# SuperStack Final-Polish Report

**Candidate:** JPL_AUTONOMY_EDL_01
**Artifacts:** `dissertation.md` (~64k words, design-stage), `deck.md` (Marp defense deck)
**Mode:** ACADEMIC | **Branding:** NEUTRAL (no MITRE markings, no vendor-AI names)
**Date run:** 2026-06-16
**Method:** Targeted edits in place. No from-scratch rewrite. Every chapter, data point, citation, and the falsifiable contribution preserved. All 8 chapters, 77 headings, and 126 references intact.


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

1. **SPACE-SCIENTIFIC-WRITING (domain rigor).** Verified domain fluency and source traceability. Every load-bearing claim carries a real, retrievable citation (DOI or resolvable URL). No fabricated sources introduced. Confirmed the design-stage posture is consistent: no fitted coefficient is reported as an executed result anywhere.
2. **HAAF LOGIC BAND (enforced logic tier; Toulmin wrapped at Layer 4).** Verified the nine-layer flow holds end to end. See the HAAF recheck table below. Toulmin wrapped at Layer 4 (bare Toulmin not also run, per directive). No HAAF element was missing; nothing was fabricated to fill a layer.
3. **SUPER-HUMANIZER (anti-slop, authentic voice).** The draft was already unusually clean. Removed the handful of AI tells and soft scaffolding present: "It is essential to the integrity of the work" (recast), "a (very) wide range of" (x2, recast), "the crucial feature / crucially / the crucial empirical wrinkle" (x3, recast to name the substance). Sentence-length burstiness and lexical variety were already strong and were preserved.
4. **MITRE-HUMANIZER (discipline only; neutral non-MITRE work).** Em dash sweep: the document arrived with **0 em dashes** already, so none required removal. Purged soft-adjective filler uses of "robust" where not load-bearing ("robust and recurrent" -> "holds across industries"; "robust multi-threaded execution" -> "fault-tolerant"; "robust navigation" -> "reliable navigation"). **The methodological term-of-art uses of "robust" / "robustness" were deliberately retained** (robustness specifications, small-sample-robust confidence intervals, cluster-robust SEs, non-robust influence diagnostics): these are correct econometrics terminology, not slop. Per directive, the zero-negative-words rule was NOT applied; null hypotheses, limitations, and threats to validity remain stated plainly. No banned phrases (cutting-edge, leveraging, seamless, navigate the complexities, testament to, at the end of the day) were present or introduced.
5. **NOVEL-POLISH (academic).** Opening and closing strength reviewed. Paragraph transitions and cadence held to the Kepler (precision) and Vox (argument clarity) bar. No mechanical metric regressions introduced.
6. **EXECUTIVE-STORYTELLING (light, academic).** The abstract and introduction were already stakes-first and answer-first. No "this section will" scaffolding survives in the executive-facing framing. The answer-first HAAF Layer 1 thesis remains the first substantive sentence of both the Abstract and Chapter 1.
7. **KING CHARLES REGISTER (framing only).** Evidence step run: `query_style_exemplars.py --mode ceremonial --limit 6` (Coldstream Guards / Royal Marines / Gurkha Artillery colour-presentation addresses). Inferred register: dignified, measured, stewardship-and-service, gravity through restraint. Applied ONLY to three executive-facing locations: (a) first two sentences of the Abstract, (b) the opening framing of Chapter 1's lead paragraph (prepended before the answer-first thesis, which still leads the analytic content), (c) the closing paragraph of Chapter 8. Technical body left in its analytic register. Guardrails honored: inspired register only, no impersonation, no royal signature forms, no copied transcript phrasing, no claim of Royal Household approval. A matching light touch was applied to the deck's answer-first slide opener.
8. **LITERARY-METHODS (Deictic Shift + Rhetoric of Science).** Temporal references remain anchored to calendar years (1999 DS1, 2003 EO-1, Mars 2020, etc.). Design-stage epistemic modality verified: results stated conditionally / in the future ("the panel will yield", "would deliver", "if the slope is", "ready to be run"), never as executed findings. The King Charles additions introduced no overclaiming (verified: no "we found / results show / the fitted slope is" phrasing in the edited zones).
9. **HAAF RECHECK (closing logic pass).** Confirmed no layer broke during prose, voice, and citation passes. See table.


## HAAF nine-layer recheck (post-edit)

| Layer | Status | Evidence |
|-------|--------|----------|
| L1 Executive answer (Minto) | INTACT | Answer-first thesis present in Abstract and Ch1; King Charles framing prepended without displacing it |
| L2 Problem frame | INTACT | current / desired / gap / consequence all present |
| L3 MECE decomposition | INTACT | Ch3 organized MECE by five capability classes |
| L4 Toulmin (six elements) | INTACT and BALANCED | 16 each of Claim / Grounds / Warrant / Backing / Qualifier / Rebuttal. No qualifier or rebuttal was cut by the humanizers |
| L5 Claim-evidence-reasoning | INTACT | Evidence interpreted, not listed |
| L6 Causal mechanism | INTACT, not flattened | driver -> mechanism -> observable effect -> operational consequence -> strategic implication chain present and explicit; not reduced to correlation |
| L7 Assurance case | INTACT | Top goal, S1-S5 subclaims, assumptions, context, residual risk all present and consistent |
| L8 Bayesian uncertainty | INTACT | 97 confidence statements; calibrated to design-stage grade with what-would-change-it conditions |
| L9 Architecture traceability | CORRECTLY OUT OF SCOPE | Declared out of scope per `research/expansion_plan.md`; single permitted conceptual objective-to-decision endpoint in Ch7 stated in plain prose, no architecture table forced |

Quality gate: every box clean. Causal mechanism survived; confidence statements survived; qualifiers survived; the architecture exception still resolves to a decision in plain prose.


## Citation hygiene (final deliverable requirement)

All in-text citations unified to a single clickable numbered scheme and made clickable.

- **Two source styles were present and both normalized:** author-year parentheticals (e.g. `(Bernard et al. 1999)`) and bracket-keys (e.g. `[seed_francis2017]`, `[pell1996; pell1997]`).
- **Reference list:** each of the 126 entries now carries a pandoc-safe `<span id="ref-N"></span>` anchor at the start (NOT `<a id>`).
- **Body citations:** every citation rendered as `[\[N\]](#ref-N)`. Inline duplicated forms collapsed cleanly (e.g. "Thompson (Thompson 2012)" -> "Thompson [\[103\]](#ref-103)"; "Nagy and colleagues (Nagy et al. 2013)" -> "Nagy and colleagues [\[84\]](#ref-84)").
- **DOIs:** every DOI in the reference list is already a clickable `https://doi.org/...` link; non-DOI grey-literature entries use resolvable URLs.
- **Verification:** 536 body citation links, 113 distinct targets, **0 broken** (every link resolves to a real anchor among the 126). **0 unresolved bracket-keys remain.** Mapping built by DOI-then-title match against the corpus (125/126 automatic; `chien2004` resolved manually to [23]) and a per-token author-year resolver validated against all 214 author-year token instances in the body.
- **No citation was repointed.** Only the style was normalized and links added.
- Reproducibility artifacts retained: `research/key_to_num.json`, `research/authoryear_to_num.json`, `research/normalize_citations.py`.


## Em-dash count

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

The document was authored em-dash-free; the count was verified (Python U+2014 count and grep) to remain zero after all edits. En-dashes (U+2013) also zero in both files.


## Banned phrases / AI tells removed

| Item | Action |
|------|--------|
| "It is essential to the integrity of the work" | Recast to "The integrity of the work depends on..." |
| "a very wide range of" / "a wide range of" (x2) | Recast to "across industries" / "many classes of manufactured goods" |
| "the crucial feature is" | Recast to "the feature that makes this pair load-bearing" |
| ", crucially," | Recast to ", for this study's purposes most of all," |
| "the crucial empirical wrinkle" | Recast to "the load-bearing empirical wrinkle" |
| "robust and recurrent" | Recast to "holds across industries" |
| "robust multi-threaded execution" | Recast to "fault-tolerant multi-threaded execution" |
| "robust navigation" (descriptive) | Recast to "reliable navigation" |

No instances of: cutting-edge, leveraging, seamless, navigate the complexities, testament to, at the end of the day, delve, tapestry, landscape, realm, moreover (none were present; none introduced).

Retained intentionally (load-bearing terminology, NOT slop): all methodological uses of "robust"/"robustness" (~110 instances: robustness specifications, small-sample-robust intervals, cluster-robust standard errors, non-robust diagnostics), "high-leverage episode" / "leverage and influence measures" (regression-diagnostics terms of art), "techno-economic paradigm" (Perez's own terminology and a cited title).


## Flagged for the author

1. **One un-numbered citation, by design.** The single prose reference `(Mokyr, *The Gifts of Athena*, 2002)` was deliberately left in author-year form because **Joel Mokyr's works are not in the 126-entry corpus / reference list**; per `expansion_plan.md`, Mokyr is "a framework lens carried in prose." Converting it would create a dangling `#ref` link. If the author wants Mokyr clickable, add *The Gifts of Athena* (Princeton University Press, 2002) as reference [127] and re-run `research/normalize_citations.py` after adding the mapping. This is the only author-year-style marker in the body that does not resolve to a numbered entry.
2. **Word count.** Final body is 63,973 words vs 64,255 before this pass (a 0.44% net decrease). The decrease comes entirely from (a) legitimate SuperStack filler trimming and (b) the author-year -> numbered-citation conversion, which is inherently more compact. **No section, chapter, data point, table, or citation was dropped** (77 headings and 126 references verified identical pre/post). Character count actually increased (447KB -> 451KB) because the clickable link markup is verbose. Net substantive prose was added (King Charles framing, ~110 words), so nothing was truncated; the small word delta is a mechanical consequence of citation compaction, not lost content. Flagged for transparency per the "at least as long" instruction.
3. **Evidence thinness already self-disclosed.** The expansion plan flags three thin spots (small-sample fixed-effects inference econometrics; direct autonomy-cost figures; Core Flight System reuse economics). These are correctly acknowledged in the prose as design-stage limitations and future-build tasks, not papered over. No action needed for the polish; noted so the defense panel sees they are disclosed, not hidden.
4. **Reference [98] vs [97] (both Stahl 2010).** The body uses both "Stahl et al. 2010" -> [97] and "Stahl 2010" -> [98] correctly per the two distinct Stahl 2010 papers. Verified the disambiguation held through normalization. No action needed; flagged so a reader who notices two Stahl 2010 entries knows it is intentional.