# What to try next on Kryptos K4

14 September 2026

The best next work is to recover a small, reproducible rule from the sculpture or its historical context, then test what that rule predicts outside the released plaintext. More unconstrained sentence generation and larger versions of the same classical searches have poor expected value. The public survey identifies substantial prior work in every major classical family, but also many overstated exclusions, incomplete input records, and proposed solutions whose apparent success comes from fitting the answer first.

The current evidence supports keeping several physical and procedural directions open. It does not support calling any of them the intended mechanism. The November 2025 disclosures identify the World Clock and historical events; they do not specify a city-list key, a clock setting, a transposition, or a compass selector. The authenticated fixed plaintext remains 24 letters. [Released-clue register](CLUES.md)

## Ranked next work

| Rank | Work | Why it can change the result | First finite deliverable |
|---|---|---|---|
| 1 | Audit proposed physical selectors with all fitted values removed | A reproducible decoder may still contain its plaintext in lookup tables. Removing those values reveals its actual predictive content. | A dependency graph labeling every constant as measured, convention, crib-derived, or plaintext-fitted, with the non-crib positions genuinely forced by the remaining equations. |
| 2 | Reconstruct a historically valid World Clock input | Modern labels and time-zone assignments differ from the original installation. A precise old input would support tests that provisional lists cannot settle. | Dated evidence for the particular sectors/labels needed by one declared extraction rule; unknown cells stay unknown. |
| 3 | Test literal instruction mechanisms from K0–K3 | The creator's earlier-text/later-text statement gives this direction a basis, but many generic key-word and route searches already cover it. | One short sequence of operations selected from source evidence before seeing its K4 output; replay against established examples and matched synthetic messages. |
| 4 | Investigate the Egypt/Berlin connection through source selection | A specific travel document, object, or message convention could supply a constrained input; a vast topical word list cannot. | A documented candidate source with a finite extraction rule and a provenance date preceding K4. |
| 5 | Reopen a classical family only for a demonstrated gap or bug | Public claims of total elimination often exceed their actual parameter bounds. | A minimal counterexample or exact unsearched cell, followed by a bounded run that names its closest prior attempt. |

These ranks reflect research efficiency, not estimated probabilities that a family is correct. Directions 2–4 need better input evidence before broad computation is justified. Rank 1 is immediately useful as a source audit; it can reject misleading evidence without pretending to exclude every related cipher.

## What the canvass changes

The technical and community catalogs contain 50 source/method records across 41 distinct normalized URLs, including context records and overlapping work. They cover periodic substitution, Quagmire variants, autokey, running keys, Gromark, matrix systems, fractionation, transposition, nulls, Morse, clocks, compass geometry, optical projection, historical texts, machine learning, and full-plaintext claims. They include older proposals as well as 2025–2026 material. The catalogs preserve source metadata, mechanisms, stated bounds, assumptions, and access limitations. They are a map of discoverable approaches, not proof that no private or unindexed attempt exists. [Combined index and overlap groups](catalog_index.json), [technical catalog](luna_approaches.json), [community catalog](luna_community.json)

Three distinctions prevent wasted work. First, a negative result for a specified alphabet, period, route, and operation order does not exclude the whole named cipher family. Second, a result fitted to all 24 clues has not predicted those clues. Third, a decoder that uses a lookup table learned from an alleged 97-letter plaintext has not independently recovered that plaintext merely because it runs forward.

The live [SolveKryptos proposal](https://solvekryptos.com/solution) now explicitly identifies its non-anchor text as a reconstruction used to build the model. Its [method description](https://solvekryptos.com/method) separates some rule-derived values from fitted tables and admits that geometry tests depend on those tables. That makes it a useful target for a dependency audit, but the full plaintext should not enter our clue register. A rule inferred from a phrase in the proposed answer remains conditional on that phrase until supported independently.

The [ogio3 spatial proposal](https://github.com/ogio3/k4/blob/main/docs/background.md) illustrates a separate problem. Its displayed shift array disagrees at 75 of 97 positions with ordinary A=0 ciphertext-minus-plaintext arithmetic for its own displayed candidate. At position 4, R minus C is 15 modulo 26, while the array prints 5. The candidate itself does align with all four released clues. This result invalidates that displayed array for those inputs; it does not disprove spatial methods or establish any plaintext. Two independent arithmetic implementations agree. [Reproducible audit](source_arithmetic_audit.py), [result](source_arithmetic_audit.json)

## How the released clues should guide work

The World Clock identification changes which physical object deserves attention. A city name, time-zone sector, rotating ring, compass direction, or meeting-place association are distinct hypotheses. Each needs its own rule and evidence. A historically correct input is essential whenever the rule depends on text engraved before Kryptos was installed. A contemporary [December 1997 restoration report](https://www.berliner-zeitung.de/article/20-staedtenamen-wurden-bei-der-restaurierung-ergaenzt-frisch-poliert-die-weltzeituhr-dreht-sich-bald-7079) describes 20 added city names and changed time-zone treatment. Recovering those differences is more valuable than cycling through guessed city lists indefinitely.

The Egypt clue suggests a source trail, not a license to try every Egyptian noun. The 1986 clue reporting and an earlier 1984 recollection should remain separately recorded. Tombs, mirrored light, relocated monuments, and the Carter text can motivate investigations, but a thematic resemblance cannot decide an extraction algorithm. Likewise, the Berlin Wall's fall could concern the message's subject or a historically selected key; the disclosure does not resolve that choice. [Source context and discrepancies](CLUES.md)

Morse deserves similar discipline. The visible fragments contain uncertain boundaries and extra marks; a normalized English phrase discards potential information. “T IS YOUR POSITION” can be treated as a visible fragment without asserting that the isolated T is a mathematical origin. The [Kryptos Project transcription](https://www.thekryptosproject.com/kryptos/k0-k5/k0.php) preserves ambiguities that several later proposals erase. Before testing a Morse mechanism, freeze whether the input consists of marks, letter groups, words, spacing, or geometric placement. Choosing among these after seeing favorable output creates hidden search freedom.

Scheidt's [2005 interview](https://www.wired.com/2005/01/inside-info-on-kryptos-codes/) describes obscuring the English-language signal and also distinguishes readable text from its covert meaning. It does not identify a standard masking algorithm or require exactly two cipher layers. This leaves room for procedural constructions while giving no support to unlimited arbitrary exceptions.

## What stays closed and what remains open

The previous campaign's exact exclusions remain intact. Its ordinary short-period columnar tests, independent column reversals, selected autokey arrangements, fixed mixed alphabets, progressive shifts, column resets, affine offsets, and inverse geometry results should not be rerun under new names. The width-31 period-6 structural fit remains a clue-consistent but incoherent result, not a lead strengthened by the new survey. [Previous campaign report](../REPORT.md), [previous admission and duplicate rules](../PRIOR_WORK.md)

Those results do not close all visual masks, every source-derived key, every fractionation system, or every sequence of operations. Conversely, the mere absence of a proof does not make a huge unexplored family worth searching. Admission requires a mechanism that is both distinguishable from prior work and constrained enough to test meaningfully. The new [coverage map](COVERAGE.md) records these distinctions at family level; individual experiment admission still requires exact parameters and input hashes.

Public search counters need particular care. The [kryptos.today plan](https://kryptos.today/plan) reports that an enormous attempt count consisted of repeated evaluation of only 96 candidates. Our accounting should count unique input/model/parameter combinations and stop when those are exhausted. Its unfinished and queued categories are historical project status, not evidence that nobody elsewhere has tried them.

## A token-efficient continuation

Use Luna for short source-specific tasks and hypothesis records. Keep source extraction, mathematical review, and enumeration separate. One worker should own each source family; a second should review a concrete claim or data extraction rather than repeat the same broad search. Workers return a compact result, provenance, and unresolved question, with detailed evidence stored in files.

Each admitted experiment should state: the closest prior attempt; the exact new difference; the source selecting that difference; finite bounds; a deterministic fingerprint; expected computation; success criteria; and a stopping rule. Enumeration belongs in code. Language-model tokens belong in selecting and auditing assumptions. Save a checkpoint at each completed batch and continue only when the result changes the next decision.

A candidate may progress from a necessary constraint survivor to a complete round trip, then to a coherent independently motivated plaintext. These are different milestones. Scoring must account for every variant tried, including failed ones. A small p-value after trying many interpretations is not evidence unless the selection process is included in its calibration. Synthetic recovery tests should use the same length, clue layout, alphabet restrictions, and search procedure as the real task.

The new native goal ceiling is 10,000,000 total tokens. It is not a target. The stopping condition for this research campaign is a useful source map, verified constraints, scoped audits, and actionable next experiments. Continuing into an unbounded search merely to consume the budget would make the research worse.

## Limits

Private group archives, deleted repositories, unavailable papers, and inaccessible source pages prevent a literal inventory of every historical attempt. Some records are based on abstracts or authors' claims rather than inspected execution artifacts; their catalog entries say so. No public method and authenticated complete plaintext were established in this survey. No new paid services or submissions were used. The recommendations above identify what would produce new evidence, rather than promising a solution from a larger token allocation.