Inspect the system behind the answer.
Basic exposes project-local commands. Reasoner adds private-alpha conversational orchestration.
- BASIC · LOCAL ALPHA
- REASONER · PRIVATE ALPHA
- ROADMAP LABELLED
CURRENT + TARGET RECORDS
What a project actually contains
The architecture defines these canonical records. The local alpha implements a subset; full schema coverage, cumulative registry and every cross-record link are not claimed complete.
- ResearchBrief
- Question, scope, constraints, permissions, ethics, budgets, stop conditions, outputs and version.
- TaskSpec / TaskResult
- Bounded objective, inputs, capabilities, sandbox, budget, approval references, idempotency key, output artifacts and failure classification.
- SourceRecord / EvidenceSpan
- Stable IDs, provider, version, retrieval time, hash, access and licence status, and an exact page, section, quote or table locator.
- Hypothesis / Decision
- Proposal, alternatives, assumptions, evidence, falsifier, status, decision owner and rationale.
- RunManifest / ArtifactManifest
- Code, data, environment, command, seed, metric and resource identity plus outputs, hashes and retention.
- ClaimRecord
- Wording, type (external, empirical, interpretive, hypothesis), status, support and contradiction edges, and limitations.
- ReviewRecord / GateReport
- Rubric, evidence snapshot, model and provider version, prompt hash, findings, severity, dissent, gate result and policy version.
- ApprovalRecord / BudgetReservation
- Actor, scope, timestamp, expiry, evidence reviewed and resource limits with spent and remaining amounts.
CLI SURFACE
What you actually type
The control plane is deliberately a short-lived, project-local CLI. Commands are listed as they appear in the public HowHow_Basic README.
python -m howhow init projects/demo --goal "Evaluate a bounded, reproducible systems question"
python -m howhow status --json
python -m howhow source search --provider arxiv --query "reproducible research provenance" --limit 3
python -m howhow experiment run experiment-run.json
python -m howhow verify --profile project
python -m howhow paper finalize- start
- Begins a project session and truthfully lists the manifest-backed capability entries and pinned sources.
- brief
- Confirmation-gated. Confirmations append immutable revisions rather than rewriting the proposal.
- idea add | rank | select
- Ranking requires three to five eligible candidates and never fabricates ideas.
- claim add | audit
- Claims are recorded, then audited against their evidence links.
- experiment run
- Accepts a JSON spec with an argument array (never a shell string), a code revision and a seed.
- paper finalize
- Renders records, runs strict verification, builds the manuscript, and only then records COMPLETE.
- integration doctor | contracts | export | import
- Reads optional checkouts. A missing checkout reports AVAILABLE_CONTRACT_NOT_INSTALLED, never a failure.
EXECUTION PROFILES
Bounded execution is not a sandbox
Two profiles exist. Only one executes anything, and neither isolates hostile code.
- RECORD_ONLY
- Records the run without executing it.
- TRUSTED_LOCAL
- Trusted local execution. It copies declared project-relative inputs into a temporary directory, supplies a reduced environment with HOWHOW_SEED and PYTHONHASHSEED, caps the timeout at 300 seconds and retains at most 1 MiB per output stream.
- OS_ISOLATED
- Disabled contract. Not available.
- CONTAINER_ISOLATED
- Disabled contract. Not available.
IN DEVELOPMENT · QUALITY GATES
Thirteen defined gates, five outcomes
The target architecture defines these gates and outcomes. Complete automated coverage and independent scientific validation are not currently available.
- 1 · Scope and safety
- Question, non-goals, ethics, consent, permissions, stop conditions.
- 2 · Literature
- Stable IDs, versions, spans, access status, date cutoff, deduplication, contradictions and retractions.
- 3 · Novelty
- Nearest-work comparison and search limitations. Never an unsupported “first”.
- 4 · Design
- Hypotheses, baseline, metric, split, seeds, confounders, ablations, analysis plan.
- 5 · Execution integrity
- Sandbox, code, data, environment and command identity, budget, logs, artifacts.
- 6 · Results and statistics
- Metric integrity, uncertainty, leakage, sensitivity, handling of negative and inconclusive results.
- 7 · Claim-evidence
- Every material claim is supported, contested, or explicitly labelled hypothesis or interpretation.
- 8 · Figures and tables
- Regenerated from recorded outputs; numbers agree with claims and raw results.
- 9 · Independent review
- Separate context and evidence snapshot with reviewer, provider and model recorded. Critical dissent blocks.
- 10 · Reproducibility
- Clean rerun, or an explicit bounded exception with its limitation stated.
- 11 · Licence and privacy
- Source, data, model, code and figure rights, retention, anonymisation and attribution.
- 12 · Package
- LaTeX, bibliography, metadata, source archive, checksums and manifest.
- 13 · Human authority
- A human has reviewed the package and decides what happens next.
READINESS SEMANTICS
What READY FOR HUMAN REVIEW does not mean
The label describes the package, not the science.
- It does mean
- The package, evidence trail, limitations and review records are available for a person to inspect.
- It does not mean
- Novel, correct, statistically significant, safe to publish, peer reviewed, or submitted.
- Why the split
- Package checks and scientific correctness are separate judgements. Collapsing them into one badge is how a reader gets misled.
- What stays open
- Inconclusive is a valid research outcome. Dissent and contradictions remain visible rather than being averaged away.
LICENCE AND ACCESS
Availability does not settle permission
A reachable page may still carry restrictions, an unclear licence, or content that cannot be copied into a research artifact. Access status, licence status and the reuse decision are recorded separately per artifact.
- Per-artifact
- Rights are recorded per artifact, not inferred from the repository that produced it.
- Unknown stays unknown
- A missing version, licence or locator leaves the record incomplete rather than silently becoming evidence.
- Software licences
- A software licence does not grant rights to papers, datasets, PDFs, models or figures.
NEXT
Read the source, not just the page.
The Reasoner PRD and Basic Mode CLI source are public. The broader HowHow product remains private alpha.