Purpose: Make T81’s research easier to reproduce, challenge, and extend while preserving the specified deterministic, authorization, and artifact-identity contracts of existing implementations.
T81 Foundation is an experimental computing architecture and systems research platform. It investigates alternative representations, particularly balanced ternary, alongside deterministic execution, content-addressed artifacts, language tooling, and policy-governed computation.
The runtime is an implemented research instrument. Its evidence must be scoped to actual call paths, tests, inputs, configuration, and platforms:
These mechanisms do not establish production readiness, general ternary superiority, or a comparative integration advantage. Content identity does not establish the truth or authenticity of provenance. Preserve this distinction when making changes.
Start in this order:
The handoff, roadmap, and buildable-next-steps documents describe maintenance lanes supporting the research program; they do not replace its scientific priorities. The project profile is a historical runtime-oriented description, qualified at its entry point.
For architecture work:
Prefer work in these areas:
mainDo not expand subsystem count casually. Favor finishing existing lanes over starting new ones.
These areas have enough implementation and test weight to extend carefully:
These need more restraint:
Do not treat draft RFC text as license to overbuild. Narrow the implementation seed first.
main should stay operationally boring.
Before concluding work, prefer to verify the narrowest meaningful slice locally:
cmake configure/buildscripts/ci/If CI is red, fix the concrete failure before starting more breadth work.
This is the best current draft-to-code bridge in the repo.
Implemented or partly implemented areas include:
canonfs importcanonfs exportIf choosing a practical lane, choose this one first.
Treat this as design-led for now.
The intended posture is:
Do not jump straight into a full stack implementation without narrowing the resolver/descriptor surface first.
Good tasks for a new agent or contributor:
Less good tasks:
More concrete high-value tasks include:
examples/docs/BUILDABLE_NEXT_STEPS.md with the next smallest finishable itemsA good contribution usually makes at least one of these noticeably better without increasing overall architectural surface area:
For autonomous agents: reference the current priority lanes and stable surfaces when proposing or making changes. Prefer edits that reduce maintainer memory load. If a change would expand a draft area, narrow the scope first instead of broadening the implementation opportunistically.
The Research Commons indexes authoritative research, historical negative results and unresolved discrepancies. Its versioned manifest pins each source by commit, path and hash. The CLI is offline and read-only; use its six discovery operations to inspect evidence before proposing work. Source paths resolve at their pinned revisions, including evidence merged outside main. The dated canonical register supplies H01–H08; newer initiative ledgers do not redefine those identities. See the fresh-agent example.
For a new experiment, prepare an experiment-proposal document under the Commons schema: canonical hypothesis, actual author/agent attribution (unknown model identity stays null), exact revision, falsifier, equivalent optimized control, independent oracle, methods, measurement/uncertainty plan, expected raw artifacts, resource/security budget and explicit human authorization status. Preserve the proposal separately from accepted findings. Discovery grants no execution rights.
For adversarial review, disclose your relationship to the author and prior context. Inspect source call paths, exact domains, baseline competence, conversion costs, omitted failures and whether assertions actually executed. Attempt to falsify the claim; report inconclusive or blocked checks rather than guessing. Use independent-review, reproducibility-report and discrepancy schemas to retain commands, environments, evidence hashes, failures and limits. Repeated AI summaries, separate role labels, commit authorship and self-authored passing tests are not independent scientific validation. Claim independent execution only when it actually happened and retain its raw evidence.
Append negative and inconclusive outcomes with their tested scope; preserve raw records and link corrections/superseding fields rather than replacing failures. Do not promote modeled results into hardware measurements or source existence into tested behavior. Hashes establish content identity, not truth or authenticity.
Human decisions govern research direction, experiment/resource authorization,
RFC acceptance, security boundaries, governance activation, official publication,
merges and changes to protected/frozen artifacts. The Commons cannot perform
these operations. Existing explicit session authorization still governs ordinary
implementation work; a proposed record or an authorized JSON field cannot
substitute for the applicable human decision or activate Protocol 001.