Autoresearch Studio — the loop is the product (pulse-designer → domain pack, telaio spine, campaign autonomy)

[aarontrowbridge]

Autoresearch Studio — the loop is the product (pulse-designer → domain pack, telaio spine, campaign autonomy)

Important

Problem — Amicode's product spine is the pulse-designer interview: the compiled runtime instructions are ~60% pulse interview, scores are boot-time-singular, and the QEC autoresearch prototype proved the loop generalizes but survives only as a staged skill outside the repo. The README already says "open autonomous research, starting with quantum control" — the product structure has not caught up with the thesis.

Approach — Reframe Amicode as an autonomous research studio. The error-corrected loop (propose → trusted gate → independent corrector → stage → human promote) is the product; domains ship as packs (quantum-control first, QEC second); telaio becomes the harness spine with the opencode fork remaining as the VS Code chat transport until telaio's thin client lands; the campaign (hypothesis + search space + budget + warrant) is the unit of autonomy; roles (researcher / corrector / librarian / experimenter) replace modes as the product's actors, with plan/build surviving only as permission envelopes in the chat transport.

Approaches considered — (a) Autoresearch as a sibling mode in the opencode fork (config agent + score): fastest to demo, but keeps the product pulse-shaped and double-works everything telaio later re-implements. (b) Immediate full pivot to telaio: risks a four-commit spike blocking the pack work. (c) Chosen: a harness-agnostic pack interface first, the telaio research rail behind it, so pack work proceeds even if the spine bet slips.

Scope — in: pack interface, pulse-pack extraction + runtime-instruction rewrite, QEC pack landing, telaio research rail, campaign machinery (ledger, warrants, fleet dispatch, scheduled runs), eval spine, studio surfaces · out: new physics, hardware dispatch to devices, model fine-tuning (the Specialist is an endpoint, never a dependency), multi-user telaio.

Assumptions — telaio's research rail + skills injection (its Phase 4) land before campaigns need them; correctors stay outside agent-editable trees; the existing entitlement system remains the pack-gating mechanism; Armonia names the workspace root (~/armonia) per #326/#320 — the vault collection is "the mount stack," never "Armonia."

Acceptance Criteria

• A domain pack loads from a single manifest (skills, templates, corrector with integrity manifest, catalog schema, onboarding score, eval corpus pointer); the pulse-designer score is one field of the quantum-control pack
• Post-refactor regression parity: the full public skill bundle stages and one transmon X-gate solve completes green
• The qec-autoresearch skill ships in the public bundle; its corrector gates are hash-pinned and live outside agent-editable paths; the staging-dir-only copy is retired as the source of truth
• A campaign entity exists in the run ledger with warrant-minted bounds (max samples, tier, spend); unattended launches outside a live warrant are refused
• One QEC campaign runs end-to-end through telaio with researcher and corrector as separately qualified tasks, staging survivors for human review
• Promotion to the catalog (pulses or codes) requires human ingest — no automated promotion path exists
• The eval spine scores research episodes per role per endpoint and reports the measured correction curve (logical-error vs shots)

Testing Decisions

Reuse-first: extension unit tests gain pack-manifest fixtures alongside the existing score-loader tests; amico-run's gate tests extend to campaign warrants; telaio's mock-transport suite gains campaign-loop tests; Prova's corpus gains research episodes rather than a parallel eval system. New surface only for the pack manifest schema (one validate() consumer per repo, per the schema package's existing pattern).

Key Decisions

• Roles, not modes — researcher / corrector / librarian / experimenter as capability profiles over qualified endpoints; plan/build remain permission envelopes in the chat transport.
• The campaign is the unit of autonomy — unattended propose/gate/stage cycles run only inside warrant bounds; warrants are human-minted; promotion is always human.
• Open studio, sell the -issimo tiers — Amicode, telaio, the vault mount stack, and Prova stay open (adoption engine + the RLVR gym asset); revenue is the entitled corrector-backed tier per pack (Altissimo solver, Intonatissimo closed-loop calibration, Piccolissimo paths, Legatissimo, Spira board codesign).
• Spira enters as vocabulary, not product — it names the ascending loop (the goal); the board-codesign repo keeps its name.
• Model strategy — K3 / Muse Spark class endpoints hold capability floors (campaign design, corrector oversight, authoring); cheap endpoints run the noisy researcher role the loop tolerates by construction; a future Specialist fine-tune enters only as an endpoint that passes the qualification battery per task — nothing depends on that bet.

Data Contracts

• Pack manifest — the unit of generality: skill list, template registry, corrector (gate binaries + integrity manifest), catalog schema, onboarding score, eval corpus pointer. Owned by the schema package; one validator consumed by extension, CLI, and CI.
• Campaign record — hypothesis, search-space reference, budget (samples / shots / spend), warrant reference, promotion rule (always stage-only in v1), and the role→endpoint qualification snapshot the campaign launched with.
• Corrector verdict — the evidence block (verify / refute / dedup / novelty), the validator hash stamp, and the threshold-check result where the domain has a second-stage decoder. Extends the QEC verdict shape; the pulse analog is the verification snapshot.

Constraints & Invariants

• The corrector is tool-grounded, independent (transversal), and measured — gate binaries are hash-pinned and live outside agent-editable trees; below-threshold autonomy is a bug, not a mode.
• Stage-only promotion: agents stage survivors; humans ingest. No warrant carries promotion rights in v1.
• The kernel never names a model; endpoints earn candidacy through the measured qualification battery.
• The physics never moves into the harness: the run gate, solvespecs, run dirs, vaults, and catalogs stay outside telaio.
• Pulse users see no breakage until the studio-surfaces workstream; every earlier workstream is regression-parity gated.

Prior Art

• The QEC autoresearch prototype: the research operating manual + the staged qec-autoresearch skill + the hunt fleet scripts (the loop, proven on a second domain).
• The 2026-08-10 thesis note (vault): autoresearch as error-corrected knowledge; the above-threshold condition.
• Telaio spike: event-sourced kernel, policy engine with capability floors and budget ceilings, qualification battery, CONTRACT-region enforcement, golden tiered-solve test.
• The autonomy spine already shipping: capability warrants, the plan ledger, critic/planner subprocess agents, fleet signal queues, capability profiles, veloce.

Source

Design-of-record: vault spec note spec-20260813-162300-autoresearch-studio (personal mount), promoted from the 2026-08-10 QEC autoresearch thesis note. Will be decomposed into workstream sub-issues WS1–WS7.

Notes

Why quantum is the training ground (carried verbatim into the spec): (1) hard by nature — infinite-dimensional spaces, hard dynamics, observability challenges; (2) reward signals are real and computable — code validation, fidelity, measurement — so the corrector can be mechanical, which is what keeps the corrector error rate above threshold (and yes, the measurements themselves need optimizing too); (3) fast experimental timescales versus robotics / drug discovery / materials — once the lab and fridge are set, classical resources are the bottleneck; true at Stanford today, contingent, and never load-bearing in the plan.

Honest caveats — telaio is a four-commit spike (the WS-ordering hedge exists for exactly this reason); corrector independence is load-bearing (the QEC gate lives in a separate repo with CI-pinned hashes — pack design must preserve that); the qec-construct / qec-verify sibling skills referenced by the staged skill do not exist anywhere yet.