Transformational Intelligence: implementation and evidence register
This is a software implementation record. The original manuscript, its PDF, LaTeX source, companion archive and existing public URL are unchanged.
Specification source: Alpha commit d0d18127e3d75b2b1bd13ca448265ba7e75fc21e.
Original paper: https://github.com/MontrealAI/AGI-Alpha-Agent-v0/blob/main/docs/Transformational_Intelligence.pdf.
| Immutable artifact | SHA-256 |
|---|---|
docs/Transformational_Intelligence.pdf |
d7170e8217e3bf6aa1b84868c1c76c5994c6995c5297ed092d7e72a6430c664f |
docs/Transformational_Intelligence.tex |
9d4a664d951d6b97f5136bf0b44f1ca458e90d585434a6a27a542694bb2ece5f |
docs/Transformational_Intelligence_LaTeX.zip |
49b89b4ca5a0db766d4e38af27d39c2ccbe98c4eec2e9b52af023fbbbca0e9b1 |
Capability and limitation register
Paths beginning successor/ below refer to
alpha_factory_v1/core/runtime/successor/. An implemented row is not a claim
that a PR is merged, a release published, or a live service qualified.
| Paper section / obligation | Implementation | Test / retained evidence | Supported scope / remaining limitation |
|---|---|---|---|
| Sections 2 and 4: separate purpose, formulation and measurement | successor/transformation.py: intent charter, typed assumptions, formulation, executable selection language and witness |
tests/runtime/test_transformation.py; full native envelope and shared JS inspector |
Bounded public-source fixed-summary task. Supplied annotations require auditing; no automatic truth extraction. |
| Section 5 Algorithm 1: diagnose, construct, compare and retain unsuccessful attempts | DiscoveryLoop, installed operator registry, signed existing Journal |
Native phase, budget, interruption and stale-binding regressions | Installed closed-language operators; deterministic discovery is explicitly labeled. General learned operator invention is not established. |
| Complementary edits | Graph representation plus compatible closure construction; isolated edits and unchanged-frame controls | Native development observations and browser comparison | Constructed mechanism example; unrestricted exact search can match it. No advantage over all automated discovery methods claimed. |
| Section 6: original-outcome bridge and structural attribution | Fixed summary artifact, source clauses and dependency checker, common structural score, unrestricted comparison | Native full-object envelope; browser downloadable summary | Structural support under supplied graph. Reader understanding and practical utility lack calibrated evidence. No unknown ceiling certificate. |
| Section 7: evaluator revision | successor/measurement.py: signed verifier disagreement evidence plus separate configured controller approval, immutable measurement epochs |
tests/runtime/test_measurement.py, test_measurement_review.py |
Distinct locally configured principals are role separation, not independent organizations. Signatures attest provenance; they cannot establish semantic truth. |
| Appendix B: confirmation arithmetic and error accounting | MeasurementLedger: journal-wide attempt sequence, explicit claim estimands, width-two paired bound, calibration correction |
Arithmetic, failed/parallel attempts, renaming, immutable fixed sample and adversarial signature-mutation tests | Requires valid sampling and independently controlled custody. One journal covers its projects; cross-journal/global service coordination and external checkpoints remain operational requirements. |
| Missing calibration | Outcome bound remains null unless a separately supported bound is supplied; calibration error events accounted per instrument epoch | Explicit unknown-calibration and corrected-margin tests | Agreement among models is not accepted as calibration evidence. No empirically calibrated reader-understanding instrument supplied. |
| Level 3: improve discovery across fresh complete campaigns | successor/campaign_experiment.py: train executable discovery policies, freeze selection, then run complete fresh native campaigns; shared initial lead agent and no archive |
test_campaign_experiment.py; scripts/reproduce_transformational_study.py; separate training/confirmation fixtures |
Completed bounded finite policy-selection example: two training and four terminal campaigns. Direct complementary search matches enumeration coverage with eight fewer search units per campaign. This is not calibrated reader benefit, independently sampled transfer or general learning-to-discover. |
| Level 4: purpose-grounded formulation change | Versioned executable transformation and comparison record | Public-source pilot, complementary controls and retained incumbent | Scoped structural result only; independently measured real-world benefit remains open. |
| Continuous reuse and SUCCESSOR Ω | successor/transformation_library.py: versioned native archives, separately signed local review/promotion, retirement, rollback and scoped cost-aware routing |
tests/runtime/test_transformation_library.py; existing successor state suite |
Local selection confers no spending, publication or execution authority. Declared logical costs are not monetary economics. Learned router transfer and independent outcome evidence remain open. |
| Appendix A: exact finite-model study | Original immutable companion archive; separate reproduction command | 50 original tests; all 15 manifest entries; byte-identical regenerated results; independent 13-horizon and 12-threshold checks | Passive 1/2, active 8191/8192 with at most 12 queries, unrestricted equality and cost controls. Not an LLM result or autonomous causal-model discovery. |
| Appendix C: public-source understanding | Guided EN/FR fixed-summary workspace, audit map, credible controls, artifact export, pause/cancel/recovery | Native tests, browser journey, accessibility/viewport checks | Browser source annotations are local data. Interactive-understanding outcome is not substituted for a fixed-summary request. |
| Computer-mediated work | Existing Alpha Computer Use engine; Fleet #53/#54; native receipts and separate Manager/Jobs adapters | Existing native/browser qualification workflows; real result-shape export regression | Scripted local execution and independently inspectable state are distinct from live-model performance, broad desktop support and economic independence. |
| Mathematical/scientific investigation | Separate successor/transformation_math.py affine-recurrence driver; executable scalar, cycle and affine-composition controls, algebraic witness and finite search |
test_transformation_math.py; examples/transformational-math/expected-results.json; actual 864-case durable campaign |
Every admitted census output is exact. Frozen affine candidate uses 43,488 units versus 30,204 for cycle detection: retain the stronger cycle control. Human-readable argument, not a proof-assistant certificate, novel theorem or general investigation system. Its own domain schema is not native source-selector v1. |
| Actual model/tool adapters | successor/transformation_model.py, pure frozen-program operator and alpha-agent successor-model; unchanged native execution/checker |
test_transformation_model.py: actual loopback HTTP fixture, full native campaign, consumed permits, in-flight HTTP cancellation, global STOP/resume revision fencing, provenance/attribution forgeries |
OpenAI-compatible local no-charge proposal adapter; closed executable program language, separate native approval. Source-tree CLI tests do not qualify a clean installed wheel or actual model weights. Reported model revision and billing remain operator/provider attestations. |
| Useful public-input pilot | scripts/reproduce_public_preparation_pilot.py; exact Manager 8b0cf7fb documentation snapshots and separate source/prerequisite checker |
tests/runtime/test_public_preparation_pilot.py; retained deterministic card, credible controls, successful inert reverse-import and recovery |
Actual public documentation reading aid with locally supplied annotations; not deployment guidance or calibrated reader benefit. Strong unrestricted search ties the constructed method. Mandatory real-weight release acceptance reuses the pinned llama.cpp/Qwen3-4B process; its exact-head result remains pending. |
| Seven transformation axes | Separate successor/transformation_source_audit.py and CLI: executed fact-list/index, repeated-query/materialization, graph-work-queue, fragment reuse, package-boundary and checked-handoff alternatives; external measurement revision |
test_transformation_source_audit.py; all 64 programs, independent wildcard/closure assertions, all 512 three-module graphs and independently recomputed operation counts |
Six actual finite static-source mechanisms plus a separately signed evaluator process. This is not general invention of arbitrary formulations. The native source-selector retains its three built-in operators; this domain has its own schema and grants no native import or reuse authority. |
| Useful static source-package pilot | Six immutable Apache-2.0 Alpha source snapshots, exact inventory/closure artifact, durable approval/freeze/confirmation, unchanged and hand-engineered controls plus executed unrestricted oracle | examples/transformational-source-audit/; 37 driver tests plus 21 measurement tests; actual CLI and signed Journal verification |
Frozen 110010 fails the supplied stale-worker confirmation; retain the negative result. All 32 checked-handoff methods succeed. Full comparison costs 4,832,867 declared units, hand control 22,745 and oracle 22,721. No oracle advantage, calibrated outcome or production-efficiency claim. |
| Shared persistent TI journey | Fleet-owned versioned project history, Alpha guided workspace and actual native/route/handoff adapters | Native browser request/result and Computer Use handoff; Fleet project/adversarial integration tests | Browser history and native evidence remain explicit. Full saved-package four-repository TI journey and external review/promotion remain final integration gates; inert historical routes grant no authority. |
| Parent lifetime and public portable history | Reserved parent-study lease checked inside ordinary native transactions; exact child request/phase/source binding; all-event public screening | Parent pause/cancel bypass, unplanned child, ordinary-history exposure and serialized structured-credential regressions | Known source reuse within the same Journal is rejected, including imported and archived source records. Semantic independence and cross-service custody remain outside these local checks. Exact inert authorization metadata is not a credential exception for arbitrary objects. |
| Ecosystem boundaries and money | Manager #1607 USDC inspection; Jobs separate AGIALPHA adapter; existing versioned receipts | Native zero-cost envelope compatibility and separate disposable currency fixtures | Current native transformation charter has zero USDC and zero AGIALPHA allowance. Positive paid drafts must refuse it. Policy doubles do not establish native paid support. |
Workstream dependencies and ownership
| Priority | Work | Owner / dependency |
|---|---|---|
| P0 | Finish existing Jobs security remediation, keep coverage and authentic scanner gates | Existing Jobs #3991 lane; no duplicate security edits |
| P0 | Qualify current four-repository execution/payment interfaces | Existing integration lane: Alpha #4836, Fleet #53, Manager #1606, Jobs #3992 |
| P1 | Correct Fleet native artifact export | Fleet #54, independent source review and automated review; hosted workflows required |
| P1 | Correct exact-Alpha-source qualification closure/index bypass | Jobs #3993, independent source review; actual byte/mode comparison and exact-head checks |
| P1 | Add durable transformation discovery and guided source pilot | Alpha discovery owner; extends existing runtime and flagship |
| P1 | Add independently revisable measurement epochs and global attempt accounting | Alpha measurement owner; separate engineering reviewer |
| P1 | Add separate inert transformation request inspection | Manager #1607 and Jobs corresponding follow-up, dependent on exact native Alpha schema |
| P1 | Add finite model, mathematical and complete-campaign discovery examples | Separate Alpha domain drivers; independent scoped reviews; original paper unchanged |
| P1 | Integrate persistent shared TI history and exact native source capture | Fleet follow-up and Alpha workspace; both currency receivers remain distinct and inert |
| P2 | Qualify a compatible immutable release set and all four published sites | Integration lead after exact-source checks, reviews, finite source profiles, packaged installation and live verification |
Coordination: https://github.com/MontrealAI/AGI-Fleet-Linux/issues/52.
Reproduction and operator boundaries
python scripts/reproduce_transformational_paper.py
python -m pytest tests/runtime/test_measurement.py tests/runtime/test_measurement_review.py --confcutdir=tests/runtime -o addopts= -q
python -m pytest tests/runtime/test_transformation.py --confcutdir=tests/runtime -o addopts= -q
python -m alpha_factory_v1.core.runtime.cli --home ./local-ti successor-transform demo --output ./local-ti-result.json
python -m pytest -q tests/runtime/test_campaign_experiment.py tests/runtime/test_transformation_public_content.py tests/runtime/test_transformation_model.py tests/runtime/test_transformation_math.py
python -m alpha_factory_v1.core.runtime.successor.transformation_source_audit_cli --home ./local-source-audit --output ./local-source-audit-evidence.json demo examples/transformational-source-audit/fixture.json --approve-free-local-audit
python -m scripts.reproduce_public_preparation_pilot --state ./new-public-pilot --output ./new-public-pilot.json --approve-free-local-pilot
Use a new local runtime directory for a disposable run. Existing journals and deployments remain read-only. Stop/pause and export semantics are tested per module; imported evidence never restores trusted approval. Retain an independent journal checkpoint and export before changing versions; hashes and local signatures cannot detect every rollback if the checkpoint is also lost.
The native source pilot makes no paid model calls and no payment transaction. Logical search/execution units are counted separately. Device electricity, engineering effort, human review, migration, maintenance, and a justified reuse horizon are not measured, so an economic break-even claim is unsupported.
The complete-campaign study has separate executable instructions in
examples/transformational-intelligence/policy-study/README.md. Its retained
demonstration evaluates three policies on two training and four terminal
campaigns. The learned direct-pair policy and unchanged enumeration have equal
structural coverage; declared search-cost utility differs by eight millionths.
The fixed-frame policy is better on the supplied rejection case. The four
paired campaign values are not treated as independent research evidence, and no
statistical or economic promotion follows. The actual reproduction took about
3.60 seconds on this machine; that timing is not a speed guarantee.
The mathematical example separately reports all 710,779 logical campaign units, including development, every control and oracle comparison. One census costs 468,283 more units than direct execution alone. Three further identical workloads would reach break-even under that declared model; the stronger cycle control still wins. These conditional units are not money, electricity or a permission to reuse a method. Local model discovery's explicit no-charge attestation is distinct from the fully deterministic free starting path.
The static source audit has its own bounded source domain and comparison epoch. Every method must produce the same exact report; supplied source is parsed and never executed. Its Journal retains failed comparisons and a bounded exposure checkpoint for both capture identities and individual source blobs. A previously observed file cannot become fresh confirmation by renaming a campaign, changing paths or adding another file. Confirmation packages are reserved before work; partial failures retain their exposure. Exact terminal replay is historical and performs no new work. The checkpoint cannot establish semantic independence, cross-Journal custody or external organizational review. Pause/cancel apply between bounded phases. The actual fixture ran in 7.88 seconds locally, without a model or network call; that timing is not a performance guarantee.
Initial deployment remains Ubuntu Terminal / Hardhat with a hidden, local,
ephemeral deployer-key prompt. Keys must never appear in command arguments,
.env, history, logs or artifacts. Governance remains the Mac Studio Ledger.
Software publication grants no Mainnet, ENS, signing, approval, funding,
activation or on-chain authority permission.
Publication state
Implementation and scoped tests are under review. This register does not certify a merged four-repository set, a published fresh release, installed mixed-version compatibility, or live website acceptance. Each final claim must name its exact commit and actual workflow/artifact evidence. Open research and unsupported capabilities above must remain visible in releases and product copy.
The proposed 1.36.0 path preserves the reviewed Computer Use parent and adds the public-source reading card, exact local-model command controls and mandatory real-weight pilot gate. Scoped passing tests and a retained source-checkout run are not installed-package, real-model, browser or connected-release evidence. The final head must still pass those gates before its version, artifacts and compatible four-repository set can be published. No general model advantage, calibrated utility or Level 3/4 superiority follows from this pilot.
Additive preparations are linked in Alpha #4839, Fleet #54/#55/#56/#57,
Manager #1607/#1608 and Jobs #3993/#3994/#3995/#3996. Manager's corrected
113-path predecessor is now public at f4a50134; its additive profile is being
reconciled explicitly, preserving the new receiver and captured imports. Jobs'
profile deliberately blocks publication while the additional scanner review is
absent. The separate scanner producer proposal does not establish a completed
scan. A normal hash-locked physical installation of Alpha 7283b4d4 and a
saved Fleet source-preview installation passed their scoped native/CLI checks.
The latter's host-browser proof retained one project through its signed return,
but used a dirty engineering archive and provides no Docker, nonroot or external
review qualification. Final compatible source pins, real-model acceptance,
hosted exact-source checks, clean saved installations and actual live Pages
verification remain outstanding. No main merge or fresh release is claimed.