Install Is Admission: truc, a Package Manager and Git-Backed Registry Where Search Reuses the Admission Predicate

Pitch

A package manager has no install step to secure because fetched source enters a composition only through the same compiler gate that checks your own code, and the registry's search is that gate run as a filter.

Abstract

Package managers fetch first and discover incompatibility or widened authority later, and they defend the supply chain with signatures and provenance. We describe truc, the component manager of the revl language, and its git-backed registry, built on a different premise: a fetched component is admitted through the compiler's own gate before it joins the assembly, so there is no separate install step. A registry query is the existing admission predicate run as a filter over candidates, with no new compatibility logic. resolve ranks the compatible set by declared authority (unbounded flag, capability count, emission count), then interface fit and direct over adapted, then evidence quality, where evidence earns rank only when something outside the publisher vouches for it (a valid attestation or a trusted publisher); an index row that disagrees with the entry's own source is refused, not demoted. When a candidate is refused, resolve can propose a safe adapter, never wire it. Human intent lives in truc.toml, proof in truc.lock, and truc reproduce rebuilds a published component and compares recomputed hashes tier by tier (source, dependency lock, IR, policy surface, backend version, independent pin, attestation, emitted artifact), with exit 0 when no tier mismatches, 1 on a MISMATCH (under --strict, also when a tier could not verify), and 2 on a usage or resolution error; cannot-verify is a tier outcome, not an exit code. truc is itself a revl composition and assembles itself as a regenerate-or-red fixpoint. We report the design, the 5-component seed registry, and test evidence; we report no user study, no large-registry measurement and no reuse-rate result.

Problem and claim

Existing managers split two questions: is this the artifact the publisher published (lockfile hashes, checksum databases, signatures, provenance) and may this artifact do what it will do (usually left to the runtime or to review). A pinned, signed dependency can still widen authority, or can collide with something already in the assembly, and in the managers this paper compares against the failure surfaces at run time (to be checked).

The claim: if components declare their authority and the compiler already decides whether a component fits a composition, then the package manager can delegate both questions to that decision. Install is admission (docs/registry.md:450). Search is admission as a filter (docs/registry-probe.md:1-20). Ranking prefers the candidate with less declared authority, and the lock row stores the capability set and emission count the assembly admitted, so a lockfile diff shows authority growth in review (docs/design/truc-architecture.md:295-330; src/revl/truc/_host.py:648-649).

The paper's scope is the ordinary path on main today. It does not claim a trust graph, a marketplace or a hosted registry. Those are designed or proposed and are listed under limits.

Contributions

1. Admit-at-add and admit-at-assemble: truc add plans, admits through compile_files(..., manifest=running) in process and vendors only on success; assemble re-admits and refuses on vendor hash drift (src/revl/truc/_host.py:333-375 (admit_all) and src/revl/truc/_host.py:581-650 (commit_add); drift check via plan.drift in src/revl/truc/components/assembler.rvl, fn assemble; tests/test_truc_add_assemble.py:118, tests/test_truc_add_assemble.py:136, tests/test_truc_add_assemble.py:154).
2. Search as admission probe: resolve calls the same _service_compatible predicate and a test compares its include/exclude decision to the real hot-swap gate on every seed Database provider (src/revl/registry.py:1808-1830; tests/test_registry_resolve.py:186, which runs over the seed Database providers; tests/test_search_as_admission.py:169 covers only the examples/ corpus).
3. Least-authority, evidence-aware ranking with verification: authority and fit first, then evidence quality; an index row that contradicts the entry's source is refused; fabricated evidence cannot outrank an honest entry (src/revl/registry.py:1766, src/revl/registry.py:2120; tests/test_registry_index_trust.py:113, tests/test_registry_index_trust.py:143, tests/test_registry_index_trust.py:269, tests/test_registry_index_trust.py:300; tests/test_registry_evidence.py:187, tests/test_registry_evidence.py:210, tests/test_registry_evidence.py:244).
4. Proposed-not-wired adapters: a refused candidate can come back as compatible-with-adapter, ranked below direct matches and with error-semantics evidence discounted when the plan merges outcomes (src/revl/registry.py:1869, src/revl/registry.py:2120; tests/test_296_resolve_adapter_ranking.py:203, tests/test_296_resolve_adapter_ranking.py:225, tests/test_296_resolve_adapter_ranking.py:312, tests/test_296_resolve_adapter_ranking.py:482).
5. Intent in truc.toml, proof in truc.lock, and truc reproduce: tier-by-tier recomputation with OK, MISMATCH and cannot-verify, exit 0/1/2, and a --strict mode where an unverifiable tier is not a pass (src/revl/truc/reproduce.py:1-70, src/revl/truc/reproduce.py:848-899; tests/test_truc_reproduce.py, 42 test functions).
6. truc assembles truc: its eight components are admitted by its own gate on every boot from a hash-pinned stage-0 lock, and a golden composition makes drift red (src/revl/truc/truc.lock; tests/test_truc_bootstrap.py:90, tests/test_truc_bootstrap.py:118, tests/test_truc_bootstrap.py:147).

Related work

Evidence on main

Numbers that exist

• Seed registry: 5 components (registry/components/: audited_database, mysql_database, pg_database, readonly_database, user_cache). Sample size for any claim about ranking is therefore tiny.
• Probe: 5 candidate providers, 3 admitted and 2 refused, per docs/registry-probe.md (one consumer, one key). Reproduced by tests/test_search_as_admission.py:169, which runs on the examples/user_cache.rvl corpus, not the seed registry.
• Test function counts by file (name-pattern count, not run): truc_reproduce 42, registry_releases 30, truc_name_jail 23, bundle 32, adapter_predicate 26, resolve_adapter_ranking 17, truc_bootstrap 17, registry_resolve 11, registry_index_trust 11, truc_apply_stack_check 13, truc_ship 13, truc_add_assemble 7, registry_evidence 6.
• Roadmap archive rows quote 17 tests for truc reproduce at landing and 15 for revl bundle (docs/roadmap-archive/12-post-v2-0-horizon-strategy-review.md:605, docs/roadmap-archive/12-post-v2-0-horizon-strategy-review.md:621); the file now has more.
• No reuse rate, no token cost per path (the registry doc names both as planned measurements, docs/registry.md:461-470), no latency table for resolve on a large registry. None yet.

Paper outline

1. Motivation: two questions package managers conflate, and the run-time failure they leave.
2. The premise: admission as the single decision; truc is a composition that calls the gate.
3. Walkthrough: add (accepted, refused G2), assemble, drift refusal.
4. Resolve: search as admission, ranking, evidence rules, refused index rows, adapter proposals.
5. Reproduce: eight tiers, three outcomes, strict mode, the independent pin.
6. Dogfood: truc assembles truc; regenerate-or-red.
7. Evaluation (to be written): reuse rate, attack cases, resolve latency.
8. Limits and related work.

Threats to validity and limits

• The seed registry has 5 entries and no attestation; the ranking test says mysql and pg tie on authority, fit and evidence and are separated by source length and name (tests/test_registry_resolve.py:99-112). Evidence ranking is exercised mostly by synthetic registries in tests/test_registry_evidence.py.
• The registry doc states that an entry's extern classifications are trusted, not verified (G8 caveat, docs/registry.md:456-459). Admission checks structure and declared authority; it does not check that a declared effect class is honest. Every response carries this in assumptions.
• The stance "the registry needs no trust beyond availability" is stated in the doc. The same doc declines authenticated publisher identity, so continuity checks compare self-asserted labels (docs/registry.md:509-512). Publisher trust is a caller-supplied set.
• Resolve reads the current release only; frozen history is stored but is not a selectable set.
• The registry is a git repository on disk. Nothing here shows behavior with many authors, name squatting or a hosted index.
• truc reproduce rebuilds with the current toolchain; a toolchain drift is reported as a mismatch or cannot-verify, which is honest but means a user must keep the toolchain in step.
• The trust graph and the verified-adapter marketplace appear only as external proposals in the roadmap archive (docs/roadmap-archive/12-post-v2-0-horizon-strategy-review.md:593, docs/roadmap-archive/12-post-v2-0-horizon-strategy-review.md:603). A reader will compare the abstract with the status tags, so the abstract is kept to the shipped path.
• The lock makes authority growth reviewable; it does not prevent a wider candidate from being chosen.

Fact-check

Claims in this draft were checked against revl main at commit 45bec431f: 60 checked, 11 corrected.

Cite this draft

@misc{revlp35,
  author       = {Thomas Moussajee},
  title        = {Install Is Admission: truc, a Package Manager and Git-Backed Registry Where Search Reuses the Admission Predicate},
  year         = {2026},
  howpublished = {\url{https://inso1337.github.io/revl/research/P35.html}},
  note         = {Research draft, not peer reviewed. Written against revl commit 45bec431f. Licensed CC BY 4.0.}
}

Text licensed CC BY 4.0, author Thomas Moussajee.