Boundary corrections meet known-solution controls

The corrected boundary map is built. Five simple key-counter policies each fail at least one applicable known-solution control. Restarting at a three-dot mark fails both the second koan's repeating-key solution and the prime-based LP2 page-56 solution. No unsolved plaintext was searched or recovered.

The result limits blanket boundary assumptions. It does not prove that the unsolved sections use the same cipher or cannot have their own reset rules.

What the new map preserves

The derived boundary map covers all 73 parsed master text/data pages, containing 15,933 runes. This means the 15 parsed LP1 text pages and all 58 LP2 pages; the LP1 title and illustration-only page are outside the parsed master. Some LP2 pages contain non-runic data.

The map contains 4,216 boundary entries, including page starts and ends, line wraps, punctuation, structural tokens and non-runic data. That number is not a count of printed punctuation marks.

Separately, it attaches 406 distinct audited physical events on 26 images: 374 black marks and 32 red marks. Identical source events found in multiple audits are deduplicated by image, color and bounding box only after their offsets and dot counts agree. Their evidence references remain attached. Different marks at the same rune gap stay separate.

The three split locations on LP2 pages 5, 11 and 20 retain their separate 3+1 or 1+3 fragments. They are not automatically converted into ordinary four-dot marks or key-reset instructions.

Two red marks at the same rune gap, as previously documented on archive 11 and 12, likewise remain two physical events. A single transcription character cannot capture that distinction.

The master file is unchanged. Its SHA-256 remains b0b2e13c3cd35512ee5955a2d9a56c22d7410e283d35dd9c90abbaa7b99279eb. The map adds evidence without altering rune values or pretending that unaudited locations have been checked. Whitespace is omitted from the event lists; the original file remains the reference for exact transcription formatting.

Establish the known answers before testing rules

The fixed plan specifies two position-dependent controls and one directly readable comparison.

Second koan, archive 14–15: 319 runes. We reproduced the documented repeating-key method with its fixed 13-rune key, continuous across both pages, and two unenciphered F exceptions. The complete decoded alphabetic reading matches the retained English translation after explicit spelling normalization. Re-encryption reproduces every ciphertext rune.

The archived implementation notes give skip positions 65 and 76. These correspond to one-based character positions in the space-containing input. Our normalized reconstruction of that input confirms that they map to zero-based rune offsets 49 and 58. Applying the published numbers directly as rune indices would address the wrong symbols. The second archive also documents the continuous two-page solution. This is reproduction of an existing solution, not a new key discovery.

LP2 page 56: 85 runes. The established prime-minus-one method, with its unenciphered F at offset 56, exactly reproduces the retained known text, including the original master delimiters and non-runic data. Re-encryption is exact.

LP2 page 57: 95 runes. Its direct reading matches the retained record. Because this method has no advancing key counter, it cannot distinguish the reset hypotheses tested here. A supposed counter test that reports success on this page alone would be uninformative.

Fixed counterfactual policies

Each comparison keeps the known key, initial phase and F exceptions fixed. There is no phase fitting or search for replacement keys. A reset means restart the same key or prime sequence at index zero before the next rune. An advance means consume one extra key position.

Only physically audited, individually aligned marks are eligible. Five four-dot groups on the two koan pages received additional manual line/gap alignments from the retained Session 61 screen; these are saved in the control alignment ledger. Unaligned candidates and unaudited literal periods are not silently promoted into observed four-dot events.

The two three-dot policies happen to produce the same mismatch total on the koan. That does not mean their decoded outputs or counter schedules are identical.

The koan's two three-dot locations are stream gaps 110 and 286; the latter is the corrected gap 60 on archive 15. Its five four-dot control gaps are 36, 71, 145, 214 and 291. The page transition is gap 226.

On LP2 page 56, the three-dot gap is 19. The audited red marks at gaps 5 and 45 can affect later runes; the final mark at gap 85 cannot and is excluded from active actions.

These numbers count disagreements with known plaintext, not low language-model scores. Every policy with an active applicable event fails its corresponding known answer. Cases with no active audited event are labeled uninformative, not counted as successful support for the policy.

Counter changes are not always letter changes

Restarting at the koan's page break changes the remaining 93 counter indices but only 71 decoded runes. Its repeating key contains repeated values, so some changed indices still select the same key symbol. The first wrong rune appears at offset 227, one rune after the boundary.

This is why the saved results report both counter changes and decoded mismatches. Merely demonstrating a new index schedule is not a decipherment or even a guarantee of different text.

What this establishes

We now have a reusable separation between three kinds of information: literal transcription tokens, measured physical marks, and proposed key operations. The three verified corrections are available to future analyses without altering the evidence underneath them.

The known controls reject the specified resets or extra advancement as universal rules for these methods. They do not eliminate a section-specific cipher, an untested trigger, a different reset phase, or a different operation. No such alternative earns support merely because this small family failed.

The practical checkpoint is now complete: future boundary experiments should consume this map, name exactly which observed events they use, and pass applicable known-answer controls. The punctuation work has produced cleaner inputs and concrete limits, not a reason to launch a larger speculative key sweep.

Evidence and reproduction

• Fixed map and control plan
• Derived boundary map
• Additional control-event alignments and method references
• Known-answer results, mismatch offsets and source hashes
• Map builder and control code
• Typography comparison
• Full LP2 three-dot screen

Run boundary_model_audit.py beside the retained inputs and lp_core.py. Requires Python and NumPy. The program checks source-image hashes, retains all source-event provenance, verifies the three overlays and three split pairs, reproduces the known baselines and records all fixed policy outcomes.

The controls are retrospective tests on known text. Manual source alignments are not an independent transcription, the image coverage remains partial, and none of these findings is claimed as new to the solving community.