Portfolio record · No. 001
FLOTILLA
24 July 2026 · Journey 0

Research portfolio memorandum

Five theses. One budget.

A falsifier-first strategy table for deciding what earns another experiment—and what returns its capital to the fleet.

The model proposes. Executable predicates decide. Operators confirm every stop. Kills remain reversible, with the original evidence and lineage intact.

5Theses charted
8Runs completed
3Courses retained
2Stops confirmed
10Decisions recorded
0Undetermined

Deterministic browser lab

Interactive strategy simulator

Temporary in-memory data only. Choose conditions, set the mandate, launch the falsifier fleet, step experiments, adjudicate stops, reallocate reserve, and inspect the lineage. Reset returns to the registered fixture.

Installed control plane

Inspect a real local workspace

The simulator below always stays in fixture mode. Start flotilla serve, then connect here for a read-only view of the persistent SQLite portfolio.

Fixture mode active

Ready to launch five equal-cost falsifiers.

Evolving thesis trajectories

Round 0

Live capital chart

1.0 reserve

Temporary event stream

Simulation lineage

  1. 00Simulator readyAwaiting launch

Strategy table

Thesis trajectories

Each course begins with the same one-unit falsifier. Survivors receive follow-up capital; stopped theses retain their evidence and a route to revival.

T-01

Feature selection beats regularization at small n

KILLED

Selected features improve paired holdout accuracy by at least 1 percentage point.

  1. RegisteredThesis v1
  2. Falsifier1.0u · seed 101
  3. KilledChallengeable stop
1.0 / 4.0 units PENDING_KILL → KILL
Open evidence dossier

Registered rule

ci_lower < 0.01

Decision deadline
2026-07-25T17:00:00Z

Stop recorded. REVIVE remains available without erasing this lineage.

Conditions under which this may be wrong

  • The fixture covers four small tabular splits only.
  • Alternative selectors and tuned regularization remain untested.

Run lineage

  1. falsifier · falsifier · 1.0u
    commit bundled-paired-measurements-v1 · data d6ef7841ca77033f9a40a65a52b660b4cfd8560b9679d70c0b09b0b115cf745b · seed 101

Decision lineage

  1. PENDING_KILLrun T-01-v1-falsifier · confirmer noneci_lower < 0.01
  2. KILLrun T-01-v1-falsifier · confirmer demo-operatorci_lower < 0.01
T-02

Augmentation helps below one thousand samples

SURVIVED

A registered augmentation improves paired accuracy by at least 2 percentage points.

  1. RegisteredThesis v1
  2. Falsifier1.0u · seed 102
  3. Ablation2.0u · seed 202
  4. SurvivedCourse retained
3.0 / 4.0 units CONTINUE → PROMOTE
Open evidence dossier

Registered rule

ci_lower < 0.02

Decision deadline
2026-07-25T17:00:00Z

Promoted after its approved plan completed without a kill.

Conditions under which this may be wrong

  • Only one augmentation family is represented.
  • Results above one thousand samples are outside this falsifier.

Run lineage

  1. falsifier · falsifier · 1.0u
    commit bundled-paired-measurements-v1 · data fcd16820a23d4c41d859857029c45195b4eabecf1a9a394ae8cbee6393f0993b · seed 102
  2. ablation · ablation · 2.0u
    commit bundled-paired-measurements-v1 · data e40bfc120adaeab2cf4aac3e574dfdc38319d975353bb809569f7c2108628409 · seed 202

Decision lineage

  1. CONTINUErun T-02-v1-falsifier · confirmer noneci_lower < 0.02
  2. PROMOTErun portfolio · confirmer noneci_lower < 0.02
T-03

Aggressive regularization improves noisy labels

KILLED

Aggressive regularization improves paired accuracy by at least 1 percentage point.

  1. RegisteredThesis v1
  2. Falsifier1.0u · seed 103
  3. KilledChallengeable stop
1.0 / 4.0 units PENDING_KILL → KILL
Open evidence dossier

Registered rule

ci_lower < 0.01

Decision deadline
2026-07-25T17:00:00Z

Stop recorded. REVIVE remains available without erasing this lineage.

Conditions under which this may be wrong

  • Noise is synthetic and class-symmetric.
  • The regularization grid has only one aggressive setting.

Run lineage

  1. falsifier · falsifier · 1.0u
    commit bundled-paired-measurements-v1 · data 1e3a8bea9ec63584c2b262bfe23ceaf2b43707aa0d00fbffb04ff6e087d36008 · seed 103

Decision lineage

  1. PENDING_KILLrun T-03-v1-falsifier · confirmer noneci_lower < 0.01
  2. KILLrun T-03-v1-falsifier · confirmer demo-operatorci_lower < 0.01
T-04

Temperature scaling improves calibration

SURVIVED

Temperature scaling improves the registered calibration score by at least 0.5 points.

  1. RegisteredThesis v1
  2. Falsifier1.0u · seed 104
  3. Replicate2.0u · seed 204
  4. SurvivedCourse retained
3.0 / 4.0 units CONTINUE → PROMOTE
Open evidence dossier

Registered rule

ci_lower < 0.005

Decision deadline
2026-07-25T17:00:00Z

Promoted after its approved plan completed without a kill.

Conditions under which this may be wrong

  • The fixture uses one shift family.
  • Accuracy preservation is tracked but not part of the kill predicate.

Run lineage

  1. falsifier · falsifier · 1.0u
    commit bundled-paired-measurements-v1 · data c99aafa89692d5a9e362fb3ca4391a5d4b8e78b706069e92a0374a54b161a5a4 · seed 104
  2. replicate · replicate · 2.0u
    commit bundled-paired-measurements-v1 · data 844c3dbadb536a570721129891b783cd17308e782cf2d1ff895179a121c4033b · seed 204

Decision lineage

  1. CONTINUErun T-04-v1-falsifier · confirmer noneci_lower < 0.005
  2. PROMOTErun portfolio · confirmer noneci_lower < 0.005
T-05

Bagging stabilizes shallow trees

SURVIVED

Bagging improves paired accuracy by at least 1.5 percentage points.

  1. RegisteredThesis v1
  2. Falsifier1.0u · seed 105
  3. Replicate2.0u · seed 205
  4. SurvivedCourse retained
3.0 / 4.0 units CONTINUE → PROMOTE
Open evidence dossier

Registered rule

ci_lower < 0.015

Decision deadline
2026-07-25T17:00:00Z

Promoted after its approved plan completed without a kill.

Conditions under which this may be wrong

  • Only shallow trees and four registered splits are tested.
  • Inference cost is not included in this falsifier.

Run lineage

  1. falsifier · falsifier · 1.0u
    commit bundled-paired-measurements-v1 · data 74224faf61daf6642fc5f55329cf785e6e8cc6cbb05903e5940c00a045c32564 · seed 105
  2. replicate · replicate · 2.0u
    commit bundled-paired-measurements-v1 · data 661d7509c96faed53cd771be5f1fb7a6c5d5c380d55c8a387e572eeb64f4e3e9 · seed 205

Decision lineage

  1. CONTINUErun T-05-v1-falsifier · confirmer noneci_lower < 0.015
  2. PROMOTErun portfolio · confirmer noneci_lower < 0.015

Capital flow

From fair floor to concentrated follow-up

Falsifiers 5.0
Follow-ups 6.0
Reserve 1.0
“Test broadly for one unit. Concentrate only after a thesis survives contact with evidence.”

The first five units fund one falsifier per thesis. Six more fund three registered follow-ups. One remains uncommitted. This fixture illustrates the allocation mechanism; it is not scientific validation.

Append-only evidence

Decision register

Open thesis register 5 entries · predicate + spend + disposition
IDThesisStatus Spend / capKill predicate
T-01Feature selection beats regularization at small nKILLED1.0 / 4.0ci_lower < 0.01
T-02Augmentation helps below one thousand samplesSURVIVED3.0 / 4.0ci_lower < 0.02
T-03Aggressive regularization improves noisy labelsKILLED1.0 / 4.0ci_lower < 0.01
T-04Temperature scaling improves calibrationSURVIVED3.0 / 4.0ci_lower < 0.005
T-05Bagging stabilizes shallow treesSURVIVED3.0 / 4.0ci_lower < 0.015
Decision timeline 43 append-only events
  1. 0109:01Thesis RegisteredT-01
  2. 0209:02Plan ProposedT-01
  3. 0309:03Plan ApprovedT-01
  4. 0409:04Thesis RegisteredT-02
  5. 0509:05Plan ProposedT-02
  6. 0609:06Plan ApprovedT-02
  7. 0709:07Thesis RegisteredT-03
  8. 0809:08Plan ProposedT-03
  9. 0909:09Plan ApprovedT-03
  10. 1009:10Thesis RegisteredT-04
  11. 1109:11Plan ProposedT-04
  12. 1209:12Plan ApprovedT-04
  13. 1309:13Thesis RegisteredT-05
  14. 1409:14Plan ProposedT-05
  15. 1509:15Plan ApprovedT-05
  16. 1609:16Budget SpentT-01
  17. 1709:17Run CompletedT-01
  18. 1809:18DecisionT-01
  19. 1909:19DecisionT-01
  20. 2009:20Budget ReleasedT-01
  21. 2109:21Budget SpentT-02
  22. 2209:22Run CompletedT-02
  23. 2309:23DecisionT-02
  24. 2409:24Budget SpentT-03
  25. 2509:25Run CompletedT-03
  26. 2609:26DecisionT-03
  27. 2709:27DecisionT-03
  28. 2809:28Budget ReleasedT-03
  29. 2909:29Budget SpentT-04
  30. 3009:30Run CompletedT-04
  31. 3109:31DecisionT-04
  32. 3209:32Budget SpentT-05
  33. 3309:33Run CompletedT-05
  34. 3409:34DecisionT-05
  35. 3509:35Budget SpentT-02
  36. 3609:36Run CompletedT-02
  37. 3709:37Budget SpentT-04
  38. 3809:38Run CompletedT-04
  39. 3909:39Budget SpentT-05
  40. 4009:40Run CompletedT-05
  41. 4109:41DecisionT-02
  42. 4209:42DecisionT-04
  43. 4309:43DecisionT-05