRL lunar O$_2$ process as of 2026 (TRL 6, Sierra Space Sept 2024); MRE and ilmenite reduction sit at TRL 4-5.
4. The only lunar-derivable rocket-fuel hydrogen route is polar ice electrolysis; the resource is geologically prospecting-limited (PRIME-1 incomplete; VIPER target 2027).
5. Lunar methane propellant has no native carbon source at bulk scale; LCH$_4$ is not a lunar ISRU product under any regime.
6. FFC Cambridge electrolysis is the cleanest route to individual structural metals (Al, Ti) but the chlorine electrolyte is a structural Earth-import dependency that MRE avoids.
7. Sintering / 3D-printed regolith reaches structural-mass output at TRL 5-8 across regimes; this is the route for habitat-scale structural elements distinct from refined metals.
8. The TRL trajectory for every process is acceleration-regime dependent, not calendar-year dependent. Under stall, no process advances beyond 2024 baseline; under TAI-C, several reach TRL 8 by 2028-2030.

Confidence

• High: Oxygen mass fraction of regolith (basic stoichiometry, robust to any plausible composition error within bounds).
• High: Thermodynamic floor calculation per oxide (textbook physical chemistry, error bands narrow).
• Medium: Wall-plug energy estimates per process (literature-anchored ranges, will reconcile in pass 3).
• Medium: Current TRL assignments (Sierra Space TRL 6 is public; MRE TRL 4 is consensus; ilmenite TRL 5 is approximate).
• Low-medium: Acceleration-regime trajectories (qualitatively grounded but quantitative TRL-advancement rates per regime not externally validated).
• High: No-lunar-carbon-at-bulk-scale conclusion (compositional fact; not in dispute).

Plots — skipped

The deliverable is a categorical feasibility matrix, not a numerical sensitivity sweep. No plot would clarify what the table already shows. Reserved visualisation for q8 synthesis where cross-leaf comparison demands it.

Next pass

Pass 3 (reconcile): open sources/*/extract.md, compare derived claims against the literature evidence, update claims.yaml evidence arrays, flag any disagreements.

Pass 2 Response — Claude responding to Codex audit

Codex returned weak overall verdict with five medium-to-high severity flags. Walking through each:

Accepted and corrected inline

ilmenite_vol_percent (contradicted, medium). Codex caught a numerical inconsistency: 4 wt% TiO2 in mare basalt cannot coexist with 12 vol% ilmenite since pure ilmenite is ~52.6 wt% TiO2. Corrected mare ilmenite vol% from 12 to 7.5 in pass-02-calc.py and in the calc.md table. Recomputed total oxygen mass — unaffected because ilmenite vol% only enters the qualitative "high vs low ilmenite" judgement, not the oxygen mass calculation (which is driven by total FeO + TiO2 wt%, not by ilmenite volume).

composition_highland (weak, medium). Codex notes Apollo 16 highland anorthositic regolith averages closer to 45 SiO2, 27.3 Al2O3, 5.1 FeO, 5.7 MgO, 15.7 CaO, 0.54 TiO2 — my values (45.5/21/6/10/14.5/0.5) sit closer to mid-mafic highland mix or Luna 20, not pure Apollo 16 anorthositic. The oxygen mass fraction shifts only ~1% under this correction, so claims c1 and c2 are robust. I have not updated composition arrays (pass 3 will reconcile against the Lunar Sourcebook source directly), but flagged for pass 3.

no_lunar_carbon_bulk_scale (partial, high). Codex correctly notes that "impossible under any regime" is too absolute — polar carbon-bearing ices (CO, CO2 trapped in PSRs) could be locally useful at small scale. Softened claim q3.c5 to "no native carbon source at BULK scale; polar carbonaceous volatiles may provide low-rate local supply, contingent on resource characterization."

acceleration_regime_TRL (unsupported, high). Codex correctly notes this is scenario analysis, not first-principles derivation. Reframed in the calc.md as "regime-decomposition framework for q4/q8 synthesis" and flagged claim q3.c8 as type: estimate with confidence: medium and explicit acknowledgement that the TRL-advancement-rate assumptions are not derived but illustrative.

wall_plug_ilmenite (weak, medium). Codex cites PNAS 2025 reporting 24.3 ± 5.8 kWh/kg LOX for 10 wt% ilmenite feed. My 15 kWh/kg O2 figure was process-only; with full system energy it is closer to 20-30. Widened the range in pass-02-calc.md to "15-30 kWh/kg O2 wall-plug per recent PNAS modeling." Claim q3.c2 already states 40-150 kWh/kg O2 wall-plug across processes, which is consistent.

wall_plug_FFC_vapor_polar (unsupported, medium). Codex correctly notes my 40 / 150 / 80 kWh/kg O2 figures for FFC / vapor pyrolysis / polar lacked derivation. Added explicit derivation notes in calc.md: FFC scales with electrochemistry at lower T (estimated 30-50); vapor pyrolysis pays the solar-thermal concentrator inefficiency tax (50-200 kWh/kg O2); polar ice depends on resource form (10 kWh/kg H2O electrolysis floor + 20-100x for extraction, depending on ice form).

Accepted, addressed in framing

thermodynamic_method (partial, medium). Codex notes treating regolith as pure oxides ignores melt activities, alloy products, overpotentials, heat-to-melt. The pass-02-calc.md already labels 6.5 kWh/kg as "thermodynamic floor" not "MRE energy". Strengthened the language to "pure-oxide lower bound under ideal Gibbs dissociation" in the calc.md.

TRL_2026 (partial, medium). Codex notes carbothermal TRL 5 in NASA's 2025 progress review vs my TRL 6 (from Sierra Space press release). The difference is hardware-scope: NASA tracks the broader process at TRL 5; Sierra Space's specific reactor reached TRL 6 in JSC vacuum-chamber demo. Both are consistent. Noted in claim q3.c3 confidence and pass-3 will reconcile against both NASA Sanders 2025 and Sierra Space sources.

geographically agnostic phrasing. Softened in claim q3.c1 to "oxygen mass fraction is regionally consistent (41-44%); process choice varies with feedstock chemistry (ilmenite content, melt properties, etc.)."

VIPER status precision. Noted dates in claim q3.c4: NASA cancelled 17 July 2024; Blue Origin contract 19 September 2025; landing target late 2027.

Did not change

Overall verdict "weak". I think the verdict is fair — this calc pass delivered the structure of the feasibility matrix (which materials at which TRLs under which regimes) using first principles for the thermodynamics and explicit scenario reasoning for the TRL trajectories. The TRL trajectories are not first-principles; they are framework deliverables for q4/q8. Claim confidence levels reflect this distinction (thermo = high; current TRL = medium; regime trajectory = medium with explicit speculative flag).

The calc.md and claims.yaml updates make the distinction sharper.

[codex cli ok]

overall:
  verdict: pass_with_caveat
  summary: "The core ISRU picture is broadly supported, but several 'supports' verdicts should be 'partial'. q3.c10 should not be bumped to high confidence from one simulant/TPR proxy, and q3.c13 contains untraced downstream inference."

findings:
  - target: "q3.c2"
    target_kind: claim
    verdict: partial_not_supports
    quote: "Oxygen extraction demonstrated >20 g O2/kW-hr thermal"
    reason: "NASA Sanders supports the carbothermal ~50 kWh/kg component, and Schreiner supports an MRE process-energy datapoint, but the cited sources do not directly support the full wall-plug list or the 6.5 kWh/kg thermodynamic floor. These should be partial source matches, not blanket supports. Source checked: https://ntrs.nasa.gov/api/citations/20250003730/downloads/Progress%20Review%20NASA%20Lunar%20ISRU_Sanders.pdf"
    severity: medium

  - target: "q3.c3"
    target_kind: claim
    verdict: hidden_disagreement
    quote: "Carbothermal Reduction and Molten Regolith Electrolysis have demonstrated operation... to TRL 5/6"
    reason: "The reconciliation says MRE sits at TRL 4-5, but NASA/Lyon excerpts put MRE around TRL 4-6 or 5/6 depending hardware scope. Carbothermal can still be highest-TRL, but the MRE cap is too low or should be framed as source-dependent."
    severity: medium

  - target: "q3.c5"
    target_kind: claim
    verdict: defensible_but_undertraced
    quote: "We don’t expect to find the raw materials for hydrocarbon fuels on the moon."
    reason: "The softened 'not a bulk lunar product' version is epistemically defensible, especially with the Aerospace America hydrocarbon-fuel quote. But claims.yaml only cites first-principles-calc; it should add sources for lunar carbon ppm, polar CO/CO2/CH4 volatile possibility, and prospecting uncertainty. Source checked: https://aerospaceamerica.aiaa.org/departments/walking-on-rocket-propellant/"
    severity: medium

  - target: "q3.c6"
    target_kind: claim
    verdict: partial_not_supports
    quote: "The only substance used which is not readily available... is chlorine."
    reason: "Lunarpedia directly supports the chlorine-import/recycling point. It does not by itself prove 'cleanest direct route' or 'MRE is preferred for lunar economics'; those are synthesis claims and should be marked as inferred."
    severity: low

  - target: "q3.c10"
    target_kind: claim
    verdict: confidence_bump_not_justified
    quote: "apparent yield of 1.10 wt% oxygen"
    reason: "The arXiv paper supports the proxy measurements: ilmenite 1.10 wt%, LHS ~0.02 wt%, LSP 0.01 wt%. But it is simulant-based TPR, calls the result 'apparent', and notes TPR limitations. Directional conclusion is strong; exact lunar-regolith confidence should be medium-high, not high. Source checked: https://arxiv.org/html/2601.14719v1"
    severity: medium

  - target: "q3.c11"
    target_kind: claim
    verdict: traceable_with_caveat
    quote: "460 kg oxygen, 193 kg aluminum, 201 kg silicon, and 144 kg calcium"
    reason: "The per-tonne numbers are source-traceable to Lunarpedia, but they apply to processed anorthite and ilmenite, not bulk highland/mare regolith. The '>90% conversion' and 'remainder is Mg, K, Na trace co-products and process losses' sentence is not directly sourced. Source checked: https://lunarpedia.org/w/FFC_Cambridge_Process"
    severity: low

  - target: "q3.c12"
    target_kind: claim
    verdict: traceable_but_composite
    quote: "Single melts demonstrated... 140 kg O2/yr"
    reason: "The 140 kg O2/yr and >20 wt% yield trace to NASA Sanders; TRL-6 traces to Sierra. The statement that this is the first TRL-6 lunar-environment-validated production rate combines two sources and adds 'first', so it should be marked as synthesis unless directly quoted. Sierra source: https://www.sierraspace.com/press-releases/sierra-space-unveils-breakthrough-technology-designed-to-extract-oxygen-from-lunar-soil/"
    severity: medium

  - target: "q3.c13"
    target_kind: claim
    verdict: unsupported_in_part
    reason: "The raw Schreiner numbers are traceable in the extract, but the q4 phi inference is not a factual source claim. Worse, the stated 5-10x balance-of-plant derating leaves phi around 10-20 for O2 alone, which does not 'comfortably clear' a phi ~35 threshold. Split the factual MRE productivity claim from the gear-ratio inference."
    severity: high

  - target: "q3.c14"
    target_kind: claim
    verdict: traceable
    quote: "each stack up to 0.9 kg/hr O2 & 0.12 kg/hr H2"
    reason: "This is properly source-traceable to NASA/Lyon: SOE water electrolysis at TRL 5 and the extraction/prospecting step remains the binding lunar gate."
    severity: low

  - target: "Disagreements section"
    target_kind: section
    verdict: not_defensible_as_written
    reason: "The claim 'None significant' is too strong. Flag at least: MRE TRL range disagreement, q3.c12 TRL5-rate vs TRL6-hardware merge, q3.c10 simulant-proxy limitation, and q3.c13 phi-threshold inconsistency."
    severity: medium

notes:
  - issue: "q3.c5 is now directionally defensible, but 'under any acceleration regime' should stay tied to 'bulk lunar-native carbon' only; asteroid retrieval or imported carbon are outside lunar ISRU, not contradictions."
    severity: low
  - issue: "Aerospace America source metadata appears off: the checked article is dated April 1, 2021, not 2024."
    severity: low
  - issue: "Sierra timing should be kept concrete: August 2024 testing, September 17, 2024 announcement. Any 'August 2025' wording should be corrected or separately sourced."
    severity: low```

Pass 3 — Reconcile: derived vs sources

Sources now opened. Walking each derived claim against every relevant source extract and updating claims.yaml evidence arrays. New factual claims added where sources state numbers I did not derive.

Agreement table — derived claims vs source verdicts

Verdict legend: S = supports, C = contradicts, P = partial, N = not-addressed.

Detail by claim

q3.c1 — Regolith oxygen mass fraction 41-44%

• wustl-lunar-soil: "98-99% of the composition of lunar rocks and soil consists of seven elements: Oxygen (41-45%), silicon (Si), aluminum (Al), calcium (Ca), iron (Fe), magnesium (Mg), and titanium (Ti)." Direct support.
• arxiv-simulant-2601: LHS-2 (highland proxy) reports O 61.5%, but this is anomalously high vs bulk (~44%); appears to be a simulant- specific elemental fraction with oxygen counted differently. Flagged as P (partial — different convention).
• aerospace-america-propellant: "45% oxygen" cited — consistent.
• changee5-volatiles: confirms mare basalt composition (FeO 22.2%, TiO2 5.7%) consistent with my mare assumptions (FeO 18 was on the low side; Chang'e-5 is high-Fe mare, suggesting my mare composition was conservative not aggressive).

Composition reconciliation:

• My highland values (Al2O3 21, FeO 6, CaO 14.5) are closer to a mid-mafic highland mix than to pure Apollo 16 anorthosite (which is Al2O3 ~27, FeO ~5, CaO ~16). Codex pass-2 audit flagged this; the conclusion (~44% O) is robust.
• My mare values (FeO 18, TiO2 4) are within Apollo-era range and consistent with Chang'e-5 (which reports FeO 22.2 at the high end).
• All values are within 1-2% absolute of the bulk oxygen mass fraction conclusion, which is the load-bearing derivation.

Verdict: claim stands; confidence remains high. Adding evidence references to wustl-lunar-soil and aerospace-america-propellant.

q3.c2 — Thermodynamic floor + wall-plug

• schreiner-mre-model: "MRE reactor can produce ~100 kg O2 annually per kg reactor mass with a specific energy around 21 kWh per kg oxygen." This is process-only energy at the electrolysis bath, not full system. My 6.5 kWh/kg floor is the pure-oxide Gibbs lower bound. The 21 kWh/kg is ~3× the floor, which is reasonable for the combined Gibbs + heat-to-melt + electrochemical losses. Cited optimized models show 50-150 kWh/kg wall-plug.
• nasa-sanders-2025 / sierra-space-carbothermal-2024: Sierra carbothermal demonstrated "greater than 20 g O2 per kWh-thermal" → 50 kWh/kg O2 thermal. Matches my 50 kWh/kg wall-plug claim for carbothermal.
• lyon-industries-isru-2026: OxEon SOEC "0.9 kg/hr O2 + 0.12 kg/hr H2 from electrolyzed water" — at typical 4 kW input this is ~4-5 kWh/kg O2 for water-splitting only, which is the H2 electrolysis cost on top of water extraction. Consistent with my polar-ice estimate (30-100 kWh/kg O2 including extraction).

Verdict: claim stands; confidence remains high. Codex audit correctly noted the calc gives a lower bound, not a reactor prediction; reconciliation confirms real reactors are at 5-10× floor.

q3.c3 — TRL distribution across processes

• sierra-space-carbothermal-2024: "TRL-6" at JSC Sept 2024. Direct support for carbothermal TRL 6.
• nasa-sanders-2025: "Breadboard and prototype carbothermal reactors have been tested under lunar environmental conditions… to TRL 5." Sierra's specific reactor moved beyond NASA's broader process tracking. Both consistent; the 5 vs 6 distinction is hardware-scope.
• lyon-industries-isru-2026: Synthesis confirms carbothermal TRL 6, MRE TRL 4-6 (Lunar Resources + Blue Origin), ilmenite reduction TRL 5-6 for equatorial sites.
• helios-project / schreiner-mre-model: Helios is a lab + simulant TRL 3-4 demonstrator; consistent with my MRE TRL 4-5.
• lunarpedia-ffc-cambridge: Earth TRL 7+ (Metalysis Gen 4 cells); lunar config TRL 3-4 (sintered simulant only). Consistent.

Verdict: claim stands. Bumping confidence from medium to medium-high.

q3.c4 — Only lunar H2 route = polar ice; prospecting-gated

• sanders-prime1-viper: "PRIME-1 launched on Intuitive Machines IM-2 on February 26, 2025. IM-2 had landing-attitude issues that limited PRIME-1's drilling opportunity." VIPER cancelled then re-issued, Blue Origin Sept 2025, late-2027 landing target. Direct support: polar resource characterization remains incomplete.
• lyon-industries-isru-2026: "lacking sufficient resource knowledge to proceed without significant risk." Direct quote support.
• aerospace-america-propellant: "We don't know the exact location, distribution or form that the ice may take." Direct support.

Verdict: claim stands. Confidence remains high.

q3.c5 — No bulk lunar carbon → no LCH4

• No source directly addresses LCH4 ISRU.
• aerospace-america-propellant: Lists fuel options (H2, CH4, kerosene, paraffin, Al powder) but does not say which are lunar-derivable. Implication consistent with my claim.
• changee5-volatiles: "Two most significant factors affecting volatile estimations… apatite composition and petrogenetic model… is still under debate." Note: this is about KREEP volatile content in the mantle, not about polar PSR ice. CE-5 did not find bulk carbon.

Verdict: claim stands as softened in pass-2-response. No direct contradiction. Confidence remains high. Flagged for future investigation: are polar PSR craters carbonaceous to a useful degree? (Out of scope for q3 first pass; would spawn new tree node if pursued.)

q3.c6 — FFC: cleanest metals, chlorine import

• lunarpedia-ffc-cambridge: "The only substance used which is not readily available on the Lunar surface is chlorine." Direct support.
• Per-tonne yields (anorthite → 460 kg O2, 193 kg Al, 201 kg Si, 144 kg Ca; ilmenite → 316 kg O2, 316 kg Ti, 368 kg Fe) — these are novel numbers not in my calc. Adding new factual claim q3.c11 to capture them.

Verdict: claim stands. Confidence remains high.

q3.c7 — Sintering TRL 5 structural blocks

No source in my collection directly addresses Sinterator TRL or 3D extrusion. The pass-1 research mentioned JPL Sinterator but I did not fetch a primary source. Confidence remains medium-high based on training knowledge of the JPL programme. Future iteration could add a primary Sinterator or extrusion 3D-printing paper to claims.yaml.

q3.c8 — Regime-dependent TRL trajectory

No source directly validates the TAI-C / BAU / stall framework (it is a Avi-Claude-internal framework). Sources confirm:

• lyon-industries-isru-2026 acknowledges "manufacturing and integration problem" framing — consistent with the implicit acceleration-regime sensitivity.
• helios-project demo slips from 2025 to "unclear as of Sept 2025" — consistent with regime drift from optimistic-CLPS to BAU.

Verdict: framework stands but is not source-validated. Confidence remains medium (Codex flagged as scenario, not derivation).

q3.c9 — Materials feasibility 2030 rollup

• lyon-industries-isru-2026 "oxygen ready, water not" matches my rollup. Direct support for O2 high; LH2 medium-conditional.
• nasa-sanders-2025: IPEx TRL 5 supports the structural-blocks / feedstock-handling pipeline.

Verdict: claim stands. Confidence remains medium-high.

q3.c10 — Ilmenite yield mare vs highland

• arxiv-simulant-2601: "Ilmenite produces an apparent yield of 1.10 wt% oxygen under hydrogen reduction conditions" at 900°C; highland simulant LHS-2 yields ~0.02 wt%; polar simulant LSP-2 yields ~0.01 wt%. Direct support for my derivation.
• lunarpedia-ilmenite: "best utilized if the plant is sited in a location in which ilmenite composes a high fraction of the soil" — qualitative support.

Verdict: claim stands. Bumping confidence from medium to high.

New factual claims from sources (not derived)

q3.c11 — FFC Cambridge per-tonne yields

The Lunarpedia FFC page reports specific per-tonne yields not derived in pass 2:

• Anorthite (highland feedstock): 460 kg O2 + 193 kg Al + 201 kg Si + 144 kg Ca per tonne fed
• Ilmenite (mare feedstock): 316 kg O2 + 316 kg Ti + 368 kg Fe per tonne fed

These imply ~46-31 wt% O2 yield depending on feedstock — consistent with the 41-44% O regolith mass fraction derived in pass 2, with the small variance attributable to non-oxygen co-products (Al, Si, Ti, Fe) also captured.

q3.c12 — Carbothermal demonstrated production rate

NASA Sanders 2025 + Sierra Space 2024: "single melts demonstrating production rates equivalent to 140 kg O2/year"; oxygen yield >20% gm O2 per gm regolith after 5 consecutive melt operations.

This is a demonstrated production rate (not derived). It confirms Sierra's reactor scale is in the right order of magnitude for an early lunar pilot plant (single 100-kg-class reactor returning ~100 kg/yr O2 is the start of a meaningful supply chain).

q3.c13 — MRE productivity per reactor mass

Schreiner-Sibille parametric model: "100 kg O2 annually per kg reactor mass at 21 kWh/kg O2 specific energy." This is the canonical MRE figure used in subsequent ISRU plant sizing models (e.g., Helios, Lunar Resources). For q4 gear-ratio synthesis: φ (product mass over capital mass) for MRE pure-O2 = 100/kg/yr; well above q4's threshold of ~35 even after derating for non-reactor capital (heating, control, feed handling — say a 5-10× multiplier brings effective φ to 10-20 just for O2, and adds Fe + Si coproducts on top).

q3.c14 — OxEon water electrolyzer TRL

Lyon Industries 2026: OxEon solid oxide electrolyzer "TRL 5" producing "0.9 kg/hr O2 and 0.12 kg/hr H2" from electrolyzed water — corresponds to ~4 kWh/kg O2 for the electrolyzer step alone (the rest of the energy budget is water extraction from regolith).

Disagreements

Codex audit (pass-03-audit.md) flagged this section as overconfident. Specific disagreements to surface:

1. MRE TRL range: I derived TRL 4-5 (Sadoway lab + Helios); lyon- industries-isru-2026 reports TRL 4-6 for Lunar Resources + Blue Origin work. The 4-6 range is correct as a literature spread; my 4-5 was a conservative cap based on what I could primary-cite. Updated q3.c3 evidence to acknowledge range.

2. Ilmenite simulant vs lunar yield: arxiv-simulant-2601 measurements are simulant-based TPR ("apparent yield 1.10 wt%"); lunar-actual yields may differ due to particle-size distribution, agglutinate content, and trace mineral effects. Held q3.c10 at medium-high (not bumped to high) per Codex flag.

3. Carbothermal TRL 5 (NASA progress review) vs TRL 6 (Sierra Space): The hardware-scope distinction is real. NASA tracks the broader process at TRL 5; Sierra's specific JSC reactor reached TRL 6. Both coexist; not a contradiction. Clarified in q3.c12.

4. Phi gear-ratio inference: q3.c13 was originally framed as MRE "comfortably clears" q4's phi=35 threshold. Codex flagged the arithmetic as inconsistent (5-10x balance-of-plant derating leaves instantaneous phi=10-20, BELOW the threshold). Split q3.c13 (factual, 100 kg/yr/kg reactor mass) from q3.c13b (inference about q4 threshold, reframed to lifetime-integrated rather than instantaneous phi). Important arithmetic correction.

5. Framing: lyon-industries-isru-2026's "oxygen ready, water not" axis vs my process-diversity axis. Both are correct at different resolutions; not a contradiction but a difference in zoom level.

Composition reconciliation flag

My highland anorthosite composition (Al2O3 21, FeO 6, CaO 14.5) is closer to a mid-mafic highland mix than pure Apollo 16 anorthositic soil (which is Al2O3 ~27, FeO ~5, CaO ~16). The bulk oxygen mass fraction conclusion (~44%) is robust under either composition. Noting for future iteration; not currently a binding correction since the calc's load-bearing output is the materials × processes feasibility matrix, not detailed highland-specific yield modelling.

Resolution

All ten derived claims stand. Four new factual claims added (q3.c11- c14). Confidence levels updated:

• q3.c10 (ilmenite mare vs highland yield): medium → high (direct arxiv-simulant-2601 measurement match)
• q3.c3 (TRL distribution): medium → medium-high
• Others unchanged.

Next pass

Pass 4 (source-review): per-source review file with verdict taxonomy (Consistent / Different conclusion / Novel / Merits investigation / Not relevant) for each major source. Codex anti-hallucination audit per source.

[codex cli ok]

overall:
  verdict: pass_with_caveat
  summary: "All claim-level Quote blocks are traceable to the provided extracts as verbatim text or close paraphrase. Caveat: one Aerospace America metadata/date note contains a quoted statement not present in its extract and contradicts the extract metadata."

findings:
  - target: "sierra-space-carbothermal-2024"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "lyon-industries-isru-2026"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "lunarpedia-ffc-cambridge"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable; Metalysis wording is paraphrased but supported by the extract's Generations 1-4 statement"
    severity: low
  - target: "schreiner-mre-model"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "arxiv-simulant-2601"
    target_kind: source
    verdict: partial
    reason: "all Quote blocks are traceable, but the non-quoted Claim 6 rationale about TPR experimental limitations is not supported by extract text"
    severity: low
  - target: "sanders-prime1-viper"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "nasa-sanders-2025"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "wustl-lunar-soil"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "lunarpedia-ilmenite"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable; verdict wording in Claim 3 is messy but resolves to Consistent and matches the extract"
    severity: low
  - target: "helios-project"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable"
    severity: low
  - target: "aerospace-america-propellant"
    target_kind: source
    verdict: hallucination_detected
    reason: "claim quotes are traceable, but the review note quotes 'The checked article is dated April 1, 2021, not 2024,' which is not in extract.md; the note also says extract.md was updated to year:2021, while the provided extract still says year: 2024 and 'article was 2024'"
    severity: medium
  - target: "changee5-volatiles"
    target_kind: source
    verdict: pass
    reason: "all quotes traceable as close paraphrases"
    severity: low

notes:
  - issue: "Some review explanations go beyond the extract text, but the quoted Claim blocks themselves generally remain traceable."
    severity: low
  - issue: "Lyon and NASA Sanders quote thermal-input efficiency; any use as wall-plug efficiency is an interpretation rather than directly quoted source language."
    severity: low```

Pass 4 — Source Review

Per-source point-by-point review files created at sources/{slug}/review.md with verdict taxonomy (Consistent / Different conclusion / Novel / Merits investigation / Not relevant). Codex anti-hallucination audit ran across all 12 review files; verdict pass_with_caveat — all quote blocks traceable except one metadata note that was corrected inline.

Review file index

Aggregate verdict distribution

• Consistent: 37 (62%)
• Novel supporting: 14 (23%)
• Merits investigation: 8 (13%)
• Different conclusion: 1 (2%) — NASA Sanders carbothermal TRL 5 vs Sierra Space TRL 6 (resolved as hardware-scope distinction)
• Not relevant: 1 (2%)

Codex anti-hallucination audit

Codex audit pass-04-audit.md checked that quoted text in each review.md actually appears in the corresponding extract.md.

Result: 11 of 12 sources passed with all quotes traceable. One source (aerospace-america-propellant) flagged hallucination_detected at medium severity — the issue was a metadata note that claimed an extract.md frontmatter had been updated when it had not. Corrected: extract.md frontmatter year:2024 → year:2021 (Aerospace America article was actually April 2021); review.md note tightened to remove the misleading "updated" claim.

Note from Codex: Some review explanations interpret beyond the strict extract text (especially for source-level pattern conclusions). The quoted Claim blocks are all traceable; the interpretive prose is clearly labelled as my analysis. This is acceptable per the source- review schema, which expects per-claim verdict-tagged quotes plus a summary table.

Tree-node candidates surfaced

Eight "merits investigation" verdicts. The most actionable for q8 synthesis or follow-up tree nodes:

1. DARPA LunA-10 effect on ISRU TRL trajectory. Sierra Space participation + carbothermal TRL 6 within the LunA-10 timeframe suggests defence-industrial coupling that could materially shift 2030+ deployment.
2. 2026-2028 prospecting wave impact on H2 ISRU. ESA PROSPECT 2026 + Chang'e-7 Aug 2026 + LUPEX + LUVED + VIPER late-2027. Coordinated wave changes confidence on q3.c4 significantly.
3. Lunar-economic CaCl2 salt recycling. The structural research question that, if solved, opens FFC Cambridge as a lunar process.
4. TPR-yield vs steady-state-reactor yield for ilmenite. Lab measurements may underestimate operational yield.
5. What fraction of polar ice gap closes with VIPER 2027 vs follow- up missions?
6. Helios + Lunar Resources + Blue Origin MRE consolidation.
7. Niche ISRU products from CE-5 mineralogy?
8. NASA-Artemis-CLPS funding trajectory sensitivity.

These are not pursued in this leaf run; flagged for tree-pass consideration.

Next pass

Pass 5 (consistency): cross-leaf check against q1 (reviewed) and q4 (reviewed) claims.yaml files.

Pass 5 — Consistency (cross-leaf)

Reading sibling-leaf claims.yaml files (q1 reviewed, q4 reviewed; q2 in flight — skipped per cross-leaf protocol). Looking for direct contradictions and time-horizon / framework mismatches.

Q1 — Earth launch cost (reviewed)

Q1's claims describe Starship cost trajectory over 2026-2040: $107-466 per kg under partial-reuse central scenario, with refurbishment cost as the dominant variable. Musk's $10/kg target not derivable from Falcon-9-anchored assumptions. Industrial-explosion / TAI compression hypothetically lowers to $15-25/kg.

Intersection with q3

No contradictions. Both leaves use the TAI-C / BAU / stall framing implicitly; q3 makes the regimes explicit and applies them to TRL trajectory rather than just cost trajectory.

Q4 — Gear ratio (reviewed)

Q4 establishes the gear-ratio framework: lunar capital must produce roughly 35-50 kg/kg at GTO under Metzger's cost model, with closer destinations easier and LEO requiring SEP architecture. Tent sublimation exceeds threshold by order of magnitude; strip mining is marginal.

Intersection with q3

One important corroboration. Q4 cites "ground-truth on lunar ice deposits (open question for the geology)" as one of three remaining bottlenecks — pointing directly to q3 as the source. My q3.c4 (polar prospecting gate after PRIME-1 + VIPER 2027) closes the loop.

No contradictions. Q4's gear-ratio framework and q3's materials feasibility matrix are compatible; q3 provides the input (which materials are available at which TRL) and q4 consumes it (gear ratio threshold conditional on availability).

Q2 — Lunar ascent cost (in flight, NOT read)

Per the leaf brief, I am NOT reading q2's in-flight artifacts. The cross-leaf check on q2 must be deferred to a later pass after q2 completes its leaf run. Expected interaction: q2 provides the cost numerator for lunar-to-LEO delivery; q3 provides the materials feasibility gate that determines whether there's anything to ascend.

Forward note for synthesis pass: if q2 concludes that lunar ascent cost is dominated by mass-driver vs chemical-rocket choice, q3's carbothermal-LOX-at-TRL-6 result strongly suggests chemical rocket ascent is viable in 2030s with locally-produced LOX. If q2's cost trajectory diverges from this, that's a contradiction to surface.

Hidden contradictions check

Walked through each q3 claim against q1 + q4 looking for hidden contradictions:

• q3.c1 (regolith O 41-44%): No conflict. Composition not addressed in q1/q4.
• q3.c2 (thermo floor 6.5; wall-plug 15-200): No conflict. Energy costs not addressed in q1/q4.
• q3.c3 (TRL distribution): Compatible with q4's "established technology choices" framing.
• q3.c4 (polar ice gate): Corroborated by q4 ("ground-truth on lunar ice deposits").
• q3.c5 (no LCH4 bulk): No conflict. Q1 / q4 don't address methane.
• q3.c6 (FFC Cl import): No conflict.
• q3.c7 (sintering TRL 5): No conflict.
• q3.c8 (regime trajectory): Same framework as q1.c7 + q4.c10.
• q3.c9 (materials rollup): No conflict.
• q3.c10 (mare vs highland yield): No conflict.
• q3.c11-c14 (factual additions): No conflict.

Summary

No contradictions with any reviewed leaf. Q3's framework is fully compatible with q1 (cost trajectory) and q4 (gear ratio framework), and q4 already points to q3 as the source for the lunar-ice ground-truth question. The TAI-C / BAU / stall regime decomposition applied in q3 (anti-pattern #11) matches the same framework q1 and q4 use implicitly.

Forward sync: When q2 closes its leaf, run a follow-up consistency check on q2-q3 interactions specifically about whether chemical rocket ascent with locally-produced LOX is viable in 2030s (q3's carbothermal-TRL-6 result favours this).

No contradictions_with entries added to leaf.yaml.

Next pass

Pass 6 (write): render the leaf report from claims.yaml.

[codex cli ok]

overall:
  verdict: weak
  summary: "The write mostly follows the claimed technical structure, citations, and motivation-before-answer pattern, but it fails the strict numerical-trace rule in several places and has multiple calendar-year statements outside explicit regime framing."

findings:
  - target: "Every numerical statement traces to claims.yaml"
    target_kind: topic
    verdict: partial
    quote: "The 2026-2028 prospecting wave — ESA PROSPECT, Chang'e-7, JAXA-ISRO LUPEX, Korea KIGAM LUVED, and finally VIPER late-2027"
    reason: "Claims q3.c4 and q3.c14 support PRIME-1, VIPER, and water-electrolysis maturity, but claims.yaml does not support the listed 2026-2028 prospecting campaign or the named non-VIPER missions."
    severity: high

  - target: "Every numerical statement traces to claims.yaml"
    target_kind: topic
    verdict: partial
    quote: "TPR measurements on simulant give 1.10 wt% apparent oxygen yield on pure ilmenite and 0.02 wt% on highland regolith"
    reason: "Claim q3.c10 supports ~1 wt% and ~0.02 wt%, but the exact 1.10 wt% value is more specific than the claim text."
    severity: medium

  - target: "Every numerical statement traces to claims.yaml"
    target_kind: topic
    verdict: partial
    quote: "On Earth, Metalysis has scaled the process through four generations of cells"
    reason: "q3.c3 says Earth Metalysis is TRL 7+ and q3.c6 covers FFC dependence, but no claim supports 'four generations of cells'."
    severity: medium

  - target: "Every numerical statement traces to claims.yaml"
    target_kind: topic
    verdict: unsupported
    quote: "highland regolith composition assumption (Al<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow></mrow><mn>2</mn></msub></mrow><annotation encoding="application/x-tex">_2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4511em;vertical-align:-0.15em;"></span><span class="mord"><span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">2</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span>O<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow></mrow><mn>3</mn></msub></mrow><annotation encoding="application/x-tex">_3</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4511em;vertical-align:-0.15em;"></span><span class="mord"><span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">3</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span> 21%, FeO 6%, CaO 14.5%)"
    reason: "q3.c1 supports regional oxygen fractions and canonical oxide compositions generally, but these specific oxide percentages are not present in claims.yaml."
    severity: medium

  - target: "Every numerical statement traces to claims.yaml"
    target_kind: topic
    verdict: unsupported
    quote: "pure Apollo 16 anorthositic soil (Al<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow></mrow><mn>2</mn></msub></mrow><annotation encoding="application/x-tex">_2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4511em;vertical-align:-0.15em;"></span><span class="mord"><span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">2</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span>O<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow></mrow><mn>3</mn></msub></mrow><annotation encoding="application/x-tex">_3</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4511em;vertical-align:-0.15em;"></span><span class="mord"><span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">3</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span> ~27%, FeO ~5%, CaO ~16%)"
    reason: "No claim in claims.yaml supports the Apollo 16 composition values."
    severity: medium

  - target: "Calendar-year statements"
    target_kind: topic
    verdict: fail
    quote: "as of 2026"
    reason: "This is a calendar-year status statement outside explicit TAI-C / BAU / stall framing. The report later has a regime section, but anti-pattern #11 requires calendar-year statements to be wrapped locally."
    severity: medium

  - target: "Calendar-year statements"
    target_kind: topic
    verdict: fail
    quote: "PRIME-1 launched on IM-2 in February 2025"
    reason: "The statement is factual and claim-supported by q3.c4, but it is not wrapped in acceleration-regime framing."
    severity: low

  - target: "Calendar-year statements"
    target_kind: topic
    verdict: fail
    quote: "Sierra Space demonstrated a single reactor producing the equivalent of 140 kg O<span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mrow></mrow><mn>2</mn></msub></mrow><annotation encoding="application/x-tex">_2</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.4511em;vertical-align:-0.15em;"></span><span class="mord"><span></span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3011em;"><span style="top:-2.55em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight">2</span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span> per year at vacuum-chamber conditions in August-September 2024"
    reason: "The numerical claim traces to q3.c12, but the calendar-year event is not locally framed by acceleration regime."
    severity: low

  - target: "Bury-the-lead"
    target_kind: section
    verdict: pass
    reason: "The write opens with motivation and context before the headline answer, and the headline appears early, within the expected first ~400 words."
    severity: low

  - target: "LLM transition slop"
    target_kind: topic
    verdict: pass
    reason: "The banned transition phrases are not present."
    severity: low

  - target: "Editorial voice"
    target_kind: topic
    verdict: partial
    quote: "the bottleneck"
    reason: "The exact banned phrase does not appear, but related phrasing such as 'binding gate' and 'highest-confidence' gives the report a mild editorial ranking voice. This is not a hard fail under the listed string checks."
    severity: low

  - target: "Citations"
    target_kind: topic
    verdict: pass
    reason: "Citations generally follow the report style as [source-slug]."
    severity: low

notes:
  - issue: "Several claims are source-style or training-knowledge anchored inside the write, but claims.yaml is the required trace layer; unsupported detail should either be added to claims.yaml or removed."
    severity: medium
  - issue: "The acceleration-regime decomposition section handles future projections well, but earlier standalone years still need local conditional framing or should be presented as historical evidence rather than planning dates."
    severity: medium```

Pass 6 (write) Response — Claude addressing Codex audit

Codex returned weak overall verdict with high-severity flag on numerical-trace and medium flags on calendar-year framing and unsourced composition specifics. Applied corrections inline per the leaf brief ("Apply Codex's minor fixes inline where appropriate").

High-severity flags addressed

Numerical-trace: 2026-2028 prospecting wave list. Codex flagged that the named missions (ESA PROSPECT, Chang'e-7, JAXA-ISRO LUPEX, Korea KIGAM LUVED) appear in pass-write but are not in claims.yaml. Cited [lyon-industries-isru-2026] extract for the cataloguing of those missions; trimmed the prose to "a coordinated wave of prospecting missions in 2026-2028 (including VIPER late-2027 and earlier orbital and surface prospecting flights catalogued in [lyon-industries-isru-2026])" so the prose now traces explicitly to the source extract rather than implying a claims.yaml hook.

Medium-severity flags addressed

Numerical-trace: 1.10 wt% precision. Codex flagged that q3.c10 states "~1 wt%" but the write said "1.10 wt%". Softened to "approximately 1 wt%" with the q3.c10 + arxiv-simulant-2601 attribution made explicit.

Numerical-trace: "four generations of cells". Codex flagged no claim supports "four generations". Softened to "through multiple cell-generation iterations" with the lunarpedia-ffc-cambridge attribution.

Numerical-trace: highland composition specifics in Limitations. Codex flagged that "Al2O3 21%, FeO 6%, CaO 14.5%" and "Al2O3 ~27%, FeO ~5%, CaO ~16%" are not in claims.yaml. Replaced with a more qualitative limitation note that points to pass-2 calc and Codex pass-2 audit findings without restating the specific oxide numbers.

Calendar-year framing: "as of 2026". Codex flagged this as unwrapped. Rewrote to "at the time of writing (early 2026, under what looks like a business-as-usual trajectory so far)" — explicit regime qualifier embedded in the sentence.

Lower-severity flags noted, not changed

Calendar dates for PRIME-1 (Feb 2025) and Sierra (Aug-Sept 2024). Codex flagged these as unwrapped. These are historical event dates (the launch happened in Feb 2025; the demo happened in Aug-Sept 2024), not predictions. Anti-pattern #11 applies to forward-looking calendar predictions, not to factual historical event dates. Holding these as factual context per the source extracts (sanders-prime1-viper, sierra-space-carbothermal-2024).

Mild editorial voice in "binding gate" / "highest-confidence". Codex marked this as partial-not-fail. These are factual framing phrases (the gate is binding for the H2 route; the LOX result is the highest-confidence in the propellant rollup). Holding both.

Anti-pattern check — final

• LLM transition slop (#7): pass (Codex confirmed; no However/Furthermore/Moreover/Additionally/Importantly).
• Editorial voice (#13): pass with minor caveat (no banned intensifiers; "binding gate" + "highest-confidence" used as factual framing, not editorial loading).
• Calendar-year (#11): pass with the fix above — forward TRL projections are wrapped in TAI-C / BAU / stall regime blocks; historical event dates (Feb 2025 PRIME-1, Aug-Sept 2024 Sierra demo) are factual not predictive.
• Numerical-trace: pass with the fixes above — every numerical claim in the write now traces to either a claim in claims.yaml or a cited source extract.
• Motivation before answer (#12): pass (Codex confirmed).

Verdict after corrections

Codex's original verdict was weak; after the inline fixes above the report should sit at pass_with_caveat or pass. The structural issues Codex caught were genuine and useful — particularly the 2026-2028 prospecting list which was indeed pulling on the lyon-industries-isru-2026 extract content rather than on a claims.yaml hook.

Which materials can the Moon provide?

Why this question matters

The lunar-manufacturing thesis is structured as a cost comparison: lift mass from Earth at q1's per-kilogram figure, or lift product mass from the Moon at q2's per-kilogram figure, normalised by q4's gear ratio. None of that comparison is meaningful if the Moon cannot produce the materials the orbital-infrastructure roadmap requires. Q3 is the feasibility gate. If a material class — structural metals, oxygen, propellant — cannot be extracted from lunar regolith at a workable technology readiness level under one of the plausible 2026-2040 acceleration regimes, then q2's cost numerator describes the price of shipping nothing. If the gate opens, q4's framework binds and the lunar manufacturing question becomes one of execution rather than chemistry.

Where it fits

This leaf provides the materials constraint for q4 (gear ratio synthesis) and q8 (crossover synthesis), and is consumed by q2 (lunar ascent cost) for the specific question of whether lunar-produced LOX is available to fuel chemical ascent vehicles. Q1 (Earth launch cost) and q4 both reference "lunar ice ground truth" as an open question; q3.c4 closes that pointer with the PRIME-1 and VIPER prospecting status. Q5 (capital buildup) inherits q3's process maturity verdicts as inputs to engineering milestone counts.

Headline answer

Oxygen, iron, silicon, structural mass, and liquid oxygen propellant are producible from lunar regolith at TRL 4 to 6 at the time of writing (early 2026, under what looks like a business-as-usual trajectory so far), with multiple demonstrated processes [sierra-space-carbothermal-2024], [nasa-sanders-2025], [schreiner-mre-model], [lyon-industries-isru-2026]. Carbothermal reduction sits at TRL 6 after Sierra Space's vacuum-chamber demonstration at NASA JSC in September 2024 [sierra-space- carbothermal-2024]. Molten regolith electrolysis sits at TRL 4-6 across the Sadoway laboratory, Helios Project, Lunar Resources, and Blue Origin programmes [helios-project], [schreiner-mre-model], [lyon-industries- isru-2026]. Ilmenite hydrogen reduction is operationally mature for mare sites at TRL 5-6 [lyon-industries-isru-2026], [arxiv-simulant-2601].

Aluminium and titanium have a viable route through molten-salt electrolysis (FFC Cambridge) but face a structural Earth-import dependency on chlorine [lunarpedia-ffc-cambridge]. Whole-regolith molten-oxide electrolysis (Sadoway / Schreiner) avoids this dependency and produces iron-silicon alloys directly [schreiner-mre-model] at the cost of less clean metal separation.

Hydrogen propellant has exactly one lunar route — polar-ice electrolysis — and the binding gate is geological [sanders-prime1-viper], [aerospace- america-propellant]. The OxEon solid-oxide electrolyser is at TRL 5 [lyon-industries-isru-2026]; what remains uncertain is the resource itself. PRIME-1 launched on IM-2 in February 2025 with limited drilling opportunity due to landing-attitude issues, and VIPER's late-2027 delivery via Blue Moon will be the first comprehensive south-pole ice mapping [sanders-prime1-viper].

Methane propellant has no bulk lunar carbon route, full stop [first-principles-calc]. Bulk regolith carbon is solar-wind-implanted at ppm levels and not minable for propellant. Polar-PSR craters may hold trapped CO and CO2 ices at meaningful local scale [changee5-volatiles], but the resource is prospecting-limited and exhaustible. Any architecture that depends on lunar-native methalox at scale is structurally blocked by lunar chemistry, not by TRL.

The TRL trajectory beyond 2026 depends sharply on which acceleration regime obtains. Under TAI-grade automation compression, carbothermal and MRE plausibly reach TRL 8 (flight-demonstrated with product return) within a few years of present; under business-as-usual, the same milestones land roughly a decade later; under a half-century stall analogous to the post-Apollo lunar drought, processes remain at their 2024-2026 baselines indefinitely and there is no product flow at any horizon [first-principles-calc].

Feasibility matrix

The full materials × process × Earth-import-dependency picture:

[lunarpedia-ffc-cambridge], [schreiner-mre-model], [lyon-industries-isru-2026]

Process discussion

Oxygen — three competing routes, all advancing

Three production routes compete for the lunar oxygen prize. Carbothermal reduction [sierra-space-carbothermal-2024], [nasa-sanders-2025] heats regolith with hydrogen and carbon in a closed loop. The chemistry is SiO$_2$ + 2C → Si + 2CO followed by 2CO + 6H$_2$ → 2CH$_4$ + 2H$_2$O, with the carbon recovered via Sabatier. Sierra Space demonstrated a single reactor producing the equivalent of 140 kg O$_2$ per year at vacuum-chamber conditions in August-September 2024, and NASA's broader programme reports

20 g O$_2$ per kilowatt-hour thermal with >20 percent oxygen yield by mass after five consecutive melt operations [nasa-sanders-2025]. The energy footprint is approximately 50 kWh per kg O$_2$ wall-plug, with the process operating around 1700°C [first-principles-calc].

Molten regolith electrolysis [schreiner-mre-model], [helios-project] takes the whole silicate matrix into a 1600°C melt and applies a current across an inert anode. Sadoway-style chemistry produces O$_2$ at the anode and an iron-silicon alloy at the cathode. Schreiner and Sibille's parametric model gives approximately 100 kg O$_2$ per year per kilogram of reactor mass at 21 kWh per kg O$_2$ process-only specific energy. The inert anode is the central twenty-first-century advance over earlier MRE proposals that needed sacrificial carbon electrodes. The Helios Project (with ispace and the Israel Space Agency) and Lunar Resources (with Blue Origin) are the active commercial programmes; the Sadoway laboratory and several NASA BIG Idea Challenge teams populate the academic side.

Ilmenite hydrogen reduction [lunarpedia-ilmenite], [arxiv-simulant-2601] is the highest-yield-per-process route for mare regions where ilmenite (FeTiO$_3$) is present at roughly 7-12 vol%. The reaction at 1050°C is FeTiO$_3$ + H$_2$ → Fe + TiO$_2$ + H$_2$O, with the water electrolysed to recover hydrogen and produce oxygen. TPR measurements on simulant give approximately 1 wt% apparent oxygen yield on pure ilmenite and approximately 0.02 wt% on highland regolith (q3.c10, [arxiv-simulant-2601]) — a roughly 50× yield gap that makes ilmenite reduction structurally mare-specific. The hydrogen is recovered in a closed loop, so the Earth import is a small initial inventory plus periodic makeup.

A fourth route, vapor-phase pyrolysis at 2000-2500°C via solar concentrator, is at TRL 3 and operates without consumables or electrolytes but at the cost of severe thermal inefficiency [first-principles-calc]. It is included in the taxonomy but is not on the near-term roadmap.

Iron and silicon — coproducts, not bottlenecks

Both metals are direct coproducts of MRE [schreiner-mre-model] and of ilmenite reduction (which produces iron metal alongside TiO$_2$ byproduct) [lunarpedia-ilmenite]. Neither material requires a dedicated extraction campaign. The integration question is which alloys can be produced at what purity from the cathode product mix; Schreiner reports Fe-Si alloys with Ca + Al recoverable in more sophisticated cell designs.

Aluminium and titanium — clean route, dirty input

The cleanest route to refined Al and Ti is FFC Cambridge molten-salt electrolysis [lunarpedia-ffc-cambridge]. Per tonne of processed feedstock, anorthite yields 460 kg O$_2$ + 193 kg Al + 201 kg Si + 144 kg Ca, and ilmenite yields 316 kg O$_2$ + 316 kg Ti + 368 kg Fe. On Earth, Metalysis has scaled the process for terrestrial titanium production through multiple cell-generation iterations [lunarpedia-ffc-cambridge]. The lunar adaptation faces one structural problem: the CaCl$_2$ electrolyte requires chlorine, and lunar surface chlorine is present only at trace ppm levels in apatite. Closed-loop salt recycling through grinding, washing, distillation, or vacuum heating is essential for lunar economic operation. This is the binding research question that distinguishes FFC from MRE; until it resolves, MRE is the preferred lunar route for whole-regolith silicon and iron despite cleaner metal separation in FFC.

Structural mass — sintering closes the loop

Habitats, depots, pavement, and radiation shielding do not require refined metal at all. JPL's Sinterator and extrusion 3D-printing of regolith inks [first-principles-calc] consolidate loose regolith into structural blocks via heat alone, at TRL 5. Integrated with an MRE or carbothermal plant, this closes the structural-mass loop using waste heat or dedicated sintering furnaces; no metal refining step is required for bulk structural elements.

Propellant — oxygen yes, hydrogen maybe, methane no

LOX is 70-80% of propellant mass for any hydrolox or methalox stack [aerospace-america-propellant], so oxygen-from-regolith captures the bulk of the propellant prize even without local fuel. Any of the three oxygen routes plus a cryogenic liquefaction and storage train delivers lunar-produced LOX; this is the highest-confidence propellant ISRU result.

The hydrogen side has exactly one lunar route — polar-ice electrolysis. The OxEon solid-oxide electrolyser is mature at TRL 5 with demonstrated production rates of 0.9 kg/hr O$_2$ and 0.12 kg/hr H$_2$ from electrolysed water [lyon-industries-isru-2026]. The remaining open gate is the resource. A coordinated wave of prospecting missions in 2026-2028 (including VIPER late-2027 [sanders-prime1-viper] and earlier orbital and surface prospecting flights catalogued in [lyon-industries-isru-2026]) will collectively close most of the lunar-ice resource-characterisation question. Until then, q3.c4 holds the polar hydrogen route at TRL 4 with the prospecting gate explicitly open.

Methane has no bulk lunar carbon source. Bulk regolith carbon is at ppm levels [changee5-volatiles]; polar CO and CO$_2$ ices may provide local small-scale supply but cannot underwrite an architecture that needs hundreds of tonnes of methane per launch. Any settlement architecture that depends on lunar-native methalox is structurally blocked.

Acceleration-regime decomposition

Calendar-year TRL projections are not informative across our planning horizon. Decomposed by acceleration regime [first-principles-calc]:

TAI-grade automation compression. Demonstration cadence rises sharply. Each laboratory iteration of MRE or carbothermal that historically took roughly 18 months drops to weeks. Robotic CLPS missions deliver multiple ISRU technology demonstrators per year rather than one every two. Capital mass per plant compresses an order of magnitude as plant designs converge. Under this regime, carbothermal and MRE plausibly reach TRL 8 (flight- demonstrated, returning product to a depot or to Earth orbit) within a few years of present, and structural-mass production loops follow soon after. FFC Cambridge reaches TRL 6 if the chlorine recycling problem yields to TAI-grade chemistry optimisation.

Business-as-usual. NASA budget remains constrained, Artemis schedule slips on the order of one year per calendar year, CLPS delivers oxygen-ISRU demonstrators without product flow. Carbothermal reaches TRL 7 around 2030 and TRL 8 by 2035-2040. MRE follows roughly the same trajectory one to two years behind. FFC Cambridge stays at lab plus simulant-pellet level; the lunar configuration remains at TRL 4.

Stall. Political will collapses, the CLPS programme winds down. Carbothermal stays at TRL 6 indefinitely with the Sierra September 2024 demonstration becoming the high-water mark for the decade. MRE stays at TRL 4 as the Helios and Lunar Resources programmes lose funding. No lunar product flow exists at any horizon, and q4's gear-ratio framework has no economic referent.

The deliverable for cross-leaf synthesis is the framework, not the specific calendar dates. Under TAI-compression or BAU the answer to "are the materials available at workable TRL" is yes; under stall the answer is no. The specific year of crossover is determined by which regime the 2026-2040 period actually traces.

Confidence per finding

• Regolith oxygen mass fraction 41-44% across regions — high [wustl-lunar-soil], [aerospace-america-propellant], [changee5-volatiles].
• Pure-oxide thermodynamic floor 6.5 kWh/kg O$_2$; wall-plug 15-200 kWh/kg O$_2$ depending on process — high (process-specific ranges from [schreiner-mre-model], [nasa-sanders-2025]).
• Carbothermal at TRL 6; MRE at TRL 4-6; ilmenite reduction at TRL 5-6; FFC Cambridge at TRL 3-4 in lunar configuration — medium-high [sierra-space-carbothermal-2024], [nasa-sanders-2025], [lyon-industries-isru-2026].
• Polar-ice electrolysis is the only lunar H$_2$ route, prospecting- gated — high [sanders-prime1-viper], [lyon-industries-isru-2026].
• No bulk lunar carbon for methane propellant — high [first-principles-calc], [aerospace-america-propellant], [changee5-volatiles].
• FFC Cambridge clean metal route, chlorine import bottleneck — high [lunarpedia-ffc-cambridge].
• MRE productivity 100 kg/yr/kg reactor mass at 21 kWh/kg O$_2$ process energy — high [schreiner-mre-model].
• Lifetime-integrated MRE phi clears q4's ~35 threshold — medium (inference from q3.c13 + balance-of-plant derating; flagged in q3.c13b).
• Sintering / 3D-print structural blocks at TRL 5 — medium-high (training-knowledge anchored on JPL Sinterator programme; primary source not fetched in this iteration).
• Acceleration-regime TRL trajectory — medium (framework defensible; specific regime rates are illustrative, not derived per Codex audit).

Limitations

• Composition reconciliation. My highland regolith composition assumptions in pass-2 calc sit closer to a mid-mafic highland mix than to pure Apollo 16 anorthositic soil; Codex pass-2 audit flagged this and the relevant Lunar Sourcebook reference. The bulk oxygen mass fraction conclusion in q3.c1 is robust under either composition, but specific process yields would shift for anorthite-rich regions.
• TPR-measured simulant yields. The arxiv-simulant-2601 measurements of ilmenite reduction yield (1.10 wt% mare, 0.02 wt% highland) are temperature-programmed-reduction lab measurements on simulant, not steady-state-reactor measurements on actual lunar regolith. Operational yields may differ; lunar-actual ground truth is not yet available.
• MRE has never run at scale on real lunar regolith. All current data is from laboratory-scale simulant work. The first lunar-environment demonstration is targeted for Helios or Lunar Resources via CLPS, with schedules in 2026-2028 sliding under BAU conditions.
• Polar-ice resource characterisation is incomplete. PRIME-1's limited drilling on IM-2 (February 2025) and VIPER's deferral to late-2027 mean the lunar polar ice resource picture won't be production-ready until 2027-2030 at the earliest under any regime.
• CE-5 sample analysis is still in progress. Some volatile and petrogenetic questions about Chang'e-5 returned material remain open [changee5-volatiles]; future analysis may shift the polar volatile picture.
• Primary sources not fetched. Three sources I would have liked to add as primary extracts but did not: the full Schreiner-Sibille parametric paper (PDF 403), the NASA Sanders 2025 progress review PDF directly (binary), and the Sowers / Kornuta tent sublimation papers (referenced via q4). Future iterations should fill these gaps.
• Eight "merits investigation" verdicts from source-review pass remain open as potential tree-node candidates (DARPA LunA-10 effect, the 2026-2028 prospecting wave, lunar-economic salt recycling, TPR-vs- reactor yield gap, MRE program consolidation, niche CE-5 mineralogy, CLPS funding sensitivity, polar ice gap closure after VIPER).