Exact state, partial observation, and reversible recovery.
Paper PDF · Methods · Reproduce · Recorded results · State contract
Interactive laboratory · Illuminated Geometry · Book of Crossroads film — Revised Edition
A small, reproducible Python environment for testing whether a system recovers its entire state, rather than merely reproducing what a narrow observation window happens to show. It stores exact rational values, the ordered reading window, and the carried reference together. It also checks recovery by replaying a complete journal from a known initial state.
Research status · release candidate 0.1.0. A deterministic software study with public synthetic fixtures, an independent implementation check, and an inspectable methods paper. This is a reusable exact-state test environment; LLM performance, compression advantage and physical information transfer were not evaluated. Noncommercial research is free under the research license. Limited hiring and engagement review is also free under section 3A. Other enterprise work, including internal research and product evaluation, requires written permission and payment.
The same reading can appear twice while the underlying state has changed. Crossroads State keeps the full field, the order in which it is read, and its reference address together, so an exact return can be checked rather than inferred from appearances.
The figure below uses four addresses to make the rule visible. The complete synthetic study uses 81. Green numbers show the reader's order; gold marks the carried reference. Relabel every component together and the observed pair stays the same. Apply the inverse and the full state returns.
Inspect the figure's exact data · Regenerate the original vector artwork
The committed synthetic study uses 512 seeded instruction sequences, each 40 steps long, on 16 independently generated 81-address fixtures. It does not distribute the private source arrays or reading windows that motivated the prototype.
| Test | Result |
|---|---|
| Full checkpoint restored exactly | 512 / 512 |
| Known initial state + complete journal reproduced the endpoint | 512 / 512 |
| Recorded inverse restored the initial full state | 512 / 512 |
| Independently implemented oracle agreed at each trace step | 20,480 / 20,480 |
| Oracle agreed on primitive/count/reference cases | 3,402 / 3,402 |
| Array, reader, and reference mutations rejected without recomputing checksum | 512 / 512 each |
| Truncated journal changed endpoint | 512 / 512 |
| Reversed instruction order changed endpoint | 511 / 512 |
| Shuffled instruction order changed endpoint | 512 / 512 |
The one reversed-order agreement is retained. Different histories can lead to the same endpoint; a checkpoint is not a record of the route taken. Likewise, two different arrays can show the same nine readings. An included example becomes distinguishable after one declared reading-window shift.
The independent oracle is a separate coordinate-dictionary implementation, written against the specification without calling the production engine. This is implementation cross-checking, not external peer review.
Python 3.10 or later; the engine and tests use only the standard library. Run these commands from the repository root:
PYTHONPATH=src python3 -m unittest discover -s tests -v
python3 scripts/reproduce.py --checkThe first command runs 17 focused tests. The second recomputes every registered
case and compares the complete machine-readable result with the committed file.
To regenerate that file deliberately, omit --check.
The CI workflow runs both commands. Its live status is reported by the badge above; the committed study records the reproducible result independently of the hosting service.
PYTHONPATH=src python3 -m crossroads_state demo --output start.json
PYTHONPATH=src python3 -m crossroads_state verify start.json
PYTHONPATH=src python3 -m crossroads_state replay start.json examples/journal.json --output arrival.json
PYTHONPATH=src python3 -m crossroads_state replay arrival.json examples/journal.json --inverse --output returned.json
cmp start.json returned.jsonThe CLI refuses to overwrite an existing output. Its built-in model is the public synthetic fixture only. The Python API accepts other finite address sets and exact operation definitions; see the methods paper.
An array is the numerical field at every address. A reader is an ordered selection of addresses, so changing its order changes how the field is read. A reference is an address carried by the state, which moves when the coordinate frame moves. A journal is the ordered list of instructions.
A passive change of coordinates moves the array, reader and reference together. The observed values stay the same even though their address labels change. Changing a reader alone can expose previously hidden values. Adding a field changes the values themselves. Keeping these operations separate makes their inverses testable.
The fixture happens to use a four-label ternary address space: each label is -1, 0, or +1, giving 3 x 3 x 3 x 3 = 81 addresses. This is a finite computational address system. No astronomical, spiritual or physical interpretation is needed to execute or verify it.
- The checkpoint contains exact rational strings and a model fingerprint. A checkpoint for another operation definition is rejected.
- SHA-256 detects changes when the original digest is retained. It is unkeyed: an editor can change the state and recompute the digest. It does not prove authorship, authenticity, correctness of a model, or historical truth.
- Exact replay requires the known initial state, the full ordered journal, and the same instruction definitions. Endpoint recovery does not recover a unique history.
- The 512 tests are deterministic software checks, not a sample of LLM behavior. Possible agent-memory applications need separate comparative experiments.
- Parsing is intended for trusted local research files. Resource limits, hostile-input hardening, durable storage and distributed consistency are not supplied by this small library.
- Methods paper and PDF
- Registered protocol
- Test results
- Engine
- Synthetic fixture definition
- Independent oracle
- Privacy manifest
The runtime has no external dependencies, network calls, model APIs, telemetry, source downloads or credentials. Rendering the optional PDF uses ReportLab; the readable paper source and all computational checks do not require it.
The original code, paper, figures and interface assets use the TASUMER MAF Research and Enterprise License 1.0. Personal study, noncommercial teaching and academic research, reproduction of experiments, and publication of research findings are permitted without a fee under its terms. Permitted copies and modifications must retain the license, attribution and change notices.
Limited hiring and engagement review is free under section 3A of the license. Prospective employers or clients may review publicly released Material solely to decide whether to employ Timothy Paul Bielec or engage Timothy Paul Bielec or TASUMER MAF for professional services. Within that review, reviewers may make necessary local copies and run only unmodified, publicly supplied examples with their supplied sample inputs on devices under their control. All section 3A conditions apply, including limits on sharing and retention. External AI submission requires the Licensor's prior written approval of the specific service and use. This exception does not permit enterprise research, product evaluation, business adoption, or training.
Other enterprise or commercial use requires separate written permission and payment. This includes a company's internal research, benchmarking, product evaluation, pilots, and employee or contractor work outside the limited section 3A review, regardless of company size or whether a product is released.
See the plain-language guide and full terms. This is a source-available research release, not an unrestricted open-source license. Separately licensed third-party works retain their own terms.
The Crossroads Laboratory is the companion browser environment: explore numerical families, instructions, return checks and Illuminated Geometry. Its November 2 geometry update includes the red thread, packing and mirror lock, reference handover and memory return, with a separate three-minute study player. Visual staging and numerical operations are labelled separately. The library in this repository continues to use only the public synthetic fixtures documented above; its test results are unchanged.
The film presents the work as a visual journey. The initiate study environment contains the private courses and books and requires access. The public laboratory does not display the reserved source square or its ordered extraction key.
rhombic contains the separate XR-001: Typed State Beats Prose study, an empirical investigation of numeric corruption during agent context compaction. Crossroads State supplies a different kind of artifact: an exact, deterministic state contract and a synthetic environment in which recovery can be checked. The 512 sequences here are software tests, not additional LLM trials or a reproduction of XR-001. The two studies retain their own methods, data and claims.
Built by TASUMER MAF · Source and tests · Custody boundary · Research and enterprise licensing

