Quality and grounding
The form quality gate (A-F grade) was removed in c754. SEIF's verification value now comes from provenance and §7 enforcement — not form-based scoring.
Form quality gate removed (c754)
The form quality gate — the seif quality-gate verb, gate quality subcommand, quality_gate
MCP tool, and the engine /v1/quality-gate endpoint — was removed. It measured the
linguistic form of text (hedging markers, stance signals) and returned a letter grade
A–F. It did not measure truth, correctness, or whether work was actually verified.
Removing it was deliberate: form-matching is not a verification signal, and calling it one was misleading. A confidently-worded wrong answer scored well; a correct, careful answer could score poorly. The grade optimized for the appearance of grounding, not grounding itself.
What SEIF verifies instead
SEIF's verification value comes from two things:
Provenance — every artifact is signed locally (Ed25519), assigned a canonical hash, and submitted to the transparency log. This creates a tamper-evident chain of custody: who produced it, when, from what inputs, in what session. The chain is independently verifiable by anyone with the evidence URL.
seif sign ./report.md # sign + get an evidence frame
seif verify <evidence-url> # independently verify the chain
seif publish ./report.md # sign → log → get a permanent verify URL§7 enforcement — delivery is verified against the norte via constrained_by links and
coherence checks. Work that is not traceable to a ratified commitment does not ship.
This is governance, not style: it checks that output is grounded in an agreed intent,
not that it uses the right sentence structure.
Grounding is conduct, not numerology
Grounding in SEIF is a way of working: verify before you assert, take one considered pass and settle, keep part of your attention on whether the method is even right, intervene minimally, and never publish without explicit human approval.
That conduct does not reduce to a form score. It is expressed through the surrounding infrastructure: memory that persists decisions, provenance that makes work replayable, and governance that ties output to ratified commitments.
Read the receipts
The full pre-registered experiment that showed form grading was neutral at best — and the positive result (compact capability map helps a cold AI reach the right verb) that did survive it — are in Honest measurement and Just-in-time expertise.