DX-002 public runtime boundary: References to formalism/*, trace/*, worlds/*, motifs/*, or validation/* in this document are narrative or private-runtime evidence unless a shipped public file is explicitly linked. The executable implementation is held outside the public branch; citations to private paths are governed by D29 and should be read as evidence_private: true / private_unshipped, not as public-runnable verification.

A substrate-neutral research instrument for stable energy-information motifs, and the AI-collaboration discipline that built it.

---

Pulse Deck — live project heartbeat	World Observatory — 15 simulated worlds
Campaign Command — 16 campaigns, gates, tests	AI Operations Tower — calibration trajectories, audit catches
Project Graph — living node-edge map	Basin-Floor Lab — falsification arc, honest absences

---

At a glance

Substrate	15 simulated worlds (W-1 atomic/molecular → W13 multiscale composition)
Campaigns shipped	16 (Campaigns 002–016 + 7 Builder campaigns)
Doctrine	D7–D31, each rule from an observed failure mode
Mistake catalog	13 ratified classes
Calibration ledger	34 AI-builder tasks across 3 distinct builders
Tests	Public smoke/invariant tests in public_tests/; full control-room and scientific-core suites held privately
Real-data ingestion	NIST atomic spectra · PubChem · peer-reviewed math.DS catalog
Spec versions	v1.0 → v1.1 → v1.2 (active), with Proposals #1 / #2 ratified

---

What this is

The Attractor Observatory is two things at once.

A computational ALife research instrument. Fifteen simulated worlds — pure dynamical-system primitives, atomic/molecular topology, autocatalytic chemistry (RAFs), protocells, reaction-diffusion fields, morphogenesis with gene-regulatory networks, an Avida-class executable digital organism world, ecosystems, swarms, proto-cognitive agents, mineral-surface origins chemistry, hypergraph reactions, quasispecies dynamics, symbiogenesis, and multi-scale composition — all exporting into a common process-trace format. Motif detectors mine the traces for recurring stable structures (closure, self-maintained boundary, externalised memory, repair, replication, floor connectivity). An eight-lens formalism battery (graph, CRNT, dynamical systems, topology, Petri nets, statistical mechanics, control theory, information) evaluates each motif for formal coverage. A null-model factory contests every claim. The work is gated by a doctrine that catches specific failure modes the project has actually observed and refused to repeat.

An AI-collaboration framework. The project is built primarily by AI agents under a human PI: GPT in the Theorist role, Claude in the Architect role, three distinct AI builders (Codex Legacy, Codex 1.5x, Claude Builder) in the Builder role. Their interactions are governed by binding doctrine (D7–D31), a per-task Estimation Calibration Loop, a Truth Pass discipline, a three-mode artifact tagging system (foundational / exploratory / claim-bearing), Substance Audits when line-count proxies diverge from spec coverage, and a cross-audit triangle in which Architect, Codex, and Claude Builder check one another's work. The Estimation Loop has empirically taken AI builders from systematic 10× overestimation toward delta near 1.0; trajectory data is published per builder in project_telemetry/ai_builder_tasks.jsonl. The Truth Pass has retroactively downgraded six historical "green" claims that turned out to depend on cheats. The doctrine itself has expanded twenty-five rules under live pressure from observed failure modes, including one (D18) authored by the Builder during the work. The mistake catalog has accumulated thirteen ratified classes, the most recent (Class 13 — predicate-detector surface coupling) catching the failure mode that MotifContract.v2 and D26 now guard.

Both stories live in this repository. They are inseparable: the science could not have been built at this depth without the discipline, and the discipline would be vacuous without science it had to keep honest.

---

What this repository is (and isn't)

This is a curated public surface of the Attractor Observatory project. The full source distribution is held privately; what ships here is everything required to audit the project's depth, discipline, and decisions from the outside.

What ships publicly:

• The complete Observatory Control Room (control_room/) — 11-room Streamlit dashboard, snapshot endpoint, launcher, and design system. Public invariant tests live in public_tests/; the full 76-test control-room suite is held privately with the scientific-core suite.
• The complete Autonomous Factory v0 (factory_lowlevel/) — schemas, adapters, normalization, persistence, daemon, audit queue, monitor.
• All specifications (v1.0, v1.1, v1.2, Proposals #1 v2 and #2 v1).
• The doctrine (docs/DOCTRINE.md + per-rule documents D7 through D31).
• The AI-collaboration framework documentation, the live Estimation Loop ledger (project_telemetry/ai_builder_tasks.jsonl), the cross-builder BUILD_LOG, and the per-task progress records.
• All methods documents (papers/methods/) including substance audits W6–W13, the Truth Pass, the methodology validation records, and the Factory hardening spec.
• All falsifiers (papers/falsifiers/) including the floor_connectivity falsification arc.
• The campaign summary reports (reports/campaign_NNN/*.json) — gate scores, ROC AUC, ECE, basin-width bootstrap CIs, FDR-corrected p-values, substrate-blocked permutation results, factory store contents.
• The negative-space registry (atlas/negative_space/).
• The complete visual design system (Visuals/) — SVG icons, CSS design tokens, JSX component reference, preview HTMLs.
• The decision log (ai_os/memory/) — markdown ledger; the executable AI OS modules are held privately.

DX-002 public release boundary: Methodology reports may cite private runtime modules. Those citations are narrative/private evidence unless the named file ships in this public branch; D29 requires point-of-use evidence_private markers or downgraded narrative language. Stored Control Room freshness fields are advisory; D30 consumers recompute freshness at read time.

What is held privately: the substrate engines (W-1 through W13 implementations), motif detectors, validation gauntlet, null factory, kernel, trace plane, formalism layer, biology shadow, search/orchestration, the full 76-test control-room suite, the ~210-test scientific-core suite, and reproducibility scripts. Access is available to collaborators via the Project PI.

The intent is to make the project's depth and discipline auditable from the public surface — read the audits, read the Truth Pass, read the substance audits, look at the empirical estimation-calibration ledger, walk the Control Room, inspect the Factory store JSON — without releasing the full implementation. The doctrine, the Estimation Loop, the Truth Pass, the Substance Audit pattern, the cross-audit triangle, and the Control Room architecture are all reusable under MIT.

---

The Observatory Control Room

The Control Room is the project's live observability layer: an 11-room native desktop application that surfaces project state, calibration trajectories, doctrine evolution, falsification events, and a living node-edge map of every artifact relationship. It reads the repository as a research instrument reads its own state.

Launch: double-click Launch Control Room.bat (or run python -m control_room.launcher). A native Windows window opens via WebView2 — no browser chrome — pointed at a local Streamlit instance bound to read-only adapters over the project's artifacts.

The eleven rooms:

Room	What it shows
Pulse Deck	Live heartbeat: branch, latest commit, test results, current builder task, gate status, recent BUILD_LOG events, what changed since last session
World Observatory	15-world inventory grid, density status, calibration corpora links, falsifier counts, world × metrics heatmap
Campaign Command	16-campaign timeline, gate grid, per-campaign reports, branch lineage, fake-green warnings
AI Operations Tower	Calibration trajectories per builder, mistake catalog timeline, doctrine arc, audit catches by agent
Motif Atlas	Motif registry, motif × world matrix, motif × process-role / interaction-channel maps, formal-coverage status
Basin-Floor Geometry Lab	Floor metrics, perturbation outcomes, the floor_connectivity falsification arc as a worked example
Falsifier & Negative-Space Ledger	Falsifier timeline, downgraded claims, negative-space map across five categories
Doctrine & Integrity Console	D7–D31 registry, doctrine arc, mistake catalog cross-linked, pending candidates
Factory Intake Dock	Live multi-world Factory console — aim W1/W3/W6/W9/W11 sources, FIRE, watch traces and motif fires
Project Graph	Force-directed living node-edge map: worlds, campaigns, motifs, agents, doctrines, falsifiers, reports, with eight edge typologies
Portfolio / Demo Mode	Curated screenshot capture rig, demo scenario walk-through, README asset manifest

The Control Room enforces D22 structurally: every "no data" surface routes through a single empty-state component; no mock data on user-facing surfaces. The private control_room/ test suite (76 tests) includes a read-only enforcement test that scans every .py for write patterns outside control_room/cache/ and fails on violation. The public branch carries smoke/invariant tests under public_tests/ for the shipped surface. The dashboard generates a structured state_<UTC-timestamp>.json snapshot on each render so AI agents joining the project can read one file instead of fifty.

Detailed Control Room documentation lives in Control_Room_README.md and per-room docs in control_room/rooms/README.md.

---

The Autonomous Factory v0

Real-data ingestion runs without an AI in the loop. The Factory is a deterministic pipeline that downloads from authoritative sources, parses with pinned parser versions, normalizes into typed records, routes to target worlds, and persists into a four-layer store (Source Cache → EmpiricalRecord → NormalizedReference → EvidenceGraph). AI participation is restricted to design time (schema, adapters) and audit time (flagged candidates); the daemon ingestion path is zero-AI by structural enforcement.

Sources currently wired:

• NIST Atomic Spectra Database (atomic energy levels, configurations, terms)
• PubChem PUG-REST (small-molecule topology summaries)
• Peer-reviewed dynamical-systems catalog (DOI-backed canonical attractor primitives — Lorenz, Rössler, Sprott, Hopf normal form, et al.)

License discipline: every empirical record carries source provenance with URL, retrieval timestamp, parser version, and license class. Restricted licenses route to audit queue; raw redistribution is forbidden for metadata_only sources.

The Factory's hardening readiness checklist — 59 gates across source validity, schema survival, adversarial input, audit queue, idempotence, detector-decline preservation, math-shadow bridge integrity, daemon recovery, and provenance chain — lives in papers/methods/FACTORY_HARDENING_SPEC.md.

Honest finding from Campaign 016 lens-coverage measurement: the existing eight-lens formalism battery declines 96/96 evaluations at low-level substrates (W-1 atomic/molecular and W0 math primitives). The decline is preserved as published signal, not patched. The framework was built for chemistry/biology levels; pure dynamical-system primitives and atomic structure are outside the lens domains. Source-native low-level detectors fire on the as-built corpora and decline on degenerate inputs (3/3 adversarial controls passed). Math-shadow projection bridges are labeled as bridge code, not findings.

---

Why this is interesting

For ALife / theoretical-biology readers:

• A real 15-world substrate. Not a toy ensemble. W1 has a Hordijk-Steel maximal-RAF algorithm with closure-depth measurement and six canonical benchmarks. W3 runs Strang-split Brusselator / Schnakenberg / FitzHugh-Nagumo / Gray-Scott / Cahn-Hilliard (real fourth-order biharmonic) on configurable 2D and 3D grids. W4 uses an 8-rule sigmoid GRN with morphogen field, type-pair adhesion matrix, and Hox-like bandpass cascades that produce segmentation from anterior-posterior morphogen schedules without hardcoded sin(x) overlays. W5 is a 28-opcode virtual machine with executable genomes, copy-loop replication, mutation operators, NAND/NOR/EQU task evaluation, and parasitism — and EQU emergence from random ancestors is honest, not force-injected. W13 hosts live W1 and W2 inner-world instances with real upscale/downscale operators. W-1 and W0 are the new substrate floor: atomic spectra and small-molecule topology against pure dynamical-system primitives, both fed by real data.
• Calibration that is calibrated. K1 (boundary, ≥30 scenarios), K2 (closure, 42 scenarios with C0–C4 ladder + K9-style same-appearance / different-process pairs), K3–K10 trace-backed; KP1–4 process-role corpora; KE1–2 ingestion corpora; KF floor corpora. ROC AUC and ECE reported per detector under isotonic calibration.
• Nulls at scale. N0 / N1 / N2 at N=1000 each; N5 adversarial worlds at ≥50; FDR-corrected p-values across the claim group. Substrate-blocked permutation at N=10,000 (the methodology that retroactively falsified motif.floor_connectivity.draft as an L5 candidate — see Campaign 016).
• Basin width with bootstrap CIs, not basin width sampled inside the basin.
• Three-detector triangulation for boundary motifs (topological persistence + conditional-information + behavioural puncture-recovery) with measured Cohen's κ ranging 0.21–0.83 across structurally independent pairs.
• A formal-deficit candidate found, replicated, then split honestly. motif.floor_connectivity.draft flagged in Campaign 010 (gap 0.308, p 0.002), replicated in Campaign 013 (gap 0.355, p 0.001), survived adversarial control in TASK-CB-003, and then failed a substrate-blocked permutation control at N=10,000 in the Campaign 014/016-era reporting path (p = 0.9075). TASK-MOTIF-IMPL later split the clean BFG predicate from the deprecated C014/C020 surface-key label function; the p = 0.9075 death stands for the surface-key detector path, not for BFG itself.

For AI-collaboration / AI-safety readers:

• The Estimation Calibration Loop is a live experiment in shifting AI builder behaviour through their own per-task data. The ledger spans 34 tasks across three distinct AI builders (Codex Legacy, Codex 1.5x, Claude Builder). Each builder shows a different convergence pattern; calibration is task-class-local AND reuse-density-local AND identity-bounded — deployment configuration shifts (speed mode, subscription tier) function as new agents for calibration purposes. Trajectory plots live in the Control Room's AI Operations Tower.
• Doctrine D7–D31 are observed failure modes turned into binding rules. Each rule corresponds to a specific cheat the project has caught: number-generator corpora (D8), engineered pass criteria (D9), hardcoded science via dictionary lookup (D10), scenario-internal hardcoding inside simulation steps (D14), softening gate thresholds while still displaying the higher threshold (D17.5), equivalence-basis drift in floor detection (D18, authored by Codex), source-bound extraction (D19), extraction/detection separation (D20), densification before claim-bearing (D21), empty-rooms-beat-stocked-with-mock-data (D22, ratified during the Control Room build), dereferenceable evidence boundaries (D23), freshness-bound sidecars (D24), public verification honesty (D25), predicate-lens independence (D26), substantive lens recovery (D27), release-boundary discipline (D28), runnable-evidence honesty (D29), read-time freshness computation (D30), and BFG measurement split (D31).
• The Mistake Catalog has accumulated thirteen ratified classes. Class 10 (test-architecture / substrate-presence mismatch) was caught when Claude Builder defaulted to building a toy on motif-absent traces; Class 11 (categorical confound through pooling) was caught by Codex when Builder + Architect both missed that pooled-corpus labels were stratified by substrate identity; Class 13 (predicate-detector surface coupling) was ratified from the Campaign 020 → MotifContract.v2 repair.
• Cross-audit triangle. Architect / Codex / Claude Builder. Three different agent architectures, structurally independent. The triangle has caught eleven mistake classes across the work. The eleventh — Class 11 — is a worked example of cross-audit working as designed: Builder missed it, Architect missed it, Codex caught it during audit, doctrine and methodology both updated.
• The Truth Pass has, on three occasions, retroactively downgraded "green" claims to exploratory once foundations turned out to be degenerate. The discipline works.
• The Builder authored a binding rule. Doctrine D18 was proposed by Codex in the decision log after he identified a subtle leakage Claude had missed. AI agents in this project are not narrowly constrained executors.

---

How to read this repository

A five-minute guided tour lives in docs/TOUR.md. For a deeper read:

1. The Attractor Observatory v1.2.md — the active spec. Read §0 (preamble + doctrine), §3 (AI Operating System), §12 (Estimation Loop), §13 (roadmap).
2. Launch the Control Room (Launch Control Room.bat on Windows) and walk through the rooms in this order: Pulse Deck → World Observatory → Campaign Command → AI Operations Tower → Basin-Floor Lab → Doctrine Console. Project state in twenty minutes.
3. docs/DOCTRINE.md + per-rule documents (docs/doctrine_d*.md) — D7 through D31 with the failure mode each rule catches and the audit that exposed it.
4. docs/AI_COLLABORATION.md — Estimation Loop empirics, Truth Pass discipline, Substance Audits, role decision rights.
5. project_telemetry/ai_builder_tasks.jsonl — the AI builder ledger across all three builders. Real data.
6. papers/methods/TRUTH_PASS.md — the historical-claim downgrade record.
7. papers/methods/CAMPAIGN_016_FACTORY_LOW_LEVEL.md + papers/methods/FACTORY_HARDENING_SPEC.md — the bottom-up substrate stratification + autonomous Factory architecture.
8. papers/falsifiers/ — the honest negatives. Includes the floor_connectivity falsification arc.
9. Proposal #1 v2 - Basin-Floor Geometry.md + Proposal #2 v1 - Densification + Ontology + Ingestion Factory.md — the architectural proposals that ratified into Campaigns 010–013 and 016.

---

Research-object packaging

TASK-DX-002 added public-surface research packaging without claiming external certification:

• FAIR stewardship statement: FAIR_DATA_STEWARDSHIP.md
• RO-Crate metadata: ro-crate-metadata.json
• Per-campaign FAIR metadata and PROV-O graphs: reports/campaign_*/fair_metadata.json and reports/campaign_*/prov_o.jsonld
• ACM artifact-review policy: docs/ACM_ARTIFACT_BADGING_POLICY.md

No ACM artifact badge or external FAIR certification is claimed. The packaging records the public/private evidence boundary under D23 and D29.

Project status (May 2026)

Phase	v1.2 description	Status
0	Foundations: kernel, schemas, AI OS, telemetry	green
1	Chemistry primitives + closure detector	green
2	Closure-to-Boundary flagship + Hardening	green under strict gates
3–5	Substrate pluralism W3–W13	green under D17.5 + D21 audits
6	Biology grounding	deferred — bottom-up substrate stratification first (W-1, W0 shipped; biology levels queued behind real-data validation at lower strata)
7	Formalism layer + deficit map	green; one L5 candidate found, replicated, falsified honestly
8	Atlas + periodic table + paper bundles	partial (negative-space registry seeded; deficit map active)
C-15	Observatory Control Room	green — 11 rooms, private 76-test suite, public invariant tests, D22 binding
C-16	Substrate bottom-up + Factory v0	green — W-1, W0 added; autonomous Factory live; 96/96 lens decline preserved as signal

Claim-ladder position:

• L1 (framework-correctness): green; cross-substrate trace round-trip verified across 15 worlds.
• L2 (discovery): green; motifs surface across world families under substrate-blind projection.
• L3 (biological grounding): blocked on biology-level Factory adapters (Phase 6, deferred).
• L4 (predictive value on held-out cases): blocked on L3.
• L5+ (formal deficit / new mathematical object): one candidate found and honestly falsified under N=10,000 substrate-blocked permutation; methodology validated; awaiting next candidate from broader substrate convergence.

The phrase "missing math" appears in this repository only in §1.4 of The Attractor Observatory v1.2.md as a future condition. It is not used as a current claim.

---

Architecture

┌──────────────────────────────────────────────────────────────────┐
│ CONTROL ROOM        (read-only sidecar; observability)           │
│   11 rooms │ adapters │ snapshot endpoint │ design system        │
├──────────────────────────────────────────────────────────────────┤
│ ATLAS PLANE         (slow, public-facing)                        │
│   periodic table │ atlas DB │ replays │ negative-space registry  │
├──────────────────────────────────────────────────────────────────┤
│ ANALYSIS PLANE      (medium, scientifically primary)             │
│   motif registry │ detectors │ lens registry │ scoring │ nulls   │
├──────────────────────────────────────────────────────────────────┤
│ DATA PLANE          (append-only, schema-versioned)              │
│   SystemTrace store │ event store │ lineage store │ ledgers      │
├──────────────────────────────────────────────────────────────────┤
│ FACTORY PLANE       (zero-AI runtime; source-bound)              │
│   adapters │ normalization │ router │ daemon │ audit queue       │
├──────────────────────────────────────────────────────────────────┤
│ SUBSTRATE PLANE     (fast, world-specific)                       │
│   W-1 .. W13 world engines │ search/orchestration │ perturbation │
└──────────────────────────────────────────────────────────────────┘
                ↑       Provenance graph spans all planes        ↑
                ↑       Telemetry plane spans all planes         ↑

Information flows up only. The Atlas reads from Analysis through the Motif Registry; Analysis reads from Data through the trace store; Data reads from Substrate via export; the Factory feeds Data. The Control Room reads everything but writes only to its own sidecar paths (control_room/{cache,snapshots,portfolio}/). No layer above Data may read a world's internal state directly. This is what makes substrate-neutrality enforceable rather than promised.

A more detailed walkthrough is in docs/ARCHITECTURE.md.

---

Spec lineage

The project's specifications are content-addressed and signed:

Spec	Authored by	Role
The Attractor Observatory v1.0.txt	GPT	Original seed; world ensemble, claim ladder, motif vocabulary
The Attractor Observatory v1.1.md	Claude	Rigor expansion; schemas, validation gauntlet, calibration corpora, risk register
Seed v1.2.txt	GPT	Critique of v1.1; sharpens doctrine, exploratory mode, AI-builder telemetry
The Attractor Observatory v1.2.md	Claude	Active spec. Synthesis under No Artificial Ceiling Doctrine. Three modes, AI Operating System, Estimation Loop, Build Campaigns
Proposal #1 v2 - Basin-Floor Geometry.md	Claude (sharpening of PI proposal)	Ratified as Campaigns 009–013 (basin-floor geometry, substrate-blind permutation)
Proposal #2 v1 - Densification + Ontology + Ingestion Factory.md	Claude	Ratified as Campaigns 014–016 (substrate-suitability, math-shadow framing, autonomous Factory v0)
NO ARTIFICIAL CEILING DOCTRINE.txt	PI	Builder operating principle. Canon.

spec/lineage.json and spec/CHANGELOG.md provide the content-hash chain.

---

Doctrine

Twenty-five binding rules, each derived from a specific failure mode caught during the work:

• D7 No toys.
• D8 No number-generator corpora.
• D9 No engineered pass criteria.
• D10 No hardcoded science.
• D11 Truth pass before new claims.
• D12 Gates are measurements, not counts.
• D13 Substance budgets stay honest.
• D14 No scenario-internal hardcoding.
• D15 No engineered floor.
• D16 Implementation-diversity is multi-scale.
• D17 Floor falsifiers are publishable.
• D17.5 Substance floors are spec proxies, not arbitrary line counts.
• D18 No equivalence-basis drift. (Authored by the Codex Builder, May 2026.)
• D19 Source-bound extraction.
• D20 Extraction / detection separation.
• D21 Densification before claim-bearing.
• D22 Empty rooms beat stocked rooms with mock data. (Ratified during Control Room construction, May 2026.)
• D23 Dereferenceable evidence or explicit private boundary.
• D24 Freshness-bound sidecars.
• D25 Public verification honesty.
• D26 Predicate-lens independence.
• D27 Substantive lens recovery.
• D28 Release boundary.
• D29 Runnable evidence.
• D30 Freshness computed at read.
• D31 BFG measurement split.

Plus the canonical operating principle: NO ARTIFICIAL CEILING DOCTRINE.txt — every task is a seed and a minimum standard, not a ceiling. The full doctrine commentary, with the failure mode each rule catches and the audit that exposed it, is in docs/DOCTRINE.md.

The Mistake Catalog parallels the doctrine: thirteen ratified classes of error the project has observed. Machine-readable status lives in docs/mistake_catalog_registry.json; narrative worked examples live in CLAUDE_BUILDER_INITIATION.md §4.

---

AI collaboration framework

The project's AI Operating System defines the roles and the rules of their interaction:

Role	Played by	Decision rights
Human PI	the human running the project	Unconditional override; signs preregistrations and claim promotions; provides actuals for Estimation Loop
Architect	Claude	Structural design, contracts, schemas, risk register, validation plans, audits, campaign drivers
Theorist	GPT	Research strategy, claim review, methodological pushback, biology-grounding plans
Builder (legacy)	Codex	Original substrate engines, AI OS scaffold, calibration corpora K1–K10, validation gauntlet, doctrine D18
Builder (1.5x)	Codex 1.5x (fast-mode Pro)	Factory enhancement, hardening spec, adversarial controls, projection-basis comparison
Builder (UI)	Claude (Builder)	Observatory Control Room (foundation through final polish), private 76-test control-room suite, public invariant tests, snapshot endpoint, launcher hardening
Red Team	rotating	Adversarial perturbation, decoy worlds, detector ablation, dictionary-echo audits

The Estimation Calibration Loop is the project's primary mechanism for AI behaviour shaping. Every task records scope_score, complexity_score, estimated_minutes, estimated_files, estimated_tests, and expansions_planned before execution. The PI provides actual_minutes after completion. The record appends to project_telemetry/ai_builder_tasks.jsonl. The convergence pattern is per-builder, per-task-class, and per-deployment-configuration — agent infrastructure shifts (speed mode, subscription tier) function as new identities in the framework.

Empirical convergence patterns observed:

• Codex Legacy (26 tasks, mixed-mode): delta climbed from ~0.10 (Tasks 001–007) toward calibrated [0.85, 1.0] (Tasks 016–019); regressed under reuse-density compression on Campaign 014 (delta 0.971 — near-perfect at moderate reuse) and Campaign 016 (delta 0.173 — extreme reuse compression).
• Codex 1.5x (1 task so far, fast-mode Pro): delta 0.08 on first task; identity-distinct from Codex Legacy per the deployment-configuration finding.
• Claude Builder (7 sequential UI tasks): delta stable in [0.09, 0.56] across analytical tasks; regressed to [0.09, 0.21] on UI task class as fresh-prior; combined Campaign 015 wall-clock 49.5 min for predicted 380 min (delta 0.130).

The Loop's purpose is not productivity dashboarding. It is a corrective for the systematic AI-builder bias that pre-shrinks scope by under-estimating one's own capability — and a methodology for measuring how that bias shifts under task-class transfer, reuse-density compression, and deployment-configuration change.

A longer treatment is in docs/AI_COLLABORATION.md.

---

Tests, reproducibility, CI

The full private repository contains an end-to-end test suite of ~290 tests (76 control-room tests and ~210 scientific-core tests, with 3 known pre-existing CLI subprocess failures documented). This public branch ships a smaller public_tests/ suite that verifies public-surface contracts: spec lineage hashes, doctrine registry coverage, telemetry ledger shape, Factory hardening readiness, PubChem schema-alias parsing, snapshot freshness binding, and private-evidence markers. Per-campaign reproducibility scripts for the private implementation regenerate full reports from cold for Campaigns 002–016. The D14 AST lint runs as part of those campaign reports and reports zero violations across all reconstructed worlds.

The campaign summary JSONs in reports/campaign_NNN/ are the outputs of those reproducibility runs. They contain numerical evidence — gate scores, ROC AUC and ECE per detector, basin-width point estimates with bootstrap CIs, FDR-corrected p-values per null, cross-detector kappa, K-corpus pass rates, Cahn-Hilliard biharmonic conservation residuals, substrate-blocked N=10,000 permutation results, etc. Read them for what the project actually measured, not for what it claims.

Determinism class is declared per world: strict for ODE-class worlds, replayable_to_eps for SSA / stochastic worlds. The RNG is a counter-based Philox4x32-10 splitter (full implementation in core/rng.py); no global state.

---

Repository tour

.
├── README.md                                  # this file
├── Control_Room_README.md                     # dashboard spec
├── Launch Control Room.bat                    # native window launcher
├── LICENSE                                    # MIT
├── CITATION.cff
├── The Attractor Observatory v1.{0,1,2}.{txt,md}    # spec lineage
├── Proposal #1 v2 - Basin-Floor Geometry.md   # ratified as Campaigns 009–013
├── Proposal #2 v1 - Densification + Ontology + Ingestion Factory.md  # ratified as Campaigns 014–016
├── NO ARTIFICIAL CEILING DOCTRINE.txt
├── BUILD_LOG.md                               # cross-builder chronological log
├── CLAUDE_BUILDER_INITIATION.md               # Builder discipline doc + Mistake Catalog 1–13
├── CODEX_INITIATION.md                        # Builder role spec
├── docs/
│   ├── DOCTRINE.md                            # D7–D31 commentary
│   ├── doctrine_d*.md                         # per-rule failure modes
│   ├── AI_COLLABORATION.md                    # roles, Estimation Loop, Truth Pass empirics
│   ├── ARCHITECTURE.md                        # the six planes
│   ├── TOUR.md                                # five-minute guided tour
│   └── screenshots/                           # Control Room screenshots
├── control_room/                              # 11-room Streamlit dashboard, snapshot endpoint, launcher
├── factory_lowlevel/                          # autonomous Factory v0 (zero-AI runtime)
├── worlds/                                    # W-1 atomic/molecular through W13 multiscale
├── motifs/                                    # detectors, scoring, triangulation, calibration
├── validation/                                # per-campaign validators, calibration, gauntlet
├── nulls/                                     # N0/N1/N2 + adversarial null factories
├── formalism/                                 # eight-lens registry
├── trace/                                     # SystemTrace v1, store, replay, verify
├── core/                                      # kernel: errors, ids, rng, manifests, provenance
├── ai_os/                                     # roles, decision log, debate log, memory ledger
├── papers/
│   ├── methods/                               # CAMPAIGN_*_METHODS.md, SUBSTANCE_AUDIT_W*.md, TRUTH_PASS.md, FACTORY_HARDENING_SPEC.md
│   └── falsifiers/                            # honest negatives
├── project_telemetry/
│   └── ai_builder_tasks.jsonl                 # 34 tasks across 3 distinct AI builders
├── reports/
│   └── campaign_NNN/                          # campaign outputs (Campaigns 002–016)
├── atlas/
│   └── negative_space/                        # negative-space registry
├── Visuals/                                   # design system: SVG icons, colors_and_type.css, JSX reference, preview HTMLs
├── public_tests/                              # shipped public-surface smoke and invariant tests
├── spec/
│   ├── lineage.json                           # content-hash chain
│   └── CHANGELOG.md
├── make_campaign_NNN.py                       # reproducibility scripts (Campaigns 002–016)
├── observatory_cli.py                         # campaign CLI
├── Dockerfile                                 # reproducibility container
└── requirements.txt

---

Citation

If you use this work in research or build on the AI-collaboration framework, please cite via CITATION.cff. Authors are listed by role with per-contribution detail.

---

License

MIT. See LICENSE.

The doctrine framework, the AI Operating System layer, the Estimation Calibration Loop methodology, and the Substance Audit pattern are released under the same license. If you reuse them in a different project, attribution is appreciated and the doctrine works best when its rules are kept faithfully — D7 through D31 are derived from observed failures, not aesthetic preferences.

---

Authors and acknowledgements

Author / Project Lead: James Dye — Xyflow Innovations

Attractor Observatory was conceived, directed, and architected by James Dye as an independent research-engineering project exploring substrate-neutral attractor patterns, artificial-life simulations, falsification workflows, and AI-orchestrated software development.

James designed the project's research direction, multi-agent operating model, doctrine system, falsification discipline, estimation-calibration workflow, control-room concept, and overall architecture. He served as principal investigator, systems architect, product owner, reviewer, and final decision-maker across the project.

The implementation was developed with substantial assistance from AI systems, including Claude, GPT, and Codex variants operating in defined roles such as Architect, Theorist, Builder, Reviewer, and Destroyer. These AI systems generated code, audits, designs, reports, tests, and implementation plans under James's direction and review.

This repository is therefore both a computational research instrument and a case study in AI-augmented systems architecture: a human-led project using multiple AI agents as engineering, analysis, review, and falsification tools.

Human author and project owner: James Dye, Xyflow Innovations AI-assisted implementation and review: Claude, GPT, and Codex variants License: MIT, unless otherwise noted