# Research amendments and operational deviations

## 1. Retrieval representation

**Recorded:** 14 November 2026, during corpus collection.  
**Original plan:** [Research plan, version 1](research_plan.md), preserved unchanged.

Direct downloads were unavailable from the local terminal. Complete numbered text representations of accessible Library of Congress pages were therefore collected through a web reader. Exact reader responses are archived in `.reader.json` snapshots. They are not original HTML bytes, and no claim of byte-level preservation of the published HTML is made.

This changes the retrieval representation, not the research question, selected essays, edition, features, counting rules, comparison method, validation, or robustness checks. It introduces a limitation: source markup and formatting cannot be checked against original HTML in these snapshots. All selected texts still come from the same Library of Congress presentation. Reader omissions and conflicting overlaps must be resolved or rejected before a text enters the collected sample.

At this decision, only source labels, textual completeness, and extraction integrity had been examined. No essay frequencies, validation scores, distances, or authorship predictions had been calculated.

The page containing Nos. 21 and 26 could not be retrieved completely. Those essays remain missing rather than being replaced. See the [collection record](corpus_collection.md) for the available material, archived evidence, and remaining work.

## 2. Main prose, appended notes, and readable copies

**Recorded:** 14 November 2026, before preparing analysis copies or calculating essay frequencies.

**Affected decisions:** Boundaries of the essay body and the stage at which counting normalization is applied.

The original plan requires the essay body and exclusion of surrounding editorial material, but does not fully specify how to handle separately numbered source footnotes. For the main analysis, the prepared text will contain the running prose after the shared reader address and before the standalone closing signature, if present. Numbered notes appended after that signature will be excluded, along with their inline numeric reference markers. Numbered arguments within the prose will remain. An unfamiliar boundary or markup pattern will require review rather than automatic deletion.

This is a clarification of the textual unit, not a claim that all appended notes are editorial. It keeps the main argument comparable across essays and prevents reader navigation from being counted as language. It also excludes potentially authorial material, which is a limitation to state in the report. All excluded notes remain available in the raw files, and each removal is logged exactly. Quotations within the retained prose remain for the main analysis.

Readable files in `data/clean/` will preserve case and apostrophes for close reading and stable quotation offsets. The already fixed lowercasing and apostrophe normalization will be applied through a shared function at counting time. No counting rule, feature, sample, scaling, validation, or decision threshold changes.

Only source structure and preparation integrity had been examined when these decisions were recorded. No essay frequencies, validation scores, authorship distances, or predictions had been calculated. The original plan remains unchanged. Preparation currently covers the 15 collected essays; the absence of Nos. 21 and 26 is not a change to the fixed sample.

## 3. Receipt of the two missing reference texts

**Recorded:** 15 November 2026, before full-sample validation or comparison of No. 49.

The user supplied plain-text copies of Nos. 21 and 26 in response to the request for the inaccessible Library of Congress page. These payloads are preserved in `data/raw/supplied/` and fingerprinted in `metadata/supplied_sources.json`. Their Hamilton labels are checked against the previously archived Library of Congress authorship table. Their association with the requested edition is inferred from the context of delivery; the full wording has not been independently compared with the live page. This transcription limitation must accompany the results.

No wording was modernized or corrected. No. 21 retains its unusual break before the closing punctuation after "four." No. 26 retains forms such as "wellinformed" and "calamaties." The normal preparation and counting rules apply unchanged. Original case, quotations, and spellings remain in the readable copies; the fixed tokenizer treats line breaks as separators.

Before importing the two texts, the partial source and preparation logs, frequency tables, manifests, blocked-status records, collection code, and dated documentation were copied byte for byte to `results/history/partial_corpus_14_november_2026/`. Its archive manifest records every copied checksum. The active records will describe all 17 essays. The previously collected 15 raw texts and prepared copies remain unchanged.

The research question, selected essay numbers, 24 words, distance formula, validation procedure, and decision threshold are unchanged. No full-reference validation accuracy or questioned-text distance had been examined when this operational addition was recorded.

## 5. Operational quotation boundaries for the predefined check

**Recorded:** 15 November 2026, after the main comparison and before either robustness variant was evaluated.

The quotation check uses the preserved readable copies and the decisions in `metadata/quotations.csv`, with input checksums and a complete 17-essay review in `metadata/quotation_review.json`. Offsets are Unicode character positions from zero, with exclusive ends. Each removed span includes its enclosing double quotation marks and internal paragraph breaks; one space replaces it. Reporting words outside the marks remain. An embedded reporting phrase inside an outer quotation is removed with that passage. No overlapping or nested double-quote spans occur in these prepared copies.

Remove clearly delimited passages presented as borrowed speech, documentary extracts, or established sayings. The 9 marked constitutional-power passages in No. 43 are treated consistently as presented documentary extracts, including shortened formulations. This is a check of the essay's presentation of quotations, not authentication of every extract against its antecedent document; some wording may be authorial adaptation.

Retain titles, explicit syntheses of opponents' arguments, generic adversary speech whose direct source is uncertain, explanatory paraphrases, and unmarked reported ideas. Record ambiguous passages rather than silently treating every quotation mark as borrowed language. The review identified 33 marked spans: 27 for removal and 6 for retention. Six essays have no marked spans. This conservative check cannot remove undetected or unmarked borrowing.

The main validation and No. 49 results were already known when annotation began. These decisions were fixed before either variant's scores were calculated. Annotation was therefore not blinded. The original features, corpus, counting rule, distance formula, tie rule, and validation procedure remain unchanged. The two checks are independent variants, not a combined change.
