PDF-derived exploratory map
Existential Puzzle Mapper
Rings
- Source signal
- Source puzzle and barriers
- Reference lenses
- Moves, outcomes, and risks
- Added pieces
Project Overview
A static website for keeping the source, map, options, and next moves cleanly separated.
Scenario Branches
Alternative directions without forcing one total map.
Simulation Workbench
Start with a working monomyth baseline. Change the role of insight, loosen the ordering, and compare alternative processes.
Synthetic experiments, not evidence about your experience. No private account is needed or transmitted. These rates describe this rule-based model, not the prevalence or truth of an ideology. Map position and node movement play no part in the results.
How to use this to develop a process
- Run the starter comparison, then inspect a successful case and an unfinished case.
- Change insight placement, or compare open and gated access. Keep the seed and budget fixed.
- Look for patterns that survive a changed condition, not just a better percentage.
- Repeat with another seed. Export the recipe or add a clearly labeled synthetic finding to the map.
Rules, role definitions, and limits
The synthetic world has eight possible solutions and three truthful binary clues. Participants do not see the solution. A check compares a claim with the synthetic oracle. Rejection eliminates only that claim, never establishes its opposite.
Every action costs one unit. Calls are noticed with 80% probability. The access setting controls the fraction of clues requiring permission; half of gated clues require two levels. Permissions, peer exchanges, and final sharing use the cooperation setting. Direct investigations use the investigation-failure setting. A peer can supply any missing clue without ranked permission. These are modeling assumptions, not measured psychological or spiritual probabilities.
The standard policy imposes ordering and waits for all clues before checking. The other policies can check earlier. Directed alternative policies explore another available action 30% of the time; the unordered control selects uniformly from available actions. Same seeds pair the worlds and action-attempt draws, not necessarily the actions taken. Policies change several rules together; attributing a difference to one rule requires a further controlled comparison.
Insight-placement labels describe the hypothesis being enacted. Only a successful check counts as a verified boon. The roles are related through events and their consequences, not independent truth values.
Limits: this prototype has a known synthetic answer, truthful clues, fixed policies, and no language understanding. Its signatures cannot determine whether a real narrative is monomythic or a worldview is Gnostic. Baseline matches are calibration, not spontaneous discoveries. Bounded failures can look like non-matches. Neither means neither detector matched, not absence of structure. Rates are descriptive counts, not confidence or belief scores.
Learning studies: repetitions, regression review, and expanded conditions
Why did the original later forecasts get worse?
Loading regression controls...
Worse means higher prediction error, not lower simulated success. The original stored
trainingError measured the retained model, not each rejected candidate.
The reconstruction below distinguishes those two measurements on the same training data.
Largest target regressions
Controlled follow-up on the original validation set
Using nine common training seeds did not remove the later regression. Spreading the same added run count across anchors and roles reduced it. This supports a coverage and model-capacity trade-off, not a single proven cause. The controls also change which settings receive attention; they do not isolate insight imbalance alone. They reuse original validation, not the original test outcomes.
Download the regression notes and expansion findings. The full control recipes, seeds and counts are included in the expanded study export.
Switching studies changes only the displayed study and learned forecast, not your recipe, simulation counts, or saved map. The original model remains the default; broader coverage was not a universal improvement.
Loading the published synthetic learning study...
A multi-output regression tree learns to forecast shared outcomes, monomyth-like matches, hierarchy matches, and their overlap for all four policies. The simulator and its definitions were not changed to improve the scores. This is supervised learning from synthetic counts, not learning whether an ideology is true.
Final unseen-scenario evaluation
Error is the mean absolute difference in percentage points from repeated simulation rates. The baseline always predicts the training average and ignores scenario conditions. Validation was reused for selection; test recipes and seeds were held apart until the model was frozen.
Selected study method and limitations
Original and expanded models on the same fresh tests
Compare predictors within each row, not this broader study's overall error with the original study's 2.68-point score. Outside-range original predictions are extrapolation stress tests only; they remain blocked in the forecast UI. Both models were fixed before these test outcomes were generated.
Inspect repeated series
Each mean pools the repetitions below. The seed range and standard deviation describe repeat-to-repeat variation, not a confidence interval or independent evidence about the source material.
Replay runs the original simulator so you can inspect its trace. Study exports include all repetition counts,
recipes, seeds, model parameters, and source-code hashes. Use npm run learn for the original study,
or npm run learn:followup for regression controls and expansion. Append :check
to either command name to reproduce and check its saved artifact.
Comparative results
Each cell counts matching cases, including failures in the denominator. Signatures may overlap.
Hierarchy inside and outside the monomyth signature
Each rate uses only cases in that column. An empty group is not a zero occurrence rate.
Learned forecast versus this simulation run
These are estimates from the frozen learning model, not additional observations. Observed counts above remain unchanged. A single seed may differ from a repeated mean.
Inspect evidence and counterexamples
Objectives and Outcomes
Possible aims, first moves, success signals, and failure modes.
Suggested Next Steps
Different ways to continue the exploration without forcing one final structure.
Pattern Lab
Non-standard pattern candidates for future data collection and conceptual mapping.
Simulated counts are kept separately from the real-observation entry counts below.
Technology / Term Key
Plain definitions for the vocabulary and mapping technology used by this project.
PDF Source Anchors
These are derived from the downloaded PDF and transcript.
Reference Lenses
These are separate comparison frameworks, not source claims.