Evidence
Anyone can put a big number on a page. Continuum's claim is that every number has a file behind it, because the artifacts are the mechanism. The ledger, the scoreboard, the paper trail, one run end to end, and the software the line has shipped.
The proof is a file you can open.
Anyone can put a big number on a landing page. Continuum's claim is stronger and easier to check: every task run since 2025 (own products, a real enterprise engagement, the studio behind it) left artifacts behind, because the artifacts are the mechanism. Verify progress with nothing more exotic than reading a directory.
The ledger: cost is recorded, not estimated
# task.yaml · the execution ledger, one per task
task: auth-token-rotation
mode: DONE branch: task/auth-token-rotation
sessions:
- mode: EXPLORING turns: 18 cost_usd: 0.24
- mode: RESEARCHING turns: 11 cost_usd: 0.14
- mode: PLANNING turns: 24 cost_usd: 0.58
- mode: IMPLEMENTING turns: 31 cost_usd: 0.72
- mode: VALIDATING turns: 12 cost_usd: 0.30
# a full task, end to end: ~$2. where time and money go,
# per mode, is a question with an answer.Which mode is expensive, whether a task was badly sized, where the week actually went: answered from the ledger, not from memory. A supervisor can be wrong about their own system and find out.
The scoreboard: done is read off the remote
$ regent scoreboard idea-unify-auth
remote: origin/main · sha 91e74f7 · read 18:04Z
auth-shared-middleware delivered
auth-token-rotation delivered
auth-docs-runbook delivered-something-else
checked 8 of 9 · 1 could not be read
(a gap, not a zero)Delivery verdicts are computed against the remote ref, never the local checkout, headed by the sha and when it was read. A task that publishes "done" against the evidence gets flagged, and an unreadable check reports as a gap instead of quietly passing.
The record: a byproduct, not a chore
tasks/auth-token-rotation/
├── brief.md what was asked, exactly
├── exploration.md the gate that opened RESEARCHING
├── research.md …that opened PLANNING
├── plan.md · todo.md …that opened IMPLEMENTING
├── validation.md what "done" was defined to mean
└── sessions/*.jsonl every session: turns, cost, eventsBy the time a task finishes it holds a written account of its own reasoning. None of it was produced as documentation; each file existed because it was the gate that let the next mode begin.
One run, end to end
One captured signal, a few lines long, went in. Out came 17 findings, 9 decisions, 22 executable briefs and 22 GitHub issues; the refinery did the expanding. A typical run completes without the human touching it at all: regent arbitrates the routine with a recorded rationale, and escalates only a scope call, a fact only you know, or an outward-facing effect.
"This abstraction let me abstract from every detail and participate only where I was actually needed."
One line, very different software.
The usual objection to a dogfooded tool is that it only works on its creator's repo. The ledger says otherwise: a Rust search engine, desktop ecosystems, a semantic engine, and the sites that argue for them, all built or hardened through the same capture-to-verified line.
A career-navigation ecosystem: desktop app, 27 CLI tools, a conversational agent.
read the work →A semantic thesaurus engine and its tooling.
read the work →An embedded search engine in Rust, and the site that argues for it.
read the work →A clinical protocol turned into an inhabitable space, not a document.
read the work →Cognitive offload: closing the open loops that tax a mind in the background.
read the work →Fluency by rhythm: a wrong answer is cut while it is still being spoken.
read the work →Continuum shipped itself, 93 tasks and counting. It also shipped this page.
The public works are argued, one thesis each, at tiphareth.com.br.