HowHow is in private alpha•Request early access
HowHow

One Research OS. Two modes.

Basic controls projects locally. Reasoner coordinates the same research episode conversationally.

  • PRIVATE ALPHA
  • BASIC + REASONER
  • HUMAN-OWNED GATES
See how it worksRequest access

THE RECORDS

Six records, each with one job

These are the durable record types HowHow is being built around. Current support varies by record; absent or incomplete fields must remain explicit rather than becoming inferred truth.

01AVAILABLE IN LOCAL ALPHA · RESEARCH BRIEF

Scope before action

The interpreted question, assumptions, source boundary, evidence plan, proposed stages, permissions, risks, budget reservation, expected artifacts and unresolved ambiguity — versioned before meaningful work starts.
  • Versioned revisions
  • Plan diff required
  • Budget reserved up front

02PARTIAL · SOURCE AND EVIDENCE SPAN

A claim keeps its lineage

An external claim points to a source record and an exact span. An empirical claim points to a run manifest, a metric and an artifact. Retrieved text is treated as untrusted data.
  • Stable source IDs
  • Exact locator
  • Access and licence separate

03LOCAL ALPHA · RUN MANIFEST

Bounded execution

Code revision, data version, environment, command, seed, metric and resource limits are recorded per run. Failed, blocked and inconclusive runs are kept instead of being overwritten.
  • Immutable evaluation
  • Failure retained
  • Idempotency key

04IN DEVELOPMENT · REVIEW RECORD

Dissent stays visible

The target review record uses a frozen context and evidence snapshot and retains provider, model, prompt hash, severity and unresolved dissent. Full independent review orchestration remains in development.
  • Separate snapshots
  • Severity recorded
  • Meta-review cannot erase dissent

05IN DEVELOPMENT · GATE REPORT

Thirteen defined quality gates

The architecture defines thirteen gates and five explicit outcomes. Complete automated coverage and scientific verification are not currently claimed.
  • Limitations retained
  • Policy versioned
  • Human authority last

06FIXTURE VERIFIED · PACKAGE

Portable by construction

ClaimLedger verifies clean-directory rebuild and tamper rejection for a deterministic product fixture. That evidence does not establish full scientific-package support.
  • Clean-room rebuild
  • Checksum manifest
  • Export independent of provider

DEFAULT SURFACE

The planned cockpit shows one next action, not agent theatre.

The full review surface is in development. Current local-alpha work is CLI-first; planned views include stage, last meaningful artifact, recommended next action, spend, blockers and record counts.

Pause

No new work is created. Safe tasks finish or cancel under policy, and a checkpoint is written.

Resume

Checkpoint, lease, budget, provider readiness, source freshness and policy are revalidated first.

Take over

Autonomy stops and the brief, files and results open for editing with an actor and an event per change.

Stop and preserve

Automation and spend stop. Every event, run, artifact and failure is kept and marked inactive — never deleted.

WORKED EXAMPLE

Follow a claim back to its source

ClaimLedger is the deterministic fixture used to prove the mechanics. It is deliberately not a scientific result.

01EPISODE · CLAIMLEDGER · DETERMINISTIC FIXTURE

Can an evidence link survive a source change?

ClaimLedger verifies that changed source bytes invalidate a recorded evidence link.
  • Append-only event history
  • Source and evidence manifests
  • Success and failure experiment records
  • Claim and review records
Inspect the episode →

02WHAT IT PROVES

Product gates, not a novel method

The fixture demonstrates that a stale evidence link fails closed, that a failed run is retained, and that the package rebuilds from source in a clean directory.
  • Fails closed
  • Failure retained
  • Clean-room rebuild

MODES AND ACCESS

Two modes, plus planned team access

Basic Mode

AVAILABLE IN LOCAL ALPHA · Deterministic, project-local CLI control with durable records and no daemon requirement.
  • Project-local records
  • Short-lived commands
  • Public CLI source
Read the CLI surface →

Reasoner Mode

PRIVATE ALPHA · Conversational orchestration that can coordinate Scouts, Classes, Reviewers and optional Secondmates around the same canonical project record.
  • Dynamic crew
  • Human-owned gates
  • Basic record beneath it
Request access →

Lab

PLANNED · Proposed seven-seat access for small research teams. Shared review and organisation policy workflows remain in development.
  • Seven planned seats
  • Team workflow in development
  • Preorder interest only
See Lab pricing →

Enterprise

DISCOVERY ONLY · Custom boundary, policy and integration requirements are evaluated before any pilot or readiness claim.
  • No certification claim
  • Environment-specific review
  • Scope before terms
Talk to us →

BOUNDARIES

What the product does not claim

  • No guaranteed novelty, correctness, statistical significance or peer-review acceptance.
  • No automatic submission to arXiv. Publication is a human decision.
  • No silent fallback, silent scope drift or duplicate execution.
  • No running untrusted code with host credentials or unrestricted network access.
Evidence re-read
Exact bytes matched

VERIFIED

Failure retained
Negative control preserved

RECORDED

Novelty
Not established

UNKNOWN

Publication
Human decision required

NOT SUBMITTED

NEXT

Inspect one episode end to end.

See a deterministic product fixture, including the parts that failed. It demonstrates product mechanics, not scientific validity.

Inspect ClaimLedgerRead the docs
Product — HowHow