Skip to content

Latest commit

 

History

History
94 lines (65 loc) · 8.03 KB

File metadata and controls

94 lines (65 loc) · 8.03 KB

M0 Contract and Test Boundary

Status: MEASURED CONTRACT CONFORMANCE / NO RUNTIME / EFFICACY UNJUDGED

Measurement date: 2026-08-08

What kind of theory is this?

FLR-H is currently an engineering research programme with an executable formal-design layer. It is not yet a demonstrated programming paradigm, a production runtime, or an industry standard.

Four claim levels must remain separate:

Level Question Current state
Formal design Do the current F/L/R/H artifacts satisfy their closed schemas, encoded invariants, and declared fixtures? M0 checker conformance measured
Mechanism Can independent code implement those contracts deterministically, incrementally, durably, and safely? bounded standalone M1 pure-F, M2 ground stratified-L, and M3 scalar-frontier-R references measured; persistent incremental maintenance, direct M2-to-M3 integration, M4–M5, and the integrated runtime are not implemented
Comparative efficacy Does full FLR-H improve held-out coding-agent outcomes over a simpler equal-budget control? Unpreregistered and unjudged
Paradigm/standard Is the abstraction independently adopted, stable across implementations, and worth teaching as a general paradigm? Unsupported

M0 therefore records checker-relative contract conformance. It does not prove global logical consistency, implementation feasibility, or practical usefulness.

M1, M2, and M3 subsequently added three bounded sibling implementation results without changing this historical M0 boundary. See M1_KERNEL.md, M2_LOGIC.md, and M3_REACTIVE.md for their implemented seams and narrower claims.

Lakatos framing

The dedicated LakatoTree programme currently carries this hard core:

  1. execution closure, not a tool-call trace or model assertion, defines completed work;
  2. Harness control is intrinsic rather than an optional wrapper;
  3. functional, logic, and reactive mechanisms are materials for a machine-checkable execution graph;
  4. no efficacy claim is admissible without an equal-compute comparison.

The v0 choices in this repository—four-valued support, stratified negation, scalar epochs, signed deltas, typed graph federation, and intent-before-effect—form a replaceable protective belt. A failed v0 choice may be repaired without silently changing the hard core. Repeated ad hoc repair, however, must be reported as degeneration rather than hidden.

Normative M0 artifacts

Diagrams and prose are explanatory views. If they disagree with a normative machine file, the machine file wins until a versioned correction is accepted.

M0 checks

Run:

python3 -m pip install -r requirements-m0.txt
python3 scripts/validate_m0.py
python3 -m unittest discover -s tests -v

The current checked bundle checks:

  • nine Draft 2020-12 schemas, including closed bindings for every normative instance;
  • ten valid protocol object kinds and thirteen single-invariant negative mutations, including rejection of F-owned authoritative risk/effect-idempotency fields and inconsistent approval bindings;
  • all four evidence states and all four default-negation outcomes;
  • multiple-support retention, last-support removal, conflict preservation, stratification rejection, and strict frontier passage;
  • canonical equivalence, ordered nested-payload separation, duplicate-set rejection, mutation separation, floating-point rejection, NFC rejection, and signed-int64 bounds;
  • 25 FSM transitions, four false-guard branches, schema/authority rejection, nine terminal paths, three immediate/deferred interrupt classes, and 18 traces;
  • one digest-bound lifecycle authority plus bounded budgets, durable intent, approval expiry, receipt re-ingestion, unknown-outcome terminal closure, no-progress, and independent success closure.

The Python checker is non-normative and replaceable. A second implementation should be able to consume the same JSON contracts and fixtures.

Test ladder beyond M0

Milestone Decisive test What would falsify or downgrade it?
M1 Functional ✅ replay/property tests in two clean processes measured for the frozen reference profile; any byte mismatch or ambient attempt reopens it
M2 Logic ✅ stepwise insert/retract ledger versus independent clean recomputation measured for the frozen ground profile; any support, conflict, provenance, rejection, or canonical-byte divergence reopens it
M3 Reactive ✅ randomized order, frontier, slow-consumer, backpressure, guard, and semantic-mutation tests measured for the frozen supplied-value scalar-time profile; any oracle, byte, bound, ambient, mutation, or authority-boundary divergence reopens it
M4 Harness crash injection around intent/send/receipt plus cancellation and fencing duplicate confirmed mutation or false SUCCEEDED
M5 Graph round-trip, migration, resolver, and composition-mutation tests required identity is lost or forbidden wiring passes
M6 Interop independent conformance runner and two real consumers contract cannot survive independent implementation
M7 Efficacy preregistered held-out equal-budget A/B with independent judge no material gain, regression, or cost ceiling violation

The full FLR-H comparison must be preregistered before measurement. A null or negative result is valid evidence and must downgrade the efficacy claim.

LakatoTree exact-readback boundary

Read-only verification found the dedicated LakatosTree_AgentParadigm_20260808 tree, incarnation 93264ef9e2734621a24b7d48d46ebfc2, at notebook assurance with partial coverage and no scientific judgment. The narrower E1 execution-closure prediction has receipt 1d714918c729f1d7d44a89294f7345b7d0a77c27b61cab07910c2c2554fd399f.

That receipt is not a full FLR-H efficacy preregistration. The standalone repository commit is also not yet bound to a LakatoTree node. LAKATOTREE_BINDING_2026-08-08.json preserves the exact readback without pretending that a local pointer mutates the tree.

No progress verdict is issued by M0. Only an independent registered judge may issue one after prospective evidence exists.

Validator portability note

The bundled fsm-design tools target their own fsm-spec/v1 dialect; this repository declares flrh-run-fsm/v1 with a full Draft 2020-12 event schema and typed actor/capability authority. The bundled loop-engineering validator expects an embedded second FSM plus the SYMPOSIUM seven-commander canon, including Jaebaeman. Both read-only invocations therefore fail on profile shape and are not counted as M0 passes. Importing duplicate lifecycle authority or unrelated planning mythology would violate this repository's contract. M0 uses a repository-local replaceable checker, while M6 explicitly requires an independent conformance implementation. Exact incompatibility readbacks are recorded in FSM_VALIDATOR_COMPATIBILITY_2026-08-08.md and LOOP_VALIDATOR_COMPATIBILITY_2026-08-08.md.