squeakwell — database recovery through cross-modal constraint propagation

Table of Contents

• What Is This?
• The Core Idea
• Why This Works
• Architecture
• Naming
  • The Concept (what this approach is called)
  • The Tool (what the binary/project is called)
  • My recommendation
• Status
• License

What Is This?

A database recovery and reconstruction tool that uses VeriSimDB’s octad model and VCL-total’s progressive type system to reassemble damaged, fragmented, or inconsistent datasets through cross-modal constraint propagation.

Traditional recovery tools work at the storage level — reassembling bytes into pages. This tool works at the semantic level — using 8 independent representations of each entity as witnesses that cross-check and reconstruct each other, tightening formal type constraints progressively until the data converges on a single consistent identity.

The Core Idea

Damaged data (fragments, corruption, inconsistencies)
        │
        ▼
  INGEST: scatter fragments across octad modalities
  (whatever you can recover goes into whichever modality has data)
        │
        ▼
  PHASE 1 — LOOSE CONSTRAINTS (VCL-total Levels 1-3)
  Accept anything structurally valid. Parse, bind to schema, type-check.
  Drift score: ~0.8 (high — modalities disagree heavily)
        │
        ▼
  PHASE 2 — CROSS-MODAL INFERENCE
  Use populated modalities to infer missing ones:
    Document text  ──► infer Graph edges (entity extraction)
    Document text  ──► infer Vector embedding (encode)
    Graph edges    ──► infer Semantic annotations (type inference)
    Provenance     ──► infer Temporal ordering (hash-chain sequence)
    Spatial coords ──► infer Graph proximity (spatial join)
    Tensor data    ──► infer Vector projections (dimensionality reduction)
  Drift score: ~0.5 (filling in, but contradictions emerging)
        │
        ▼
  PHASE 3 — CONFLICT RESOLUTION
  Where inferred modalities contradict observed ones:
    Provenance chain   → which source is more authoritative?
    Temporal versions  → which is more recent?
    Semantic types     → which is type-valid at higher levels?
    Cardinality proofs → which produces the expected row count?
  Drift score: ~0.3 (contradictions resolved, consistency improving)
        │
        ▼
  PHASE 4 — TIGHTENING (VCL-total Levels 4-6)
  Apply null-safety, injection-proofing, result-type checking.
  Any data that can't satisfy these levels is flagged for human review.
  Drift score: ~0.15
        │
        ▼
  PHASE 5 — CONVERGENCE (gate ceiling L6; VCL-total Levels 7-10 are future vcl-ut work)
  Cardinality bounds, effect tracking, temporal consistency, linearity.
  The reintroduction statements for an entity are gate-admissible at level k
  (ceiling 6 from VCL text; see docs/ROUTE-PLAN.adoc section 4.4).
  Drift score: approaches 0.0
        │
        ▼
  OUTPUT: reconstructed database with:
    - Per-entity confidence scores
    - Provenance chain showing reconstruction history
    - Proof certificates for each type level achieved
    - Residual drift report (what couldn't be resolved)
    - Human review queue (ambiguous cases)

Why This Works

Eight witnesses are better than one. When you have 8 independent representations of the same entity, damage to any one can be compensated by the others:

If you lose…​	These modalities can reconstruct it
Graph edges	Document text (NER), Semantic annotations, Spatial proximity
Vector embeddings	Document text (re-encode), Tensor projections
Document text	Graph edges + Semantic annotations (template generation)
Temporal history	Provenance chain (hash-chain ordering)
Provenance	Temporal versions (timestamp ordering)
Spatial coordinates	Graph proximity, Document georeferences
Semantic types	Infer from Graph + Document + VCL-total type checking
Tensor data	Vector embeddings (expand dimensions), raw data re-ingest

Progressive typing is the ratchet. Each VCL-total level constrains the solution space further. Level 1 accepts almost anything. Level 10 demands formal proof. Data can only become more consistent through the levels, never less. This is constraint propagation applied to database recovery.

Drift detection becomes the convergence metric. Instead of alerting on drift (normal VeriSimDB behaviour), drift score becomes the objective function: recovery is complete when cross-modal drift approaches zero.

Architecture

• Ingest engine (src/ingest/) — Accepts damaged databases, fragments, exports, dumps, partial backups. Maps whatever it finds into octad modalities. Handles: SQL dumps, JSON/CSV fragments, corrupted binary files, partial WAL logs, orphaned indexes, detached TOAST data, broken foreign keys.

• Reconstruction engine (src/engine/) — The core. Runs the 5-phase progressive constraint propagation:

  • Phase 1: Loose acceptance + schema inference

  • Phase 2: Cross-modal inference (fill missing modalities)

  • Phase 3: Conflict resolution (provenance + temporal + semantic arbitration)

  • Phase 4: Type tightening (VCL-total levels 4-6; gating planned, the stub does not yet invoke vclt-gate)

  • Phase 5: Convergence (no gate level beyond Phase 4’s planned L6 ceiling; VCL-total levels 7-10 are future vcl-ut work)

• Invariants (src/abi/mod.rs) — the engine is designed so that the following hold. No proof lives in this repository and the Rust engine itself is unproved; the theory behind rungs R2–R10 (identity preservation through every round, guarded fill inflationary but not monotone, the conflict-raising counter-model, fold permutation invariance, bounded termination, the fibre-intersection and hub-ceiling lemmas) is mechanised in its estate homes and ledgered with prover receipts in PROOF-STATUS.adoc; the prover homes per estate PROOF-POLICY are in docs/ROUTE-PLAN.adoc:

  • Constraint propagation is monotonic (data only gets more consistent)

  • Cross-modal inference preserves entity identity

  • Conflict resolution is deterministic (same input → same output)

  • Type ratchet is sound (passing level N implies passing levels 1..N-1)

• VCL-UT integration (planned) — The vclt-gate binary (hyperpolymath/vcl-ut) is the certifier. Phase 1 is to gate at L1–3 and Phase 4 at L4–6, the L6 ceiling vcl-ut certifies from source today; Phase 5 adds no further gate level (L7–10 are future vcl-ut work). The current engine stub prints the level names and does not yet invoke the gate, so no phase certifies anything today.

• Output — Reconstructed database in VeriSimDB format (all 8 modalities populated), with per-entity confidence scores, proof certificates, and a human review queue for ambiguous cases.

Naming

Working title: octad-recover. Seeking a better name. Candidates below.

The Concept (what this approach is called)

Name	Rationale	Feeling
Cross-Modal Constraint Convergence	Technically precise. Recovery through converging constraints across modalities.	Academic — good for a paper title
Progressive Semantic Recovery	Captures the progressive tightening and semantic (not byte-level) approach.	Accessible — good for explaining to people
Octad Convergence	Short, ties to VeriSimDB’s octad model.	Branded — ties it to the ecosystem
Identity Reconstruction via Drift Minimisation	Captures the core mechanism: entity identity rebuilt by driving drift to zero.	Precise but wordy
Multi-Modal Witness Recovery	Each modality is a "witness" to the entity’s identity. Recovery = reconciling witnesses.	Evocative — "witnesses" is compelling
Constraint Propagation Recovery (CPR)	Like CPR for databases — resuscitation through constraint propagation. Also a medical metaphor that fits ambientops.	Memorable — CPR is a great acronym, ties to hospital model
Type-Ratcheted Reconstruction	The progressive type levels act as a ratchet — data only tightens, never loosens.	Technical — captures the mechanism clearly

The Tool (what the binary/project is called)

Name	Rationale	Feeling
resurgam	Latin: "I shall rise again." A database that resurrects itself from fragments.	Beautiful — scholarly, memorable, works as a CLI name
reconvene	Data fragments reconvene into a coherent whole.	Clear — English, accessible, verb-as-name
coalesce	Fragments coalesce into identity. Also a SQL function (COALESCE) that handles nulls — fitting.	Perfect double meaning — SQL + recovery
octad-recover	Descriptive. Does what it says.	Safe — clear but not inspiring
witness	Each modality is a witness. The tool reconciles witnesses.	Evocative — but might clash with other tools
converge	The drift converges to zero. The data converges on identity.	Simple — strong verb, clear meaning
palimpsest-db	A palimpsest is a manuscript where earlier writing shows through. Recovering underlying data.	On-brand — ties to your Palimpsest License. Metaphorically perfect.
anamnesis	Greek: ἀνάμνησις — "recollection, remembering." The database remembers itself.	Scholarly — fits the Greek naming in your ecosystem (Ephapax, Phronesis)
mneme	Greek: μνήμη — "memory." Simpler than anamnesis. The memory-recovery tool.	Short — pronounceable, memorable
ektasis	Greek: ἔκτασις — "extension, stretching out." Extending fragments back to wholeness.	Thematic — but less intuitive
synapsis	Greek: σύναψις — "connection, junction." Reconnecting disconnected data.	Good — ties to the cross-modal linking concept
cpr	Constraint Propagation Recovery. CPR for databases. Ambientops hospital model.	Perfect fit — medical metaphor, short, memorable as CLI command

My recommendation

Concept: Constraint Propagation Recovery (CPR) or Multi-Modal Witness Recovery

Tool: resurgam (scholarly, memorable, "I shall rise again") or cpr (ties to ambientops hospital model, instantly memorable, explains itself)

Status

Concept stage (2026-11-05; proof ledger 2026-11-09). Architecture sketched, phases defined, engine and client are stubs. No proof lives in this repository and nothing about the Rust engine is proved; the first mechanised theory results for the recovery invariants (rungs R2–R10) landed in verisimdb, echo-types and tropical-types and are ledgered with receipts in PROOF-STATUS.adoc. The route to a working recovery, the decisions still open, and what remains to be created in verisimdb, vcl-ut, echo-types and tropical-types are in docs/ROUTE-PLAN.adoc. The crate, repository and clade are named squeakwell; the naming table above is kept as the record of the alternatives.

License

SPDX-License-Identifier: CC-BY-SA-4.0