Reading the output

Terraform can be red/green because its semantics are binary. Groundhold's are not — the whole system refuses to collapse "I don't know" into a boolean — so the output speaks a slightly richer, closed visual vocabulary. You can read it at a glance without understanding the system; this page is the whole glossary.

The banner: one word, always last

Every verb ends its human output with one word. It answers the only question you have at a glance: can I walk away?

banner	color	it means	do
PROVEN	green	every hard constraint satisfied under accepted evidence	—
CONVERGED	green	reconciliation reached its proven fixed point	—
APPLIED	green	the sealed plan's mutations committed	—
SEALED	green	a hash-pinned plan exists and is executable	—
OK	green	the procedural verb completed	—
VIOLATED	red	something is proven false	fix the world or the contract
INVALID	red	the input document does not parse/validate	fix the document
BLOCKED	yellow	missing knowledge on a hard constraint — not a failure	gather evidence: probe, observe
STALE	yellow	the world moved since the decision was made	re-observe, re-plan
REFUSED <code>	blue	a gate held, deliberately	follow the code (groundhold explain <code>)
DIED	red	apply failed mid-flight	groundhold resume
CORRUPTED	red inverse	the ledger is damaged	groundhold repair; nothing proceeds

Two deliberate oddities, both load-bearing:

• A refusal is blue, never red. The gate holding is the system working. If refusals looked like errors, you would learn to resent and bypass the gate — so they look like a guard, not an accident.
• A non-green banner always names its culprit: BLOCKED: c-rto unknown — recovery.rto requires probe verification, never the bare word.

Machine consumers ignore banners entirely and route on exit codes and the JSON code field — banners are explicitly not a machine interface. Verbs whose stdout is machine output (JSON, ndjson, plans) print the banner on stderr instead, one final line.

Verdict rows: shape first, color second

Everything stays readable with color stripped (CI logs, NO_COLOR, --color=never, color-blindness) because the SHAPE carries the meaning:

glyph	ascii	verdict	next action
✓	OK	satisfied — proven true	—
✗	X	violated — proven false	fix the world or the contract
?	?	unknown — not known yet	go look: probe, observe
∅	NA	unverifiable — not knowable this way	change the contract or the method

? and ∅ are different shapes because their remedies differ: unknown is fixed by gathering evidence, unverifiable only by changing the question. The ASCII set appears automatically under non-UTF-8 locales or --ascii.

Provenance renders as brightness. A verdict resting on inferred/assumed values keeps its glyph and color but renders dim, with the [inferred] marker: green-but-dim means satisfied, standing on sand.

The output teaches as it blocks

When a constraint is in a friction state (violated, unknown, unverifiable), the row gains one indented line straight from the vocabulary — what the attribute means and why configuration alone may not prove it:

? c-rto   unknown   requires probe verification; not evaluable from the candidate alone
    recovery.rto — time to restore service after failure; probe-only —
    a value here is a claim until a restore test measures it

Satisfied rows stay terse — the happy path never lectures. And every noun the system emits has one place to ask about it:

groundhold explain recovery.rto        # vocabulary attribute
groundhold explain consent-required    # machine error code

Freshness is part of the claim

"Converged" is always "converged as of some observation". Human renderings carry the age of the evidence they rest on, and a downstream management console shows it on every green badge — a green state is only as fresh as its stalest proof.