# MVUEH: first parallel research round

**The four-agent campaign is complete. MVUEH remains unsolved.** The round produced a faster validated attack, a fresh handwriting assessment, better-grounded plaintext hypotheses and a stricter independent verifier. It completed **1,119 target experiments**, with **172,622,500 overlapping evaluations**. These counts are neither unique keys nor a fraction of the full Enigma search space.

The coordinator integrated the agents' findings and executed follow-up experiments within this round. All launched computations and all four agents finished. No background search, recurring task, new standalone task, external correspondence or purchased compute remains pending.

## What the parallel work changed

| Investigation | Result | Practical consequence |
|---|---|---|
| Fresh evidence reading | A 15-letter blind neighboring control matched the published transcription. The revised target reading differs from the phase-two preferred reading at six positions. | Both literal readings were tested separately; neither is promoted to a definitive transcript. |
| Historical context | An earlier explicitly N.F. → Ib message begins `BETRIEBSSPRUQX`. Ib-addressed examples do not require an encrypted address prefix. | Tested the service-message phrase family; downgraded the assumed `ANROEMEINSBERTA` address rationale. |
| Attack engineering | 37 paired comparisons retained byte-identical candidates; median elapsed ratios were 1.33× on repeated start-search batches and 1.39× on fixed-start controls. | Independently reviewed and used the faster engine in a bounded MVUEH pilot. |
| Independent review | Found that the old arithmetic audit did not independently enforce a claimed crib or uncertain-letter sets. Added and tested that check. | Future positive crib candidates must satisfy both machine arithmetic and the actual experiment constraints. |

The [protocol](evidence/experiments/campaigns/parallel-01/PROTOCOL.md) records ownership, independence and concurrency. The agents used separate working directories. The evidence agent initially received no preferred ciphertext or plaintext clues; its initial reading and control were frozen before source comparison. The historical agent did not inspect that target reading. Fresh model reasoning reduces anchoring but does not substitute for an independent human handwriting expert.

## Handwriting state

The fresh reader's revised preferred version is:

```text
MVUEH ICRVS ORMCC WQTAT
YEVFX DBZGG SNXWL PSYWZ
YBCBS WULST BRCUG ODVJU
SLSOO MJQJZ SXSEB ZPEYM
DNXJY TC
```

Relative to phase two, body positions 27, 37, 43, 44, 47 and 49 change from M/T/R/R/T/V to N/B/L/S/R/U. Positions are 1-based and exclude MVUEH. Alternatives and visual reasons remain in the [character catalogue](evidence/experiments/campaigns/parallel-01/evidence/character-alternatives.json). The final comparison was exposed to the public received transcription, so only the preserved initial reading and 15 control letters qualify as blind work.

The header still favors GTA/KCI; QTA/KCI is retained as a low-confidence visual possibility. KCC remains a prior alternative in the computational pilot. The fresh preferred body differs from the published received transcription at twelve positions. That discrepancy could involve copying, transmission, handwriting interpretation or the public transcription. Without the received scan, this round cannot choose among those explanations. [Evidence report](evidence/experiments/campaigns/parallel-01/evidence/REPORT.md), [published July messages](https://cryptocellar.org/bgac/g-army-july-1941.html).

The bounded source search again found the outgoing NF88/61 facsimile but not the original received No.172 image. A transient web-tool failure was resolved; the outgoing PDF is accessible. The public July compilation contains the typed entry, not a replacement received facsimile. No maintainer was contacted. [Search log](evidence/experiments/campaigns/parallel-01/evidence/SEARCH_LOG.md), [outgoing facsimile](https://cryptocellar.org/bgac/spruch/10071941-088-061-out-nf.pdf).

## New target computations

All grids below assume standard Enigma I, reflector B, ordinary indicator/body coupling and exactly ten plug pairs, except where a supplied board fixes those pairs. No target rotor order or ring triple has been established.

| Search | Exact bounded scope | Completed runs | Outcome |
|---|---|---:|---|
| Address continuation | Order 521, ring indices [32,160), both indicators, `ANROEMEINSBERTA` at every fitting offset, phase-two eight erasures | 256 | No partial board survived. |
| New crib work | Service-message prefix Q/CH forms across all 60 orders at JQH/MRP, plus order521 across all 52 JQ/MR right-ring settings; small X-ending, all-offset, Ib-location and artillery-prefix pilots | 704 | No completion satisfied the declared letter sets. |
| Header-copying-error screen | All 151 zero-or-one-character variants of GTA/KCI; two same-day boards, order521, all17,576 rings; fixed phase-two native body | 151 | Leading retained outputs incoherent; no solution accepted. |
| Optimized heuristic pilot | Two literal bodies × KCI/KCC × JQH/MRP; order521, starts [0,256), faster indicator-only seed method | 8 | Leading retained outputs incoherent; no solution accepted. |

The **960 crib jobs** total 167,312,500 overlapping crib/settings evaluations. Of these, **596 jobs had zero partial boards**: their negatives tolerate arbitrary A–Z choices at all erased positions. The remaining **364 jobs depend on the declared letter alternatives**; they reached partial boards but no compatible completed board. Every negative still depends on the crib and other observed letters, header, rotor/ring assumptions and machine model. The queue recorded one execution per job and checked actual inputs and hashes.

The **151 header hypotheses** account for 5,307,952 evaluations and 4,530 retained records. They include QTA/KCI and GTA/KCC, but not the two-error QTA/KCC combination. This is a scored screen using known boards, not an exhaustive unknown-board attack. Its positive SIPVX calibration recovered the known ring/start under a nominated correcting header; the board and order were supplied. [Header experiment and limitations](evidence/experiments/campaigns/parallel-01/review/REPORT.md).

The **eight heuristic runs** account for 2,048 ring/start states and 240 retained records. They deliberately use the faster method that previously recovered only three of six controls at the correct state. Failed heuristic runs provide no exclusion proof. The start window is AAA through AJV inclusive; nothing is inferred about the remaining starts. The [pilot files](evidence/experiments/campaigns/parallel-01/coordinator/heuristic/summary.json) include every output, independent checks and review decisions. Only the top30 combined outputs were individually assessed for language; all240 re-encrypt correctly, as arbitrary-key decryptions normally do.

## Attack improvement and verification

The optimized hillclimb skips duplicate cable-swap descriptions, computes cable-count changes directly and removes a redundant scan of locked connections. It preserves move order, scoring, indicator constraints and final plug count. It improves speed, not mathematical coverage or the strength of a failed search.

All 37 paired output files were byte-identical. Six controls recovered their exact boards with zero and four additional substitutions; five of six recovered with seven, retaining the earlier WNFGI failure. A deliberately answer-centered diagnostic varied three ring candidates and64 starts for CFYZR, recovering the exact ring/start/board at global rank1. Rotor order and the restricted search window were supplied. This is not evidence of full unknown-key recovery or an estimated MVUEH success probability. [Attack report](evidence/experiments/campaigns/parallel-01/attack/REPORT.md).

The reviewer separately rebuilt both engines, repeated a nontrivial32-start comparison and reran 7,151,146 move-equivalence checks. Sixty planted-key fixtures across all rotor orders tested the constraint solver at double-step boundaries and non-AAA rings. Six semantic-verifier tests included deliberately invalid crib and erased-letter candidates that the old arithmetic audit would accept. The existing eleven machine/constraint tests also passed. [Independent review](evidence/experiments/campaigns/parallel-01/review/REPORT.md).

The final coordinator audit rechecked **4,770 retained target records**, input/output/source hashes, queue uniqueness and preservation of the eleven frozen baseline files. Attack and review lane checksums covered313 and799 files respectively. The military frequency table was unchanged and may include material related to the controls; no held-out statistical success estimate is claimed. [Machine-readable completion audit](evidence/experiments/campaigns/parallel-01/SUMMARY.json).

## What should receive the next round

1. **Calibrate the deeper attack on uncertain ciphertext before expanding its key grid.** Represent the six disagreements between current readings as alternatives and test recovery on known messages with comparable ambiguity. Compare this with the two literal-reading approach. The faster pilot's limited recovery rate makes a much larger run premature.
2. **Separate transcription uncertainty from indicator uncertainty.** The original received No.172 facsimile and an independent human reading remain valuable. A request draft exists, but sending it requires explicit authorization; no search result proves that the image is unavailable privately.
3. **Expand only a declared, calibrated search.** The historical lane's eight-crib catalogue supplies testable hypotheses, not known target words. The new service-message prefix is better grounded than a mandatory Ib address, but its absence under a small assumed key grid does not identify the correct branch.

These are recommendations, not background jobs. The next experiment should demonstrate recovery under the uncertainty model before spending substantially more time on an unrestricted ring/board search.

See [reproduction instructions](evidence/experiments/docs/PARALLEL_ROUND1_REPRODUCE.md), the [research log](evidence/experiments/docs/RESEARCH_LOG.md), and the four lane reports for sources, commands, failed hypotheses, candidate records and code.