Working Vision — A1-PH2: Foundation

Arc: A1 — Self-Contained Cell. Phase: A1-PH2 — Foundation.

Essence

A1-PH2 builds Remoir's foundations: settled documentation structure, a unified stratification grammar, and a clean prototype code base. The aim is a Remoir whose foundations are clean enough to build a knowledge graph on without first having to refactor anything.

What this phase is

Laying the foundation for easy iterations in following phases. The settled vocabulary, the canonical references, the framework that the next phase can build on without negotiation. Documentation reads as truth; code reads as intention; structure reads as inevitability.

A1-PH2 is:

The foundational docs: working-vision, first-principles, north-star, object-model, milestones, heuristics, glossary, theory.
The stratification grammar: Arc → Phase → Milestone → Task plus the cross-cutting refs, codified in /paradigm/stratification.md.
A PM scaffolding authored with real content: north-star metric, strategy, JTBD, OST, OKRs, spec template, metrics dashboard.
A clean code base: capture-app, review-frontend, transcribe-worker, Supabase schema.
An exit gate with five falsifiable claims.

A1-PH2 is not:

A feature push. The code re-implements the audio capture-and-review surface cleanly; no new capabilities until A1-PH3.
A production launch. The output is a clean prototype, not a production system.
Trusted-circle or multi-user work — cross-arc; out of scope for this phase.

Milestones

See docs/milestones.md for the canonical slug-alias table and per-milestone outcomes. Summary:

Code	Slug	What it ships
M1	M-first-principles	first-principles.md
M2	M-object-model	object-model.md (four-layer model, confirmation gate, equivalence-class clustering)
M3	M-north-star	north-star.md
M4	M-working-vision	this doc
M5	M-heuristics	heuristics.md
M6	M-glossary	glossary.md (includes stratification-vocabulary section)
M7	M-theory	theory entries
M8	M-stack-eval	per-component stack ratification (ADRs)
M9	M-capture-surface	capture surface code base
M10	M-review-surface	review surface code base
M11	M-transcription-pipeline	transcription pipeline code base
M12	M-data-storage	data-storage schema + setup
M13	M-pm-content	PM artifacts authored
M14	M-exit-gate	exit-gate scoring + A1-PH3 working-vision v0 draft

Documentation milestones (M1–M7) feed Stack Evaluation (M8). M8 and the code milestones (M9–M12) interleave: M8 ratifies stack components by ADR, and the M9 Capture→Enrich→Review vertical-slice spike de-risks the enrichment dynamic, its findings informing the outstanding M8 component ADRs. PM content (M13) trails M1–M3. M14 closes the phase.

In-phase contract

Commits to

Every Remoir concept (term, code, principle, layer) resolves through exactly one source-of-truth doc.
Foundational docs read as settled truth — no dates inline, no revision blocks, no provisional markers.
The four-layer object model (L1 Substrate / L2 Emergent / L3 Structural / L4 Deliberate) is ratified and propagated.
A unified stratification grammar covers all build-progress and architectural codes.
A PM scaffolding (north-star metric, strategy, JTBD, OST, OKRs, spec template, metrics) is in place with real content, plus solo-scale and team-scale notes.
A clean code base implements the audio capture, review, and transcription surfaces — the full capture → store → enrich → review → annotate → re-enrich loop on one media type. Enrichment produces minimal, per-capture emergent structure — a transcript plus the shallow candidate entities and tags the enrichment pass derives from it. The loop's purpose is to de-risk the two-speed enrichment dynamic (does AI-proposed structure converge under user confirmation?), not to build the A1-PH3 enrichment engine.
Data storage holds Foundation's audio in a lean substrate core behind a portable storage contract — an abstraction over Supabase, not Supabase-specific logic spread through the codebase. The schema is forward-designed: a source_type discriminator (audio, calendar_event, reference, stream) and optional enrichment-context slots hold the other capture types without later refactor; only audio rows are created this phase. The specific stack is ratified at M8.
The review surface renders that minimal L2 output — transcript plus candidate entities and tags — for one capture at a time, and offers the lightest feedback affordances: confirm a candidate (writes an L1 confirmation act — the L2→L3 gate) or free-text correct it (writes an L1 correction act). Each act lands as new L1 substrate the next enrichment run re-reads.

Does not commit to

Capture channels beyond audio. Foundation captures voice memos only; photo/slide capture, person/URL capture, OCR/vision, and image storage buckets are out of scope. The forward-designed schema reserves slots for the other source types, but no non-audio rows are created this phase.
Cross-capture or corpus-wide clustering, knowledge-graph display, entity merge/disambiguation UI, and L4 Insight authoring — A1-PH3 (Enrichment) territory. Foundation's emergent layer stays shallow and per-capture, and its review affordances are confirm + free-text correct only — the minimal mechanism that makes convergence observable, not a structured-feedback or graph surface.
Trusted-circle / multi-user features — cross-arc.
Cross-cell discovery, canonical-entity resolution at scale, federated ownership — cross-arc.
Production-grade AI enrichment (A1-PH3 territory).
Knowledge-graph experimentation (A1-PH3 work).
Production hardening — later phase.
Pricing / business-model design — cross-arc.

Exit gate

A1-PH2 closes (M14) by scoring five falsifiable claims. Pass opens A1-PH3 design work (drafting working-vision-A1-PH3.md v0). Fail triggers an A1-PH2 extension or a re-scope.

Claim 1 — Foundational coherence

Any Remoir concept (term, code, principle, layer) resolves through exactly one source-of-truth doc, identifiable in under 30 seconds via docs/file-registry.md or /paradigm/stratification.md.

Test. Pick 20 random terms from the active corpus. For each, time the resolution. Count failures (>30s, or no single source).

Pass threshold. ≤2 of 20 fail.

Claim 2 — Trust-loop reliability

Over a 30-day window of daily real use, the capture pipeline records zero data-loss events and zero unrecoverable-state events; weekly trust hits sustain ≥10 with let-down rate ≤15% and all 5 JTBD trust shapes touched within the window.

Test. Data-loss events surface in friction-log.md (entries tagged capture-pipeline). Trust-hit count, let-down rate, and trust-shape coverage come from docs/pm/metrics.md weekly rows.

Pass threshold. 0 data-loss; 0 unrecoverable-state; weekly trust hits ≥10 sustained; let-down rate ≤15%; all 5 trust shapes touched.

Claim 3 — Architectural readiness for A1-PH3

A1-PH3 work can begin without first revisiting any A1-PH2 foundational doc. The working-vision-A1-PH3.md v0 draft cites A1-PH2 docs only as inputs, never as "to be refactored."

Test. Inspect the A1-PH3 v0 draft. Grep for phrases like "to be refactored," "needs updating," "still uses old vocabulary," "pending rewrite." Zero hits = pass.

Pass threshold. Zero refactor-flag hits.

Claim 4 — Operational sustainability

A 14-day disengagement returns to a coherent project state — STATUS.md current, todo.md actionable, no orphan in-flight work.

Test. Empirical if a 14-day gap occurs naturally during the phase. Otherwise simulated at M13 via a cold-read: a fresh session opens the project and asks "what's the next action?" — the answer must come from STATUS.md + todo.md alone, in under 5 minutes.

Pass threshold. Cold-read succeeds.

Claim 5 — Enrichment convergence

Over the spike's ~2-week window of daily real captures, the two-speed enrichment loop converges: new captures predominantly resolve against already-confirmed entities rather than re-spawning candidates; when slow-path refinement fires, 100% of confirmed entities persist (refinement reshapes only the unconfirmed layer, never re-litigating settled structure); and the correction rate per capture trends down as confirmed anchors accumulate.

Test. From cell.l1_records + cell.l2_emergent: (a) fast-path resolution rate — repeat entity mentions attached to a confirmed entity vs. minted as fresh candidates; (b) confirmed-entity persistence across each slow-path refinement pass; (c) corrections-per-capture (L1 correction acts ÷ captures) over the window.

Pass threshold. Confirmed-entity persistence = 100% (hard); fast-path resolution is the dominant outcome for repeat entities; corrections-per-capture trend is downward across the window.

Failure handling

If exit-gate claims fail at M14:

Claim 1 (coherence) fails. Foundational docs need more work. Open a follow-up milestone covering the specific incoherences.
Claim 2 (trust-loop reliability) fails. The system has real issues — capture friction too high, retrieval too unreliable, or both. Extend the relevant code milestone or open a hardening pass.
Claim 3 (A1-PH3 readiness) fails. A foundational gap surfaces while drafting A1-PH3 v0. Open a follow-up milestone for whichever doc carries the missing scope.
Claim 4 (sustainability) fails. Likely means STATUS.md and todo.md aren't serving as orientation aids. Open a follow-up milestone to fix the orientation surface.
Claim 5 (enrichment convergence) fails. The two-speed dynamic does not converge — corrections do not trend down, or slow-path refinement disturbs confirmed structure. This is the platform's load-bearing bet failing: it reshapes the enrichment commitments (ADR-010 superseded) and the A1-PH3 scope, not just a code milestone. Re-scope rather than extend, and feed the finding into working-vision-A1-PH3.md v0.

How this doc relates to others

docs/vision/north-star.md — long-lived, phase-independent. This doc cites it as the strategic anchor.
docs/vision/working-vision-A1-PH3.md v0 — drafted at M14 as part of the exit gate. Its readiness is itself Claim 3.
/paradigm/stratification.md — the canonical reference for the codes used in this doc.
project-management/okrs-A1-PH2.md — three objectives, two to three key results each, mapped onto the five exit-gate claims above.