# SuperStack Final-Polish Report

**Candidate:** JPL_ASTRO_EARTH_09
**Documents:** `dissertation.md` (revised in place), `deck.md` (revised in place)
**Mode:** ACADEMIC. **Branding:** NEUTRAL (no MITRE markings, no vendor-AI names).
**Date:** 2026-06-16
**Method:** targeted edits only; no from-scratch rewrite. Every chapter, data point, citation, and the falsifiable H0/H1 contribution preserved.


## Passes run, in canonical SuperStack order (stack_order.md)

1. **SPACE-SCIENTIFIC-WRITING (domain rigor).** Verified domain fluency and source traceability. All cited sources are real, drawn from the candidate's own `research/corpus.jsonl` (148 entries), each with a resolvable DOI or stable URL. No source added, removed, or re-pointed.
2. **HAAF LOGIC BAND (enforced logic band; Toulmin wrapped at Layer 4).** Rechecked the nine-layer flow across all eight chapters. Confirmed intact: answer-first executive theses (14 explicit "the chapter's answer / stated first" markers); problem frames (current/desired/gap/consequence) present in every chapter opening; full Toulmin per major claim (warrant 67, qualifier 55, rebuttal 35, grounds 58, backing 28 occurrences); claim-evidence-reasoning prose, not evidence dumps; named causal mechanisms (181 "mechanism" occurrences, driver-to-effect-to-consequence form preserved); confidence statements calibrated to the design-stage grade (61 occurrences, all "conditional/moderate" capped); the Assurance Case Summary (top goal G0, five subclaims, assumptions A1-A5, context, residual risk) present and consistent. Layer 9 (architecture traceability) correctly remains out of scope per the document's own Section 2.4.4 and `research/expansion_plan.md`; it was not forced. No HAAF element was missing; none was fabricated.
3. **SUPER-HUMANIZER.** Document was already largely free of AI scaffolding. Removed the one residual generic AI-tell word ("landscape," 1 instance) and an awkward inline citation annotation ("indexed in the corpus as the Copernicus uptake study"). No "delve / tapestry / realm / moreover / it is important to note / in conclusion" instances were present. Sentence-length variation and authentic academic voice retained.
4. **MITRE-HUMANIZER (discipline only; neutral non-MITRE work).** Em-dash sweep: zero em dashes before, zero after (a prior pass had already removed them; re-verified by grep on the final file). Banned-phrase purge: no instances of robust-as-marketing, cutting-edge, leveraging, seamless, navigate the complexities, testament to, at the end of the day. The 121 occurrences of "robust/robustness" are domain-legitimate econometric terms (robustness battery, heterogeneity-robust estimator, Rambachan-Roth robustness region) and were correctly retained. The zero-negative-words rule was NOT applied: null hypotheses, limitations, and threats to validity are stated plainly, as academic prose requires.
5. **NOVEL-POLISH (academic).** Strengthened the opening of the Abstract and Introduction and the close of the Conclusion (see King Charles, below). Held prose to the precision (Kepler) and argument-clarity (Vox) bar. No metric regressions introduced.
6. **EXECUTIVE-STORYTELLING (light, academic).** Made the Abstract and Introduction openings stakes-first (the public-trust stewardship stake now leads, before the methodological thesis). No process-label scaffolding ("this section will") introduced; payoff arc preserved.
7. **KING CHARLES REGISTER (framing only).** Ran the mandatory evidence step (`query_style_exemplars.py --mode ceremonial --limit 6`) and inferred the dignified, measured stewardship-and-service public-address register from six Royal address exemplars. Applied it ONLY to three executive-facing framing spots, leaving the technical body in its analytic register:
   - Abstract, first sentence: data "held in trust," terms of release as a matter of honoring that trust.
   - Introduction Section 1.0, opening framing sentence: a mission as "an instrument of public service," stewardship resting on measured benefit.
   - Conclusion Section 8.4, closing sentence: the reproducible method offered "in the same spirit of stewardship and service," left open for others to verify, contest, and build upon.
   - Deck slide 1 subtitle: a one-line stewardship framing.
   Guardrails honored: inspired register only, never impersonation; no royal signature forms; no copied transcript phrasing beyond the inferred register; no claim of Royal Household approval.
8. **LITERARY-METHODS (Deictic Shift + Rhetoric of Science).** Cold-reader audit confirmed: temporal references anchor to calendar dates (the 2008 Landsat switch, the early-1990s EOSDIS posture, the observation window are all dated in prose); entities are introduced before pronominal reference; epistemic modality stays in conditional/future terms ("the design would estimate," "expected sign," "illustrative and design-stage") and never asserts executed findings. Removed one ambiguous citation-as-annotation.
9. **HAAF RECHECK (closing logic pass).** Confirmed no layer broke after the prose, voice, and citation passes: every major claim retains its warrant and qualifier; no causal mechanism was flattened to a correlation by the humanizers (the distribution-before-publication ordering signature and the relabeling falsifier are intact); the calibrated confidence statements survived; the Layer 9 out-of-scope decision still resolves to the plain-language objective-to-decision link rather than a forced architecture table. Reference: `haaf_quality_gate.md`.


## Em-dash count

| File | Before | After |
|---|---|---|
| dissertation.md | 0 | 0 |
| deck.md | 0 | 0 |

A prior pass had already eliminated em dashes; this run re-verified zero via `grep` on the final files. No en-dashes or horizontal bars present either.


## Banned phrases removed

| Item | Count removed | Note |
|---|---|---|
| "landscape" (generic AI-tell) | 1 | replaced with "set of curated databases" |
| Awkward inline citation annotation ("indexed in the corpus as...") | 1 | reduced to the clean citation marker |
| em dashes | 0 | already zero on entry |
| robust / cutting-edge / leveraging / seamless / testament to / delve / tapestry / moreover | 0 | none present (the 121 "robust/robustness" hits are legitimate econometric terms, retained) |


## Citation hygiene (final deliverable requirement)

This was the largest mechanical task and revealed a genuine pre-existing inconsistency that was fully corrected.

**Problem found.** The body intermixed three incompatible citation schemes: (a) author-key tokens such as `[eysenbach2006]` (Chapters 3, 5, 7); (b) prospectus seed-numbers `[1]`-`[36]` (Chapters 1, 2, 6, 8), where the extension numbers 27-36 meant *different sources in different chapters*; and (c) a corrupted Chapter-4-local numbering that matched neither the prospectus crosswalk nor the final author-sorted reference list (e.g., body `[25]` intended North, but author-sorted entry 25 is an unrelated cannabis-policy study). None of the three resolved directly to the only complete, anchored list, the author-sorted 148-entry reference list.

**Resolution.** Unified every in-text citation to the author-sorted 1-148 scheme:
- All 101 distinct author-key tokens mapped to their reference number via DOI matching against `corpus.jsonl` (100% matched, saved in `research/_keymap.json`).
- Prospectus seeds `[1]`-`[26]` mapped via the document's own crosswalk; chapter-local extension numbers `[27]`-`[36]` and all Chapter-4-local numbers resolved per-occurrence by in-prose author context (saved in `research/_nummaps.json`).
- Rendered every body citation as a clickable `[\[N\]](#ref-N)` marker (688 citation links; multi-source groups deduplicated and sorted).
- Anchored all 148 reference entries with a pandoc-safe `<span id="ref-N"></span>` at the start of the entry.
- Made all 148 reference DOIs/URLs clickable `[url](url)` links.

**Verification.** Zero unresolved citation markers in the body (zero stray author-keys, zero bare numeric brackets). All 111 distinct in-text reference numbers resolve to a defined anchor; all 1-148 anchors defined; zero broken links; zero out-of-range targets. Sixteen high-stakes citations spot-verified to point to the correct source (North=85, Callaway and Sant'Anna=14, Rambachan and Roth=97, Goodman-Bacon=35, Wulder 2019=122, Wulder 2022=77, Landefeld=62, Henderson=41, McMillan-Rodrik=76, Wilkinson FAIR=120, Hemati=82, and the NASA grey-lit entries 135/138/145/147). No citation was re-pointed to a different source; only the style was normalized and links added.


## Flagged items for the author

These do not affect any claim, number, or the falsifiable contribution, but the author should confirm them because the original Chapter 4 numbering was corrupted and a few referents had to be inferred from prose context rather than read from a clean number:

1. **Chapter 4 NASA grey-literature group citations.** The original Chapter-4-local numbers for the EOSDIS / DAAC / data-policy grey literature did not match the final reference list and could only be resolved by context. They were mapped to the best-matching NASA Technical Reports Server entries: `[27]`->147 (NASA ESDS: Open Data), `[34]`->143 (Earth Science Data Stewardship), `[37]`->145 (Stewardship / long-term archives), `[47]`->144 (Role and Evolution of NASA ESDS), `[31]`->138 (US data policy for EO from space). Please confirm each grey-lit citation points to the intended NASA report; the candidate corpus contains 13 "Anonymous" NASA entries (135-147) and the original draft's intended subset was not recorded.
2. **Chapter 4 `[40]` (relabeling falsifier) -> 76 (McMillan and Rodrik).** Resolved to the relabeling authority the sentence is about, since the original number matched no clean source. Confirm this is the intended citation rather than a data-citation-norms source.
3. **Chapter 4 `[43]` (DAAC policy / mission-regime records) -> 74 (Rousi et al.).** Lowest-confidence single mapping; the original number was corrupted and the prose names a generic source class. Confirm or re-point.
4. **Chapter 4 `[105]` (cloud / programmatic access, time-varying) -> 29 (Consoli et al., cloud-processing framework).** A Sentinel cloud-format paper (entry 103) is an alternative; confirm.
5. **Uncited reference entries.** 37 of the 148 corpus entries are never cited in the final body. This is consistent with a corpus assembled broader than the narrative needs and is not an error, but the author may wish to prune the reference list to the cited set, or add the uncited sources where relevant, before final submission.


## Convergence

- Logic band (HAAF): quality gate clean; no failure-mode pattern surviving; Layer 9 correctly excluded.
- MITRE-humanizer: zero em dashes, zero banned marketing phrases, correct terminology.
- Citation hygiene: zero unresolved markers; 688 clickable in-text links; 148 anchored, clickable reference entries.
- Length preserved: 66,867 words (at or above the original; no section, chapter, data point, or citation dropped).
- HAAF recheck: no regression; warrants, qualifiers, mechanisms, and confidence statements intact.