Useful Alpha: a capability worth keeping / un outil utile à conserver

Open the flagship · Ouvrir en français · Implementation

Start with an opportunity worth developing: valuable digital work, a useful software improvement, a testable research direction or a rigorous mathematical argument. Define who benefits, compare strategies and ask whether a different formulation could produce a better result. Keep the work product, evidence and lessons together. Mission preparation is free and needs no account, model or wallet. A configured native executor and model perform the Computer Use work. An optional catalog worked example remains available locally without a provider.

Version 1.35.0 adds the Computer Use path below while preserving the catalog journey. The matching published manifest and acceptance evidence identify delivery; a source version alone does not. Published v1.34.0 retains its original catalog scope and immutable evidence.

Computer Use: Level 3 and Level 4

The flagship starts with Discover my next capability. Its default mission examines the public Computer Use runtime at released v1.34.0 for an implementable improvement to redundant work, latency or operating cost. The requested result is a source-linked patch proposal and a reproducible before-and-after validation plan, preserving task limits, stop controls and accurate accounting. Keeping the present design is a valid result when the evidence does not support a change. The proposal is useful engineering work; tests, savings and performance gains must not be described as observed until the corresponding execution has occurred.

The research lane develops a concrete mechanism and an experiment that could disprove its advantage. The mathematics lane explores an adaptive portfolio of proof strategies, asking for explicit assumptions, competing proof routes, a checkable argument or a precise unresolved gap. Add relevant public primary sources before making literature or novelty claims. These lanes specify substantive work; they do not connect a code executor or formal prover. The same acceptance conditions apply when testing an improved strategy or reframing.

The digital-work opportunity lane starts from permitted nonpersonal source material and asks who would benefit, what concrete deliverable would help and which alternative deserves testing. Supply an exact starting brief or approved sources. Missing evidence calls for an investigation plan, not an invented discovery, measured improvement or revenue claim. Each lane retains its own acceptance conditions and matching work surface.

Capability	What the added path does
Computer Use	Prepare a bounded work surface for a configured model to inspect screenshots and propose mouse and keyboard actions
Level 3 · Meta-strategic intelligence	Record competing strategies, select a baseline or combination to test, retain observations and propose an improved method
Level 4 · Transformational intelligence	Identify an unnecessary restriction and test a different formulation against the same required outcome and acceptance criteria
One project history	Retain prepared attempts, runner returns, owner reviews and existing catalog work under the same project identity

Choose a valuable digital-work opportunity, an open mathematical question, a public software bottleneck, a decisive research experiment or your own mission. Name the capability, supply the exact starting material, describe the opportunity, explain who benefits and list what every approach must deliver. Missing evidence calls for an investigation plan, not an invented discovery, measured improvement or revenue claim.
Record two to six competing strategies, a proposed combination and an alternative problem formulation. Choose the baseline, combined strategy or reframing for this attempt. These are hypotheses until tested.
Review the supplied work surface, exact public source URLs, input rights and action/time/review limits. Choose Prepare my capability mission, then Take this mission to my executor. Preparation is free and performs no model call.
Use the native Computer Use runner to review and authorize the exact mission with an explicitly selected model and immutable container image. Provider credentials stay with the native operator. The static page does not operate a remote computer.
Import the runner's computer-return.json. Inspect its actual work product and retained screenshots, check each acceptance condition and record your review and actual review time. Keep failed or interrupted attempts too. A model's completion message alone is insufficient.
Prepare another authorized approach and compare reviewed attempts with the same outcome, work surface, sources, rights and limits. Save the complete workspace or download the complete project archive to retain catalog work, strategy attempts, returns and lessons together.

The executable goal asks the model to report strategies considered, the method actually tested, observations, a proposed improvement or combination, and unnecessary restrictions worth reframing next. Tested results and untested ideas must remain distinct. A proposed reframing earns its place through reviewed useful work; it cannot silently relax acceptance conditions, enlarge permissions or declare itself the winning method. The complete strategy slate stays visible in the native goal, while execution is restricted to the selected approach. Considering an alternative does not authorize executing it in the same attempt.

The current native actuator is a disposable Chromium browser in a Linux Docker container. It supports the current Responses computer action interface and an alternate compatible vision/JSON supplier. The operator selects an available model and qualifies it for the task; a model name alone does not establish capability. For difficult scientific reasoning, consider gpt-6-astra; for source inspection and Computer Use with a different cost/capability tradeoff, consider gpt-6.1-sol. These are current documented choices, not a universal ranking or an automatic provider selection. See the task-fit model guidance. Full desktop control, arbitrary code execution, formal proof checking, personal profiles, private data, logins, payments and message sending are outside this adapter. A browser-observed proposal or argument is not an executed patch, verified theorem or measured economic improvement.

Imported returns establish consistent historical bytes. The browser does not authenticate signer trust or independently certify the result. Verify native receipts against a separately trusted source key and review the actual artifacts. Owner acceptance does not approve payment, settlement, institutional admission or operational authority. USDC and $AGIALPHA commissioning retain their separate routes and requirements.

The existing catalog example remains available without a provider. Local control tests and scripted browser checks do not establish live-model qualification or paid outcomes for the new path. Those require authorized provider trials, retained usage/cost evidence and review before a release can claim them.

Maintained local catalog tool

The existing catalog example runs locally without an account, installation, model, API key or wallet. Open Maintained example · local catalog tools for the catalog workflow. Its supplied inputs are synthetic; your own inputs can be public, synthetic or licensed and nonpersonal. This tool remains alongside the primary capability-development journey.

Choose Run the maintained catalog example to run the supplied catalog example.
Inspect the joined catalog, unmatched rows and discrepancies. Check that the selected keys and fields express the task you intended.
Inspect correctness, operation counts, observed timings and the comparison limits. A strategy can lose; that result remains useful evidence.
Download the useful files, your project record and the working offline tool. Reopen the saved HTML in a modern desktop browser to continue locally. Save a new copy after changes.

Choose Use the catalog tools with my data to paste or select two CSV files. Select each catalog's key; common non-key columns are compared as exact text. Inspect identifiers and duplicates before relying on the output. A reconciliation finds the differences expressed by its contract; it does not establish that either catalog is complete or factually correct.

Keys are exact, case-sensitive, nonempty strings. Whitespace and numeric-looking strings are retained; 001, 1 and 1 are different keys. Duplicate keys produce every matching left/right pair, rather than silently deduplicating. The joined catalog contains matching pairs; unmatched rows are exported separately. Inputs support quoted commas/newlines and a UTF-8 BOM. Each table is bounded to 200,000 UTF-8 bytes, 1,500 data rows, 32 columns and 4,096 characters per cell. Output is capped at 10,000 matching pairs.

CSV downloads prefix formula-like cells with an apostrophe for spreadsheet import. The exact JSON retains the original strings. Use that JSON when exact original cell values matter.

File	Use
`joined.csv`	Every exact matching pair, with left/right column names
`unmatched-left.csv`, `unmatched-right.csv`	Input rows with no matching key on the other side
`conflicts.csv`	Differences in shared non-key fields for matching pairs
`exact-result.json`, `comparison.json`	Original cell strings, task and engineering observations
`report.html`	A readable local result and comparison report
`project.json`	Retained task history for inspection and continuation
`successor-tool.mjs`	The reusable Node tool for another pair of local catalogs

The browser additionally saves the whole workspace as one self-contained HTML file. This working browser copy and the standalone Node tool are separate ways to keep the capability.

What Alpha means here

Alpha is an advantage worth building for a particular mission. A working join and a discrepancy report can be useful before anyone has measured whether they justify replacing the current workflow. The question is whether the resulting capability helps its intended user enough to warrant its full costs.

For a different mission, use the opportunity notebook to record the goal, current approach, credible alternative, measurable target, supplied evidence and budget. Ask what current evidence, expertise and feasible experiments make achievable now. These records stay typed hypotheses; recording one neither retrieves new sources nor executes a domain experiment. The catalog capability is a maintained provider-free local tool; the Computer Use path adds a separate project-bound handoff to the native model runner.

Stage	What this flagship provides	What requires further evidence or a separate decision
Find Alpha	A valuable mission, beneficiaries, acceptance criteria, explicit strategies and a bounded native handoff	Actual usefulness, adoption constraints and benefit
Develop in a Gym	Model-guided browser work with retained results; maintained local catalog experiments remain available	Connected code/proof execution, verified novelty or scientific certification
Examine a frozen candidate	Bound inputs and implementation identity with inspectable local comparisons	Independently administered protected work and a credible external comparator
Prepare successor continuity	A portable working tool, project inputs and retained observations	Earned Specialist status, institutional admission and operational authority

The developmental Gym helps build and inspect capabilities. Specialist qualification requires a frozen candidate to pass the agreed constraints and outperform Current and the best credible alternative on protected work. Separate accountable authorization is required before operational use with authority. The local flagship does not complete those independent steps.

Read the comparison correctly

The comparison uses the same bounded reconciliation semantics and inputs for each maintained strategy. The four strategies are a nested scan, hashing the right table, hashing the smaller table and sort/merge. Hashing the smaller table is an established heuristic; this example does not invent a new join algorithm. Operation counts describe those implementations. Observed browser timings describe one device and run; browser optimization, load and measurement order can affect them. A strategy with fewer counted operations can take longer. Inspect both results and correctness before drawing a conclusion. The reported indexed-row counts are not measurements of memory consumption.

These are engineering comparisons against the implemented strategies, not a demonstrated advantage over every spreadsheet, database or existing commercial product. Selecting the fastest observed strategy after seeing the same workload is development evidence. Fresh protected evaluation under a frozen protocol is a separate stage.

Any user-supplied time, volume, price or cost assumptions are estimates. They do not become measured human time savings, realized revenue or accepted customer value. Count setup, integration, review, repair, maintenance and unwind when deciding whether to adopt the capability. Unknown costs and burdens remain unknown.

Construct a strategy, then test a different question

After the first result, the optional research experiment constructs executable join plans from a finite, validated grammar. Plans combine key partitioning, conditional small-product scanning, bounded hashing and an exact uniform-key shortcut. This is composition of known primitives, not unrestricted algorithm invention. The search executes the constructed plans; it does not merely rename the four initial comparison methods.

Set an explicit auxiliary-row budget and inspect the metered work. Auxiliary rows count retained row references, not bytes or process memory. The comparison includes competitive bounded block-hash alternatives as well as the original approaches. Infeasible alternatives are identified. Construction, checking and validation have costs of their own; the work metric is an inspectable proxy rather than an adoption or economic measurement.

The candidate is frozen before a fresh local synthetic check. Its inputs and generator remain inspectable and under local control, so this check is not protected evaluation. If a candidate does not establish the declared improvement, the incumbent remains. Save the research record to retain its plan, source identity, comparisons and rejection or improvement evidence. An included duplicate-heavy example makes the shortcut's narrow scope concrete; it does not establish improvement on arbitrary catalogs.

The executable reframing asks a different question: which product keys need a corrective decision? It groups unmatched occurrences and conflicting pairs into a per-key action queue, preserving references to every original exception. A coverage check verifies those references against the original reconciliation. Compare the number of exceptions with the number of grouped actions, and download the queue. Grouping changes the output contract; it neither repairs the source catalogs nor measures human review time. It can reduce repeated decisions when several occurrences concern the same key, and can correctly show no reduction.

These are bounded working examples of strategy construction and task reframing. A general ability to discover better scientific methods or fundamentally new business opportunities requires its own environment and evidence. The retained synthetic research example contains an exact mission, frozen plan, fresh confirmation and all 131 exception references grouped into two proposed actions. This is a deliberately constructed local example, with a changed review contract and no human time-saving claim.

Continue through the native workbench

The catalog adapter uses the existing SUCCESSOR workbench runtime and native Jobs compiler. It binds the exact catalog project and cycle to the work specification, executes the installed adapter locally, and independently rechecks the returned catalog outputs before accepting a continuity record. The record retains the original project identity and historical cycles. The learning archive and offline workspace preserve the associated research, action queue and native return evidence.

This is an unfunded native workbench path. Existing Manager/Fleet project receivers do not install this catalog executor; a native workbench result is not a completed Fleet execution or paid settlement. Separate USDC and AGIALPHA work briefs remain available with their own readiness boundaries. See the native continuity guide for the supported formats and exact limits.

Keep and continue your work

The downloadable project retains the task, inputs and observations. Imported historical observations are unverified history; rerunning produces new local evidence. A saved file or checksum establishes neither independent trust nor current permission. Keep the exported files before clearing browser data. Project and learning imports are bounded to 2 MB, including retained evidence. A full archive is refused rather than silently pruned; keep the existing downloads before beginning another separate project.

The retained tool preserves the mission and its working method as material for an eventual authorized SUCCESSOR Ω. Saving it does not establish that status. It does not require a particular model provider. A future implementation can replace models or methods while retaining permitted evidence, skills and operating memory. A changed capability needs fresh applicable proof and a separate authorization; predecessor proof or authority does not transfer automatically.

The broader guided project journey and five-family workbench remain available. Their native formats and execution boundaries are unchanged. A flagship project export is not automatically a Jobs/Manager/Fleet execution request or native backup. USDC and $AGIALPHA commissioning retain their separate configuration, budgets, acceptance and settlement requirements.

Premier résultat en français

Périmètre de la version 1.35.0 : le parcours Découvrir ma prochaine capacité part d'une opportunité logicielle, scientifique ou mathématique utile, avec bénéficiaires et critères d'acceptation explicites. La mission initiale analyse le code public du runtime Computer Use pour proposer une amélioration réalisable et un plan de validation avant-après reproductible. Aucun test exécuté ni gain économique ne doit être inventé. Vous pouvez aussi choisir une opportunité de travail numérique et fournir une note de départ ou des sources autorisées non personnelles. Précisez bénéficiaires et livrable utile; sans preuves suffisantes, demandez un plan d'investigation. La liste des stratégies accompagne la mission, mais seule l'approche choisie est autorisée. Au niveau 3, proposez plusieurs stratégies, testez une méthode ou une combinaison et conservez les observations pour l'améliorer. Au niveau 4, remettez en question une restriction inutile et testez une autre formulation qui respecte toujours le résultat requis. Les idées non testées restent des hypothèses.

La préparation dans le navigateur est gratuite et n'appelle aucun modèle. L'exécution utilise un modèle configuré séparément et Chromium dans un conteneur Linux isolé, avec autorisation explicite de la mission. Importez ensuite le résultat, examinez les livrables et les captures, puis consignez votre avis dans le même historique de projet. L'espace complet conserve aussi les travaux de catalogues existants. Aucun secret fournisseur n'appartient à la page; l'import ne prouve ni qualification indépendante ni autorisation de paiement. Le contrôle universel d'un ordinateur et l'usage de profils personnels ne sont pas fournis par cet adaptateur. L'exécution de code et la vérification formelle de preuves nécessitent des connexions distinctes; une proposition ou un argument candidat ne constitue pas un correctif testé ni un théorème vérifié. La version publiée v1.34.0 conserve le périmètre de rapprochement de catalogues décrit ci-dessous.

Dans Exemple maintenu · outils locaux de catalogues, lancez aussi l'exemple synthétique. Examinez le catalogue rapproché, les lignes sans correspondance et les différences. Pour utiliser vos propres catalogues autorisés, sélectionnez deux fichiers CSV et leurs clés; les champs communs sont comparés comme du texte exact. Téléchargez les résultats, le projet et l'outil autonome pour les conserver.

Après ce premier résultat, construisez et testez une stratégie dans la grammaire proposée, ou exécutez le regroupement des divergences en décisions par clé. La candidate reste soumise au budget déclaré et aux alternatives implémentées. Une absence d'amélioration conserve la méthode actuelle. Le contrôle synthétique nouveau reste public et local, et le regroupement change le contrat de sortie. L'archive d'apprentissage conserve ces éléments avec le même projet. Le relais natif utilise le workbench et le compilateur Jobs; il reste non financé et n'installe pas un exécuteur de catalogues dans Manager ou Fleet.

Les comparaisons locales décrivent les méthodes exécutées et les mesures de votre appareil. Les hypothèses économiques restent des estimations; elles ne prouvent ni économies réalisées ni valeur client acceptée. Le Gym sert au développement. Une qualification Specialist exige une évaluation protégée indépendante du candidat figé contre la méthode actuelle et la meilleure alternative crédible. L'autorisation opérationnelle reste une décision distincte. Conserver l'outil et son historique ne transfère aucune permission.

Development and verification

With Node 22.17.1 or later, run the example from the repository root:

node examples/successor-alpha/cli.mjs --example --out alpha-example

To use your own permitted local catalogs, explicitly choose both key columns:

node examples/successor-alpha/cli.mjs --left left.csv --right right.csv --left-key id --right-key id --out alpha-result

Run the focused engine checks with:

node --test examples/successor-alpha/*.test.mjs

Keep the output directory and inspect its artifacts before integrating the result. The CLI runs the same maintained reconciliation engine without a provider or blockchain request. It does not produce an independent qualification or a production grant.

Run the exported Node tool from any directory with another pair of permitted inputs:

node successor-tool.mjs left.csv right.csv id id new-result

The tool reruns all four maintained methods and exact checks. It writes new artifacts; it does not reuse historical timings as proof of the new run.

The separate retained plan download exports the strategy selected by the optional research experiment. Import it as an ESM module to apply that exact plan to fresh CSV strings:

import { reconcileCatalogs } from './successor-retained-plan.mjs';

const result = reconcileCatalogs({ leftCsv, rightCsv, leftKey: 'id', rightKey: 'id' });
console.log(result.joinedCsv, result.conflictsCsv);

It rechecks exact output and the declared resource budget on the new inputs. It refuses inputs the retained plan cannot process within that budget. The module performs no file, network or wallet operations; its caller decides how to read and save data. A held experiment exports the incumbent rather than the rejected candidate.

The maintained implementation is in examples/successor-alpha/. Its generated browser entrypoint is docs/successor/alpha/index.html. Keep source changes, the generated page and its source manifest together; do not edit bundled JavaScript inside the generated HTML. Follow the example's reproduction instructions and run its focused checks, browser/accessibility checks and the repository's required exact-source checks before publication.

Source availability, a local passing check and a published Pages deployment are separate facts. The route becomes public only after the relevant source has passed its required checks and Pages has deployed it.

Supplied visual reference

The flagship uses the owner-supplied purple and gold institution illustration, encoded as WebP for delivery. examples/successor-alpha/artwork.json records source and delivered-byte identities. The binary gate admits only this exact asset hash; the general restriction is preserved.