# Reify Phase Two: Technical Specification — r1 (reconciled against live repos) *Revision of [reify_phase2_spec.md](reify_phase2_spec.md) incorporating [review_2026_07_11.md](review_2026_07_11.md). Changes from r0 are marked ⟲. The governing philosophy is unchanged: analogical not statistical, trajectory not state, inspectability over confidence, honor failure.* --- ## Governing Constraint (read before implementing anything) Schema pressure will concentrate here. The predictive layer consumes **only** membrane output (Logos coordinates); it has **no write path** to filing, vivification, or the fabric tree. Enforce this in code, not convention: - Prediction modules import from the coordinate layer only. No imports from filing/promotion code. - ⟲ Enforced by test: `tests/test_membrane.py` asserts the prediction module's import list contains no filing/promotion modules and that no prediction function signature accepts `raw_text`. The flat repo root makes convention the only wall; the test is the wall. - No prediction function receives raw account text. - Any proposal to "normalize accounts for better matching" is rejected by design. Matching quality improves by growing the precedent base, never by reshaping inputs. ⟲ Precedent for this constraint working: the 2026-07-10 contamination fix (operators commit `3944ecf`) — machinery vocabulary had colonized the emergent tree; the fix confined filing to left keywords. Same membrane, same discipline, already proven once. --- ## 1. New Container: the Restoration Vector Sibling to `inf_*.json`. Trajectory, not snapshot. **File:** ⟲ `res_<8-hex>.json` (hash id, matching the `inf_285ae7ab` convention — not sequential). Lives beside the account's inference container; indexed, but the human-walkable path is authoritative for reading. ```json { "res_id": "res_9f3c21ab", "account_ref": "inf_285ae7ab", "scale": "institution", "signature_initial": { "act_type": "assertive", "cooperative": "honored", "transmission": "archive", "resonance": "illusion", "authority": "sovereign", "utility": "narrative", "social_field": {"grid": 0.85, "group": 0.3, "quadrant": "hierarchical"}, "structural": {"layer": "institution", "scale": "institution"} }, "intervention_trace": [ { "type": "legal.remedy.exoneration_motion", "actor": "human_operator", "ts": "...", "notes_ref": "optional pointer, never inline prose" } ], "signature_terminal": { "resonance": "harmony", "...": "remaining dimensions re-read after outcome" }, "delta": { "movements": { "resonance": ["illusion", "harmony"] }, "tension_score": -0.42 }, "outcome": "restored | partial | unresolved", "_src": ["ostrom", "dunbar"] } ``` Rules: - `signature_*` fields hold coordinate values in Logos vocabulary. They are copied from the membrane (a normal logos re-read), never computed here. - ⟲ **`delta` replaces r0's `tension_delta`.** The four tension types (structural/conflict/functional/resolution) are architectural node types, not per-account measurables — functional tension is "rules in CLAUDE.md," not a number an account carries. The delta is therefore: (a) `movements` — per-dimension coordinate changes over the eight logos dimensions, each expressible and enum-validated; (b) `tension_score` — the scalar left/right divergence delta, already measured by the pipeline. If typed per-account tension is ever actualized (it is listed unactualized in pillars/FLOW.md), the delta can grow — by deliberate schema change at the membrane, not by example drift. - ⟲ **Terminal vocabulary decision:** terminal signatures reuse the existing enums (`resonance: illusion → harmony` is expressible today). Where restoration has no honest existing value (e.g. restored authority — current enum `sovereign|tribal|occult`), the enum grows by a deliberate, versioned addition to `config/coordinates.json`, reviewed as a membrane change. `logos_fused` validation stays authoritative; no value enters via a `res_` file first. - `outcome: unresolved` vectors file under `unresolved/` (honest-latency sibling of `unclustered/`). Never coerced to `partial`. - `resolution/` fabric paths autovivify only when a vector with `outcome: restored` is promoted. The vector does not create fabric structure directly — it is offered to the normal categorize→promote flow. ⟲ That flow now exists as code: `promote.py` (both repos, 2026-07-10); `unresolved/` reuses its empty-paths holding-pen pattern. --- ## 2. Scale (⟲ rewritten — substantially already built) r0 treated scale as an unbuilt blocking dependency. In the live repos: - `logos.structural` **already carries** `scale` and `layer` per inference (Sutton: `scale: institution`), emitted by `structural_operator.py` / `logos_fused.py`. - `cross_scale.py` (vivify-operators) **already consumes it**: `scale = structural.scale or structural.layer`, builds Psi–Phi–Omega coordinate fingerprints, and surfaces cross-scale isomorphic pairs — r0's `cross_scale_analogy` concept, running today. Remaining work, in order: 1. **Enum unification.** r0 proposed `person | dyad | group | institution | culture`; the structural layer vocabulary already in use is `self | dyad | small_group | local_network | institution | global`. **Use the existing vocabulary** — one scale language across `inf_*`, `res_*`, and cross_scale. Do not introduce a parallel enum. (Mapping for r0 readers: person→self, group→small_group, culture→global.) 2. **Require on containers.** `scale` is required on every new `res_*.json`, copied from the account's `logos.structural.scale`. 3. **Backfill = a structural_operator pass** over inferences lacking `logos.structural`, not a new mechanism. Where the operator cannot assign confidently, `"scale": null` — scale-pending, honest latency applies to metadata too. 4. Cross-scale matching rule (unchanged from r0): a precedent at scale S is a **direct precedent** only at scale S; at any other scale it is emitted as `cross_scale_analogy` with mandatory flag — a weak hypothesis, rendered differently (§5). --- ## 3. Right-Side Intervention Ontology (unchanged in substance) The one place a fixed schema is correct. Typed, versioned, boring on purpose. ``` intervention/ legal.remedy.* (exoneration_motion, appeal, settlement, ...) cognitive.cbt.* (reframing, exposure_protocol, ...) procedural.* (record_correction, review_mandate, ...) relational.* (mediation_session, facilitated_dialogue, ...) ``` - Namespaced dotted types; extension by adding leaves, never by loosening the grammar. - An intervention record is a **trace of what was done** — machinery vocabulary, right-side only. - Effect evaluation happens exclusively in Logos coordinates (`delta`, terminal signature). No intervention type carries a "success rate" attribute; efficacy lives in the vectors, contextualized by signature and scale. - Membrane contract: intervention types cross into a `res_*.json` only inside `intervention_trace`. Nothing in the fabric tree, filing vocabulary, or coordinate space may reference an intervention type. - ⟲ File as `config/interventions.json` beside `config/coordinates.json` — both are membrane schemas, versioned the same way. --- ## 4. Matching: the Precedent Reader ⟲ **Name decided: `precedent_operator.py`.** r0's alternative `resonate` is rejected by the abstraction's own terminology discipline — the empathy axis must not be overloaded. ⟲ **Build on `cross_scale.py`, do not start fresh.** Its Psi (resonance + cooperative), Omega (banded tension), and dynamic (conflict schema/behavior/terrain) fingerprint is the signature representation; its store-walk and pair-surfacing are the neighbor mechanics. The precedent reader extends it from *pairs of accounts* to *account → restoration vectors*. Pipeline (read-only against the store): 1. **Input:** coordinate signature + scale of a new high-tension account. 2. **Neighbor search:** distance over the eight dimensions within same-scale vectors. Distance metric is per-dimension and ordinal where the lineage supports it — but the metric is machinery and its output is never shown as a number to the human reader. 3. **Partition:** direct precedents (same scale) / cross-scale analogies (flagged, via the existing cross_scale channel) / none. 4. **Compose:** emit a **prospective reading** — a second-order composition in Logos vocabulary describing candidate coordinate movements, e.g. `resonance: illusion → harmony`, each movement anchored to its precedent vectors by `res_id`. 5. **No-neighbor case:** emit honest latency — "the field cannot yet read a path here" — and file the query itself as demand signal for corpus growth. Never extrapolate. Recursion hook: when new vectors are promoted, the precedent reader's neighbor index rebuilds as part of the existing corpus re-read. Self-evolution = the reachable-delta map refining, nothing more exotic. --- ## 5. Output Structure: the Prospective Rendering Reify-phase output. Logos as communication: the grammar of the rendering **is** coordinate movement. ⟲ All movement values must be enum-valid per `config/coordinates.json` — the rendering inherits membrane validation; examples updated accordingly. ```json { "rendering": "prospective", "account_ref": "inf_285ae7ab", "scale": "institution", "field_state": { "pressure": ["authority", "resonance"], "blockage": ["transmission"], "openings": [ { "movement": { "resonance": ["illusion", "harmony"] }, "precedents": ["res_9f3c21ab", "res_44a1c0de"], "standing": "direct", "reading": "accounts with this signature moved toward restoration when authority rebalanced and resonance realigned" }, { "movement": { "authority": ["sovereign", "tribal"] }, "precedents": ["res_b2276e01"], "standing": "cross_scale_analogy", "scale_of_precedent": "small_group", "reading": "a weak hypothesis carried across scale; read the precedent before acting" } ] }, "human_gate": true } ``` Rendering rules (unchanged from r0): - **Never a probability, score, or ranking number in human-facing output.** Openings are ordered by precedent standing (direct before analogy) and precedent count — ordering is presentation, not verdict. - Every opening links to walkable precedent nodes. Inspectability is the substitute for confidence intervals. - `human_gate: true` is structural and permanent on prospective renderings: the system points, the human decides. No auto-apply path from rendering to intervention. - Visual layer consumes this JSON; the JSON is the contract, the visual is presentation. --- ## 6. The Closed Loop (unchanged from r0, grounded) 1. Prospective rendering issued → human operator selects and enacts an intervention (right side). 2. Intervention trace recorded against the account. 3. On observed outcome: coordinates re-read (normal logos flow), terminal signature captured, `res_*.json` written with honest `outcome`. 4. Vector offered to categorize→promote (⟲ `promote.py`, live). `restored` may grow `resolution/` paths; `unresolved` files under `unresolved/`. 5. Corpus re-read fires; precedent index rebuilds; operators refine. Failure handling is the integrity test of the whole phase: an `unresolved` vector is a first-class citizen of the store. If reporting pressure ever pushes toward reclassifying failures, the precedent base rots and prediction silently degrades into wishcasting. --- ## 7. Build Order (⟲ revised for what already exists) 1. **Scale unification + backfill** — adopt the existing structural vocabulary; run structural_operator over untagged corpus; `null` where unconfident. (Mostly an operator run, not new code.) 2. **⟲ `refile.py` pass** — pipeline plumbing both phases depend on, twice-observed gap (2026-07-12): `promote.py` only moves *out of* `unclustered/`; nothing re-files an already-filed inference whose `category_paths` change afterward. Observed causes: a synonym merge re-normalizes keywords and shifts `paths[0]` (the delayed_vindication→delayed_justice merge left Washington at a retired address until moved by hand); a clean re-categorize empties the paths of previously-filed accounts (contamination-fix aftermath). Spec: move a filed inference **and its sibling files** (`reify_*.md`) to current `paths[0]`, retire emptied directories, demote to `unclustered/` when paths go empty. Complement of promote: promote = unclustered→filed; refile = filed→refiled/demoted. Idempotent, dry-run, same shape as promote.py. 3. **`res_*.json` container** + `unresolved/` path + promotion wiring (reuses `promote.py`). 4. **Intervention ontology** — `config/interventions.json`, typed, versioned. 5. **`precedent_operator.py`** — extend `cross_scale.py`: same-scale matching first, cross-scale analogy channel second. Plus `tests/test_membrane.py` (import wall). 6. **Prospective rendering** JSON contract, then visual field on top. 7. Loop closure: outcome capture → vector promotion → index rebuild. Each step is independently useful; none requires the later ones to justify itself. ⟲ Step 1 is smaller than r0 assumed; steps 1–3 together still upgrade the store from case library toward functional model. ⟲ Step 2 pays off immediately in phase 1: every approved synonym merge triggers the re-normalize → re-categorize → refile sequence.