# Session 63: What the punctuation corrections actually change

**The corrections do not rescue the earlier arithmetic searches.** Replaying the affected clause-reset tests changes some candidate scores, offsets and continuations. One section-8 candidate rises from fifth to second place, but neither affected section gets a new overall winner or a coherent reading. No new plaintext was verified.

This closes the specific dependency question raised in Session 62. The three observed separator-class corrections matter, but their numerical effects are confined to code that distinguishes those marks from ordinary word separators. They do not change rune values or word divisions.

## Evidence versus the compatibility hypothesis

The [fixed plan](boundary_impact_plan.json) uses the three overlays already recorded in the [source-aware boundary map](boundary_model.json). The original transcription remains untouched.

For this retrospective comparison only, the program builds an in-memory version in which each of those three literal hyphens becomes a semicolon. That is the older scripts' convention for a three-dot mark. It is **not** a finding that the mark means a grammatical semicolon or tells a cipher to reset. The unresolved split 3+1 and 1+3 pairs remain untouched.

| Correction | Location | Consequence for the reviewed experiments |
|---|---|---|
| Archive 15, gap 60 | Solved second koan, LP1 | Outside the older scoring-model training pages; Session 62 already tested its known repeating-key solution |
| Archive 55, gap 28 | LP2 page 38; section 7, rune gap 1149 | Changes clause-counter indices for 29 runes, until the next existing reset |
| Archive 63, gap 96 | LP2 page 46; section 8, rune gap 1689 | Changes clause-counter indices for 197 runes, crossing the page boundary and ending at the next existing reset |

Offsets are zero-based rune gaps. The 29 and 197 figures describe counter changes on the complete sections, not guaranteed changes in every decoded letter. Removing first clauses changes the locations within the shortened input but leaves the affected interval lengths the same.

## What remains invariant

The [script inventory](boundary_impact_inventory.json) records hashes, imports and relevant short literals for **82 retained Python files**. That inventory aids dependency review; it is not a formal proof about every program.

The numerical checks establish **37 exact equalities** between the original and compatibility inputs. These include all 73 page-level rune streams, both historical and current lexical word records, word-length fingerprint blocks, structural unit contents, unchanged training sources, and the Session 27 feature matrices. The total remains 15,933 runes across the parsed corpus. The unsolved LP2 text remains 12,956 runes and 2,896 current lexical words, including the short nine-rune segment.

The current word count differs from the older 2,899-token count because Session 42 joined three apostrophe suffixes. This session changes neither count under its respective convention.

| Earlier work | Dependency and result |
|---|---|
| Rune frequencies, immediate-repeat counts and rune-only feedback bounds | Rune values and section contents are identical |
| Word schedules, repeated words, singleton constraints and word resets | Both parsers return identical words, values and offsets before and after these three corrections |
| Word-length fingerprints, including the corrected LP1 page | The underlying blocks and lengths are identical; this does not repair that older experiment's separate apostrophe convention |
| Continuous, line, page, paragraph and word counters | Both counter indices and reset flags are identical under the tested options |
| First-clause removal | Its literal full-stop rule removes the same runes |
| Session 2 selective-F beam search | Targets only LP2 pages 0–2 and 54–55, outside the corrections |
| Session 3 shuffled calibration | Targets only sections 0 and 4, outside the corrections |
| Literary and cuneiform keys using continuous, page, paragraph or word schedules | Their rune, word and counter inputs are unchanged |
| Red-mark policies and numbered-list extraction | The corrections alter neither the red-event ledger nor the literal full stops or selected list endpoints |

Across the Session 2 grid, **384 counter/flag comparisons** cover eight substantial sections, two scopes, six reset rules, two punctuation-advancement choices and two F policies. **16 comparisons change**, all clause-reset variants in sections 7 and 8. The other 368 are identical.

There is an important scope caveat for **Session 27**: its predefined punctuation set was `. , & §`. It excludes both hyphens and semicolons, so its 21-feature numerical matrices remain exactly unchanged. Its result must not be generalized to a test of every physical three-dot mark. Adding that physical class would define a new analysis, not correct the execution of the old one. Raw displayed gap strings can also change even when a numerical feature does not.

## Session 2: replay only the affected clause branches

First, reconstruction of the original calculations reproduces both saved top-ten lists, including their scores and text. This checks the replay against the retained record.

Then **33,280 existing clause parameter settings** are reevaluated, 16,640 per affected section. These are the original eight streams, offsets 0–64, signs, F policies, punctuation-advancement options and two input scopes. No key family or range is added. Unchanged branches are carried into the merged rankings with their original ordering for ties.

Each clause family chooses its best offset using the original prefix criterion. There are 256 such family winners per section; these are not 256 independently discovered candidate solutions.

| Effect | Section 7 | Section 8 |
|---|---:|---:|
| Family winners whose selected offset changes | 0 / 256 | 123 / 256 |
| Family winners whose prefix score changes | 0 / 256 | 256 / 256 |
| Family winners whose suffix score changes | 256 / 256 | 123 / 256 |
| Overall best candidate, including text and scores | Unchanged | Unchanged |
| Top-ten parameter identities and ordering | Unchanged | Changed |

The difference has a simple cause. Section 7's correction is after its selection prefix, so it can change continuations without changing prefix selection. Section 8's correction lies inside its selection prefix. Some families therefore choose a different offset, which can also change their later continuation even after the counter schedules have rejoined.

The strongest section-8 clause candidate keeps the same square-stream parameters and offset 22. Its prefix score improves from **−7.688924 to −7.681172**, moving it from rank 5 to rank 2. Its continuation score remains **−7.793231**, and its text remains incoherent. The overall winner still has the higher prefix score of approximately **−7.679740**.

These are scores from the small retained rune-language model. A higher score is not a probability of successful decryption. The altered leading passages and continuations were inspected; no coherent new reading emerged. This was not an exhaustive human inspection of all 33,280 outputs, and the search does not exclude other cipher families.

## Session 3: rankings are mathematically unchanged

Session 3 selects parameters using only the first **256 runes** of each substantial section. Both corrections occur later. The audit also checks the less obvious dependency: full-section counter maxima determine which offsets fit the finite key streams. **The eligible offset sets remain identical for every stream and both F policies.**

Together, these equalities establish that all candidate prefix scores, eligibility and rankings are unchanged under the original rules. There is no reason to repeat the full parameter search or the unchanged random calibration.

All 40 retained top-twenty baseline decodes on the two affected sections reproduce exactly. Their **seven clause-reset candidates** are replayed with fixed parameters; the other 33 retained candidates are unchanged.

| Section | Retained rank | Changed decoded runes | Old suffix score | Revised suffix score |
|---|---:|---:|---:|---:|
| 7 | 11 | 29 | −7.797547 | −7.788903 |
| 7 | 14 | 29 | −7.801129 | −7.808041 |
| 7 | 17 | 25 | −7.833230 | −7.826581 |
| 8 | 5 | 177 | −7.795866 | −7.797056 |
| 8 | 6 | 177 | −7.795866 | −7.797056 |
| 8 | 17 | 191 | −7.802766 | −7.791473 |
| 8 | 20 | 185 | −7.785190 | −7.792044 |

Some suffix scores rise and others fall. The changed regions remain incoherent. The two matching rows at ranks 5 and 6 use two previously enumerated stream conventions that produce the same output here; they are not independent supporting results.

## What this closes, and what remains open

The source corrections warranted a bounded sensitivity check. They do not overturn the reviewed experiments' lack of a verified decryption. The historical reports remain preserved, with links to this follow-up where their numerical details are affected.

The known-answer tests in Session 62 also remain relevant: simple resets at observed three-dot marks break both tested position-dependent known solutions. Applying those resets to unsolved sections here is a sensitivity hypothesis, not evidence for a universal rule.

The punctuation cleanup and its impact audit are now complete at the stated scope. The next cryptanalytic question should concern a mechanism that can explain both the adjacent-repeat deficit and the repeated six-rune passage. Further arbitrary key sweeps are not justified by these modest score changes. Existing negative results and recovery controls must constrain that next test, rather than restarting the same families under new names.

## Evidence and reproduction

- [Fixed impact plan](boundary_impact_plan.json)
- [Retained-script inventory and hashes](boundary_impact_inventory.json)
- [Invariant checks, changed indices, candidate deltas and full revised leading outputs](boundary_impact_results.json)
- [Reproduction code](boundary_impact_audit.py)
- [Session 62: source map and known-answer controls](../sessions/62.html)
- [Original Session 2](../sessions/2.html)
- [Original Session 3](../sessions/3.html)

Run `python3 -B boundary_impact_audit.py` beside the retained inputs and helper scripts. Requires Python and NumPy. It verifies the immutable master hash, reconstructs the compatibility view from the source map, checks the declared invariants, reproduces saved baselines and writes separate derived results. The combined-stream helper reconstructs its existing finite prime/Fibonacci streams during import.

This is retrospective sensitivity analysis on previously explored material, not new independent evidence or a claim of first discovery in the solving community. The master SHA-256 remains `b0b2e13c3cd35512ee5955a2d9a56c22d7410e283d35dd9c90abbaa7b99279eb`.