01
Discovery
query → retrieve → deduplicate and version → read → extract spans → citation and contradiction expansion → coverage audit
From question to review.
Five visible stages connect scope, evidence, runs, failures and human review.
- INTAKE → EVIDENCE → RUN → REVIEW
- PARTIAL WORKFLOW
- FAIL-CLOSED
TARGET EPISODE MODEL
Five stages, each with an exit condition
This is the workflow HowHow is implementing. Current local-alpha coverage is partial; a stage should advance only when its required durable record exists.
- 01INTAKE
Brief
Bound the question, define non-goals, and reserve permissions, budget and stop conditions before work starts.
- Question
- Scope
- Policy
- 02LITERATURE
Gather
Retrieve permitted sources, preserve versions and access status, and extract exact evidence spans with contradictions.
- Sources
- Spans
- Contradictions
- 03EXPERIMENTING
Test
Compare candidates, reproduce a baseline where allowed, and run only a locked, bounded intervention plan.
- Baseline
- Evaluation
- Artifacts
- 04REVIEW
Review
Map claims to evidence, inspect failures and dissent, and revise anything that outruns its support.
- Claims
- Critique
- Limitations
- 05READY FOR HUMAN REVIEW
Handoff
Package the record for rebuild and human judgement. Ready-to-review is not the same as novel or accepted.
- Manifest
- Checks
- Decision
BOUNDED LOOPS
Research loops back, so the loops are capped
Discovery can return after critique, a reviewer can reopen a literature query, and weak evidence can send work back to design. Each loop carries its own ceiling.
02
Idea and novelty
gap → candidates → nearest-work comparison → feasibility, risk, cost and falsifier → direction approval
03
Experiment
design → implementation → execution-boundary review → baseline → run → metric validation → analysis → diagnosis → ablation or reproduction
04
Evidence
claim → support or contradiction → provenance audit → confidence and limitations → revise the claim or open a task
05
Writing
outline → claim-linked draft → citation, number and figure audit → review → revision
06
Governance
gate → pass, pass-with-limitations, revise, block, reject or human escalation
PLANNED LOOP LIMITS
Every ceiling should be declared before the loop starts
Iterations
Maximum iterations per loop before the work pauses or waits for a person.
Wall time
Declared before the loop starts, not discovered afterwards.
Compute budget
API, GPU and storage reserved up front with a hard stop.
Retries
Automatic only for idempotent, retryable failures. Unknown side effects block.
Concurrency
Bounded; serial degradation keeps every gate and record intact.
No-progress detector
Repeated equivalent actions without new evidence pause the loop.
STATUS MODEL
Twenty explicit states, plain language
The architecture defines these states so failure and uncertainty do not become success. Coverage varies in the private alpha; INCONCLUSIVE remains a valid research outcome.
- INTAKE
- BRIEFING
- SCOPING
- LITERATURE
- CANDIDATES
- WAITING_FOR_HUMAN
- BASELINE
- EXPERIMENTING
- ANALYZING
- WRITING
- REVIEW
- REPRODUCIBILITY
- PACKAGING
- READY FOR HUMAN REVIEW
- PAUSED
- BLOCKED
- FAILED
- INCONCLUSIVE
- CANCELLED
- ARCHIVED
CONTROL MODEL
Three controls HowHow is designed to preserve
Pause
No new work is created. Safe tasks finish or cancel under policy, and a checkpoint is written.
Take over
Autonomy stops and the brief, files and results open for editing. Every change carries an actor and an event.
Stop and preserve
Automation and spend stop. Every event, run, artifact and failure is kept and marked inactive — never deleted.
NEXT
See the record a reviewer receives.
The ClaimLedger fixture shows the verified product trail, including the deliberate failure. It is not a scientific episode.