PortentRequest access
How it works

Tool · Agent · Repo

Three parts, one behavior: the language model formats, the tool decides, the repo remembers. It is why the same input always produces the same output.

The Tool

deterministic

The model formats. The tool decides.

Pure deterministic functions do the deciding: sizing math, failure classification, plan synthesis. A compiled pattern table maps a log signature to a codified fix. Same input → byte-identical output, every time.

  • No formula or benchmark number is ever invented by the language model.
  • A compiled detect → fix pattern table (109 patterns) turns a known signature into a known remedy.
  • Determinism is tested: identical inputs are re-run and byte-compared in CI.

The Agent

orchestrator

Collects inputs, routes to tools, knows when to stop.

The agent is the orchestrator. It gathers the inputs a tool needs, routes work to the right tool, parallelizes independent steps, and stops for a human at every irreversible or costly step.

  • Autonomous where it's safe; human-gated where it isn't.
  • Runs the bring-up sequence in dependency order so a step never precedes its prerequisite.
  • Prepares the risky action and hands you exactly one command.

The Repo

version-controlled corpus

Not a prompt. A corpus.

Every fix comes from a git-backed, coverage-ratcheted knowledge base, not from a prompt. Each lesson is retrievable, actionable, and attributed to the real event that produced it.

  • 342 canonical lessons, each traced to a source event.
  • A coverage ratchet keeps every lesson retrievable, actionable, and attributed; the bar never lowers.
  • Unattributed folklore is rejected at the gate.
The closed learning loop

Every run makes the next one smarter

Success or failure, each run is mined for lessons, categorized on four axes, scrubbed of anything sensitive, and verified against a zero-false-positive gate. Only after a human approves is it promoted into the canonical corpus, where it's recalled on the next run.

closed loopcoverage ratchet01Mine02Categori…03Scrub04Verify05Human-ga…06Promote07Recall
  1. 01Mineevery run, pass or fail
  2. 02Categorizelayer × substrate × framework × subject
  3. 03Scrubstrip IDs, IPs, hostnames, keys
  4. 04Verifyzero-false-positive gate
  5. 05Human-gatea person approves the promote
  6. 06Promoteinto the canonical corpus
  7. 07Recallfeeds the next run
Lanes

The agent runs in three modes

The same agent at three levels of authority: read-only recall, a plan you review, and a fully orchestrated bring-up that stops for you at every irreversible step.

Answer a question

Read-only recall over the corpus. “What's the known EFA misconfig list?” “Which NCCL algorithm did we settle on at 32 nodes?” It retrieves attributed lessons and does not touch your infrastructure.

input: a question → output: sourced answers
no side effects

Plan capacity

Given a goal, it synthesizes a plan: nodes, fabric, framework, and the ordered bring-up sequence, with the deterministic sizing tool doing the math. You get a plan to review before anything runs.

input: a goal → output: a reviewable plan
no side effects until you approve

Orchestrate a bring-up

Executes the plan layer by layer, verifying each before the next. Every irreversible or billable step pauses and hands you one command. Ends with a provenance-backed report.

input: an approved plan → output: a verified cluster + report
human-gated at every irreversible step

The model formats; the tool decides. In every lane, no number is guessed and no irreversible step runs unattended.