Build the Next Discovery Lab: paper-to-code map

Source-inspection baseline: Alpha 6ab02a24b1f59306fb25e0e8da73b6210a2fe363, inspected 21 October 2026. SOURCES.json binds the exact PDF, TeX and reusable source bytes. This map distinguishes existing software from the new integrated workcell. It is not a qualification result or a claim that the seven papers share one model.

Research contract

The new campaign is a finite synthetic workcell. Supplied semantics, primitive operations, construction grammars and task families are inputs. Measurement programs, reasoning interfaces, joint realizations, conditional policies, boundary contracts and platform constructions must be computed outputs. The purpose, evaluator and grammar are frozen before construction. A stronger successor receives a new charter revision while preserving old obligations.

The papers' numerical results have different assumptions and units. None of their headline counts or savings is an end-to-end result of this campaign. Independently implemented finite checkers provide software evidence. They do not establish physical validity or external scientific peer review.

Source and implementation map

Research source and required mechanism	Existing code at the inspection baseline	New workcell obligation
Transformational Intelligence, sections 2-7 and appendix D: distinguish purpose, inherited means and measurement; executable realization, purpose bridge, competent alternatives and complete cost	Native alpha_factory_v1/core/runtime/successor/transformation.py and transformation_library.py; browser examples/transformational-intelligence/; existing capability history validators	Bind one versioned charter to all six artifacts, keep failures and alternative winners, and invalidate dependent evidence when a real constraint changes. Reuse the existing history instead of creating another project database.
Out-Learn, sections 2-8: synthesize sampling times, polarity and correction weights; learn a gain interval from a reference; jointly select calibration/use programs for the use horizon	Existing nuisance-channel typed-runtime.mjs can express supplied shielding/reference primitives. It does not construct the revised paper's polynomial-drift measurement program.	Add a bounded schedule/sign constructor, raw reference/query observations, contrast and gain checks, interval-valued response, scope checks and charged calibration. Changed drift or gain scope must trigger rejection or reconstruction.
Out-Think, sections 2-8: derive answer conflicts, close them under permitted operations, emit encoder/contextual decoder/updaters, and allow retirement of obsolete labels	Existing capability proposal/witness modules contain no general conflict-closure or canonical updater constructor	Add finite source/context/operation tables, explicit update-consistency checks and executable data programs. A code that only answers now must fail continuation qualification. Count resident labels, tables and retained source archive separately.
Out-Design, sections 3-9: construct a behavior-equivalent joint realization, qualify full independent input behavior and resource use, release a useful worker, and realize a continuous-service transition	Existing apparatus installation guards are not a joint-service compiler or occupancy transition checker	Feed the actual reasoning program into the service realization; independently check every admitted independent input combination, worker instruction/storage limits and every handover. Compare equally cheap releases at useful and useless locations and extra staging when admitted.
Out-Strategise, sections 2-5: couple observation selection with an explicit common construction valid throughout the residual belief; honor ordered checkpoints	Existing constructor searches nuisance-channel installations, not diagnosis/common-route policies over fault worlds	Add exact bounded observation/construction search with residual beliefs, robust leaf routes and resource reservations. Retain cases favoring full diagnosis and no-observation construction. A newly mandatory checkpoint must invalidate routes that bypass it.
Out-Execute, sections 3-6: derive complete boundary value functions from complete factor scopes; rebuild affected components, stop only on full semantic equality and materialize current witnesses	Existing capability-package.mjs binds nuisance ASTs and certificates; generic historical imports confer no authority. No revised complete finite boundary-profile compiler is installed.	Add scoped finite factors, checked decomposition, full profiles and attaining witnesses. Reuse only compatible contracts; reject hidden cross-component scopes and scalar-only summaries. Perform current acceptance even when an external value is unchanged.
Out-Evolve, section 2 and appendix A: construct a platform whose response has qualified bounded interaction order; freeze it, evaluate a family, then qualify a compatible stronger successor	witness.mjs already measures low-Hamming configurations and uses subset inversion in a supplied polynomial world. typed-runtime.mjs, constructor.mjs, independent-checker.mjs, capability-package.mjs and evaluation.mjs implement a separate bounded nuisance-channel constructor/reuse study.	Add an explicit bounded isolation/reference/modularization grammar and a structural semantic checker. A good sparse fit alone cannot establish absent interactions. Freeze the platform before threshold screening, inverse design and constrained selection. A new control must preserve the old restriction and requalify the stronger capability against scratch.

Paths in the table without a prefix are under examples/successor-capability/. Existing modules are reuse candidates, not evidence that a new receiver or an integrated stage is already implemented. The new runtime, UI and seams require their own source-bound tests and published acceptance evidence.

New bounded implementation paths

The additive Discovery Lab implementation is separate from the papers' original standalone experiments. Its source baseline and final qualification are recorded by the feature PR and release evidence; the source-inspection manifest above continues to identify the unchanged research inputs.

Path under examples/successor-capability/	Responsibility and scope
discovery-lab.mjs	Closed finite charter, bounded schedule construction, measured interval relation, executable state representation, compiled joint decoder/updater services, conditional fault policies, current finite factor contracts, installed platform semantics and charter invalidation. The full response cube is acquired and checked; no sparse end-to-end acquisition saving is claimed.
discovery-lab-checker.mjs	Separately implemented moment, interval, state/update, independent service-input, continuous-service, residual-world, checkpoint, complete factor/profile and full-cube interaction checks. The checker has a finite work cap and returns inconclusive when it cannot complete.
discovery-lab-baselines.mjs	Independent determinant-based measurement construction, pair-distinguishability closure, actual compiled realization, exhaustive finite policy route choices, direct factor enumeration and installed platform checks. It constructs and independently qualifies complete alternative six-stage witnesses. Its exactness concerns the declared finite grammar, not arbitrary lab architectures.
discovery-lab-contract.mjs	Shared bounded data language and zero-cap unsubmitted commission request, binding mission revision, purpose, charter, model, source and evaluator identities. It grants no execution or payment authority.

The unrestricted finite reference combines independently generated measurement, representation, realization, policy and installation choices. It enumerates every route combination in each observation partition; because the downstream completion price is monotone in policy cost under these supplied semantics, retaining the cheapest route assignment per partition preserves the joint minimum. Each admitted composed campaign receives the current independent six-stage checker, rather than inheriting a declarative tie. Search and checking costs remain charged separately from its explicitly priced horizon objective.

The fixed-architecture control executes raw measurements and corrected responses for every finite configuration on each use. Full-state reasoning emits an actual identity encoder and updater tables. Other controls check separate service occupancy, full/no diagnosis, complete recomputation, compatible incremental contracts, unchanged fixed certificates and a complete six-stage cold campaign. The scratch comparison points to the already executed, source-bound unrestricted reference, which was constructed without parent assets under the same current charter. Its measured search, checking and acquisitions are charged once in that reference row; the scratch row does not manufacture a second unexecuted campaign or double-charge an identical replay.

Mechanisms that must remain explicit

Measurement and uncertainty

The revised Out-Learn model is y = g * polarity * response + polynomial drift + error. The signal is constant within a block; gain is positive and stable across the qualified reference/use scope; drift coefficients can change between blocks. Those are supplied simulator assumptions, not discoveries by the constructor.

For degree d, the paper supplies an exact finite program grammar of d+2 readings. Barycentric weights are normalized to have unit response and zero moments through degree d. Every emitted program must check those equations. Program cost charges readings, internal polarity changes and last sample time. Calibration and use are selected together using the intended number of uses. An insufficient search cap means inconclusive, not proven infeasible.

Raw reference evidence yields a gain interval. Each query yields a feasible response interval and a checked error bound. Passing that result to reasoning cannot turn an ambiguous threshold into an exact answer. A changed drift degree, gain, reference or measurement scope invalidates the corresponding bridge.

Executable reasoning and actual released capacity

A decoder is only qualified on its admitted contexts. Updates receive the current code and operation, without hidden source access. Every modeled state must satisfy update(encode(state), operation) = encode(operation(state)). Pairwise compatible states do not automatically form a valid merged partition. A static coloring of future conflicts does not automatically supply an updater.

Re-encoding a retired representation requires retained sufficient source evidence or paid new observations. Source archive bytes and access are not free because the active label is small. Additional operations can erase compression benefits.

Joint service execution must preserve each service on independent inputs. Worker savings must come from the emitted realization and its counted resources. A vacancy must have an actual admitted route to the installation site. Every phase retains service availability; an empty-worker icon is not a release certificate.

Planning, reusable execution and stronger successors

Out-Strategise's deferred-construction result assumes nondestructive tests of the unchanged plant and no earlier benefit, deadline or observation from preparation. A workcell where installation enables a test needs the current installed state in the search. Its compiler cannot reuse that normal form without those premises.

Out-Execute profiles contain every boundary assignment, including infeasible entries, with a current attaining witness for every finite entry. A factor touching two private groups invalidates an unexpanded decomposition. Equality of one selected optimum is insufficient for universal reuse. Full profile equality can retain an outside decision while requiring a different current local witness.

The revised Out-Evolve response-class result is conditional on the constructed composition and readout. A higher-order monomial can vanish at every low-Hamming calibration point and still violate the model elsewhere. The small workcell must establish its structure from full supplied semantics and independent finite checks, and charge that qualification separately from calibration.

For a compatible new control, the old response at the new control's zero setting must remain identical within the declared interval contract, and the interaction bound must remain valid. Merely preserving control names is insufficient. Nonzero calibration error must be propagated through subset inversion; a compact coefficient vector is not an exact response oracle. A failed successor leaves the accepted predecessor usable only within its original scope.

Companion artifact custody

The main tree contains the seven research PDFs and TeX entry points. The foundation ZIP additionally contains frame_lab.py, test_frame_lab.py, frozen results and a manifest. The companion was extracted to a fresh temporary directory; all 50 tests passed, all 15 manifest entries matched, and regenerated results.json and accuracy.csv matched byte for byte. FOUNDATION_REPRODUCTION.json records this separate inherited-study reproduction. It is not new integrated-workcell evidence.

The revised manuscripts identify standalone companion package files:

Paper	Manuscript-relative companion examples	Availability at the inspection baseline
Out-Learn	code/experiment.py, code/laboratory.py, code/reference.py, code/check_artifacts.py, results/final/primary.json	Not present at those paths or as an Out-Learn companion archive in the inspected Git tree
Out-Think	code/core.py, code/reference.py, code/portable.py, results/reasoning_interface.json	Not present at those paths or as an Out-Think companion archive in the inspected Git tree
Out-Design	reproduce.py, original/revised computational packages described by the manuscript	No corresponding revised Out-Design companion archive in the inspected Git tree
Out-Strategise	code/model.py, code/planner.py, code/checker.py, results/hero_policy.json	Not present at those paths or as an Out-Strategise companion archive in the inspected Git tree
Out-Execute	code/contracts.py, code/independent.py, code/store.py, run_storage.py	Not present at those paths or as an Out-Execute companion archive in the inspected Git tree
Revised Out-Evolve	reproduce.py, operator/economic laboratories and exact protocols	Existing installed browser nuisance-channel modules are a different bounded implementation; no revised interaction-structure laboratory is installed by the manuscript reference

The public Out_Evolve.tex is an include-only root. Its section inputs and references.bib are absent at this baseline. The complete published PDF, including the interaction-structure appendix, was read instead. This workstream does not repair or replace the active research PDFs. Newly written bounded constructors must be identified as new implementations; manuscript result counts remain reported research results until their actual companion bytes are located and rerun under their own protocol.

Qualification and comparisons

The campaign must retain a strong unrestricted planner with the same primitives, fixed architecture with more search, full-state reasoning, separate service realization, full and no diagnosis, full recomputation, equally capable incremental contracts, fixed certificates, and successor from scratch. Ties and cheaper accepted competitors are valid outcomes. Each stage needs a negative control and an ablation; no requirement says each stage must help every fixture. Public seeds are reproducible regression inputs, not protected holdouts.

The resource account distinguishes acquisitions, search/compilation/checking, memory/storage, worker use, installations and measured elapsed time. Cold start, recurring use, amortization, successor increment and cumulative lineage totals must remain separate. Failed candidates, independent checks and requalification are charged. Inherited construction is charged once in cumulative lineage cost. Synthetic currency caps remain zero.

Manager USDC and Jobs AGIALPHA are alternative routes for one logical commission. Receiver admission and executor competence need exact schemas, source/runtime identities and resource/verification coverage. A draft remains unsubmitted. An unprovisioned executor remains unprovisioned. Local synthetic replay, qualified local execution and real external execution require distinct labels.