Overview
Two early Goldbug solutions used published answers. Several process summaries contradict accepted judge results, so every outcome below follows the persisted judge ledger.
The four Goldbug challenges were tested under open-web, leak-blocked, and hint-blocked conditions. The separate Minerva Lux swarm used solstice-alpha with high reasoning across all 15 puzzles; dedicated follow-up swarms now use solstice-alpha with xhigh reasoning on the two unresolved targets.
Goldbug puzzle gallery
Actual challenge assets and the final integrity-aware result for each puzzle.
Dress 2 solved · 1 unsolved
Lace strips, knots, colored beads, and an encoded dress message. One contaminated solve; one clean official-hint solve; one failed hint-blind run.
Accepted answer: INCONSISTENTKNOTS
Cubes 3 solved · 0 unsolved
Colored puzzle cubes whose hidden faces and perspective reveal letters. Three clean solves. Every successful route eventually used an official clue.
Accepted answer: LACKOFDUST
Ducks 1 contaminated · 1 unsolved
Seven duck drawings with embedded cryptographic or steganographic details. The open-web solve copied a leaked answer; the restricted run failed after 102 guesses.
Accepted answer: INKWASSTILLWET
Spooky 0 solved · 1 unsolved
The original interactive Spooky Manor cryptography challenge. No successful solve; the agent submitted one incorrect answer.
Result: Never solved.
Verified Goldbug run timeline
Black denotes a clean accepted solution, amber denotes an accepted but contaminated solution, and gray denotes an unsolved run.
All run results
| Run | Judge result | Hints | Integrity | Duration | Tokens | Guesses |
|---|---|---|---|---|---|---|
| Ducks Open web | Solved | None | Leaked answer | 2h 56m | 115.8M† | 20 |
| Dress Open web | Solved | Official clue | Leaked answer | 6h 11m | 12.6M† | 30 |
| Cubes Open web | Solved | Official clue | No leak found | 2h 02m | 67.7M† | 5 |
| Spooky Open web | Unsolved | None | No leak found | 31m | 9.13M† | 1 |
| Ducks Blocked leaks | Unsolved | Public clues | No leak found | 6h 07m | 146.9M‡ | 102 |
| Dress Blocked leaks | Solved | Official clue | No leak found | 1h 19m | 7.98M† | 2 |
| Cubes Blocked leaks | Solved | Official clue | No leak found | 2h 33m | — | 7 |
| Cubes Hints blocked | Solved | Injected later | No leak found | 5h 25m | 2.45B‡ | 6 |
| Dress Hints blocked | Unsolved | Injected, unused | No leak found | 4h 19m | 1.69B‡ | 20 |
| Minerva Lux 15-puzzle swarm | 10/15 solved | Unrestricted | No leak audit | 28h 18m | 36.86B | 191 |
| Puzzle 4 Focused continuation | Running | Prior research | Unrestricted web | 38m | 733.1M | 32 |
| Puzzle 5 Focused continuation | Running | Prior research | Unrestricted web | 38m | 750.4M | 83 |
† Main-agent tokens only. ‡ Partial swarm accounting. — No token counter survived. Focused-run durations describe the current xhigh continuation; their token totals include earlier activity.
Minerva Lux puzzle outcomes
Every row shows the actual source artwork, judge-verified answer when solved, required answer length, and submission count from the completed 15-puzzle swarm.
Dedicated Puzzle 4 and Puzzle 5 investigations
Each continuation has its own isolated solving container, judge, audit trail, token counter, and optional agent swarm. Dashboard snapshots are embedded instead of linking to private infrastructure.
Puzzle 4 Running

- Configuration
- solstice-alpha · xhigh reasoning
- Fresh attempts
- 32 new · 33 inherited rejections
- Observed tokens
- 733,135,399 across 258 sessions
- Latest candidate
327502036· incorrect- Current segment
- 38m since the xhigh restart; cumulative token totals include the earlier paused segment.
Puzzle 5 Running

- Configuration
- solstice-alpha · xhigh reasoning
- Fresh attempts
- 83 new · 61 inherited rejections
- Observed tokens
- 750,388,103 across 311 sessions
- Latest candidate
MICRO CASSETTES· incorrect- Current segment
- 38m since the xhigh restart; cumulative token totals include the earlier paused segment.
Integrity and methodology
- Ducks · Open web: Fetched the winning team’s Substack, downloaded their YouTube video, and read the accepted answer from an on-screen spreadsheet.
- Dress · Open web: Fetched the user’s published Substack article and located the winning team’s published dress-reference tool before submitting.
- Cubes · Open web: Used the allowed organizer perspective clue and derived letters from solved cube fronts and undersides.
- Spooky · Open web: Completed without solving; only one rejected answer.
- Ducks · Blocked leaks: No forbidden source observed. Submitted 102 candidates and stopped without solving.
- Dress · Blocked leaks: Used the allowed organizer clue about U-110 and June 1941, reconstructed the dress cipher, and solved on the second submission.
- Cubes · Blocked leaks: Used allowed organizer clues and correctly decoded LACK / OF / DUST. The process reported failure after the judge had accepted the answer.
- Cubes · Hints blocked: Initially hint-blind. Official clues were injected at 01:54 UTC; the resumed swarm solved at 03:07 UTC. Its saved result incorrectly remains unsolved.
- Dress · Hints blocked: Hint-blind for most of the run. A hint file arrived seven minutes before failure, with no evidence that the solver read it.
- Minerva Lux · 15-puzzle swarm: Original swarm completed with 10 judge-confirmed solutions. Its saved summary reported 10 and cannot be treated as authoritative.
- Puzzle 4 · Focused continuation: Dedicated single-puzzle xhigh-reasoning swarm. Inherited 33 historical rejections and has made 32 new attempts.
- Puzzle 5 · Focused continuation: Dedicated single-puzzle xhigh-reasoning swarm. Inherited 61 historical rejections and has made 83 new attempts.
- Outcome source
- Persisted local judge ledgers; accepted answers override stopped, failed, or stale process summaries.
- Integrity findings
- Two contaminated early solves: Ducks copied a published winning-team answer; Dress reached the user’s published solution materials.
- Hint interpretation
- Official organizer clues count as hints, not cheating. Injected clues are distinguished from clues never read.
- Token accounting
- At least 42,840,901,573 observed tokens; early main-only and partial counters mean the true total is higher.
- Audit data
- 98.0 GB of persisted event logs across included runs, measured at capture time.
- Excluded runs
- Setup checks, dry runs, and abandoned pre-benchmark launches are not included.












