What LCF claims, and where it stops.
This describes essence, not incarnation — what the system must do and must remember. An earlier prototype was an incarnation; it was discarded. What follows is what survived.
Stage-level error attribution
LCF borrows its error vocabulary from the Total Survey Error tradition. The point of the split is attribution: when a verdict is wrong, the framework must be able to say where it went wrong — in the shapes, in the kernel that applies them, or in the accept/reject decision relative to ground truth. Collapsing these into one number destroys that diagnostic power.
Specification error
The shapes and rules are themselves poorly written or incomplete — the wrong constraints are enforced, even if enforced flawlessly.
Mitigated by: adversarial review of shapes before they enter the substrate
Measurement error
The specification is correct, but the system applies it incorrectly — an evaluator bug, a parsing fault, a mishandled edge case.
Mitigated by: testing against deterministic mocks; recomputable verdicts
Decision error
The verdict is wrong relative to ground truth — Type I false reject and Type II false accept. Observable only in the Lab, never from the pipeline in isolation.
Mitigated by: measured, not asserted — the paired-arm harness
A measurement error may or may not produce a decision error. A misapplied rule can still land on the correct verdict; a correctly-applied rule on top of a bad specification can produce the wrong one. The gap between the two layers is itself a reported diagnostic — it tells the steward whether to chase a wrong verdict into the kernel or into the shapes.
Principle index
Instrument the seam; don't trust it.
Meaning to form is probabilistic and permanent. Everything else is in service of not trusting that crossing.
No loss is silent.
Residue is mandatory. Content fidelity and structural validity are independent axes — an extraction that drops an "unless" without flagging it is worse than one that flags it.
Kernel judges claims; the policy engine enforces the world.
The kernel can check that a record asserts a review relationship. It cannot know whether the review board actually convened. Keep the line clean.
Time is bitemporal.
Transaction time (when the system knew a rule) and valid time (when it is in force). Grandfathering is two shapes with non-overlapping windows — decided by a triple in the graph, never by date logic in the kernel.
Wrap, don't mutate.
The core ontology stays pristine; extensions import and add. When an extension needs a noun the core already has, reference the core's class — never shadow it. Separate on disk, unified in the loaded model.
The Governor adjudicates the gold.
Author input blind to the desired verdict. The LLM is a proposer, not a gold author — it makes the exact seam errors LCF exists to distrust. Record how every row's gold was established.
What LCF does not claim
LCF does not solve the seam, and does not establish that adjudicated reference classifications are infallible.
It does not claim that demonstrated error rates generalize beyond the studied corpus. Reliability across domains, models, prompts and independent adopters remains an empirical question.
Deterministic results are statements about the kernel and the shapes. They do not measure real-extraction fidelity — and no document may cite fidelity until a live arm is run and published.