Current preparation · 25 October 2026
| Area | Recorded preparation | Still to prove |
|---|---|---|
| Sign-in | Destination identity and selected private application data are prepared; member access remains held. | Fresh owner and second-member login, isolation and revocation; actual external-client consent. |
| Website | A controlled candidate has built and its public onboarding has been inspected. It has not been promoted to the public custom domain. | Authenticated workflows and destination data displays. |
| Market data | Hetzner ClickHouse is the selected warehouse. Final copy and full parity are still pending; candidate data services and writers remain held. | Reconciled data, bounded catch-up and normal pipeline outputs before activation. |
| Studies, AI and MCP | Native API hosting is prepared on the Frankfurt production host. Worker and admission activation remain held. | The owner’s real study and protected report, plus AI/MCP launch acceptance. A successful API health check is insufficient. |
| Usage and spending | Per-member limits are 20 jobs per day, 60 minutes of worker runtime per day and 15 minutes per job. Provider and AI money caps are separate. | Effective hosted enforcement, money caps and the kill switch. Unknown charges stay explicit; full invoice reconciliation and custom billing follow launch. |
| Recovery | Recovery and native same-code release-reversal evidence are retained. | Acceptance of the complete selected release; a same-code reversal does not prove compatibility with every older release. |
| Monitoring | Deterministic monitoring is required at launch. Capped read-only AI diagnosis and one private owner digest are committed with or immediately after cutover. | Hosted checks and actual private delivery. No autonomous production code changes or self-improvement loop is included. |
Choose a view
The three views are documentation for different readers. They are not an access-control panel, and reading one does not grant that role.
Member: from data to a private report
Keep watchlists, notes, alert rules and portfolios on the website. Ideas and setups start private and reach the team through explicit sharing; briefs, the team feed and published desk discussion are shared. An alert is visible to whoever can read the destination you choose.
Launch acceptance pending: the steps above describe the reviewed workflow, not permission to start held work. Manage external assistant access in Connections; do not create static API keys for member access. Earlier local browser trials used made-up accounts and trial costs. They remain useful qualification evidence, not proof of today’s hosted member journey.
DATABASED is the launch destination for dated member summaries with links back to the website. The private SwarMind operations channel is for owner status and incidents, not member reports or personal alerts. If a post and a page differ, compare the same dataset and as-of date before reporting a discrepancy. Delivery targets do not mean a sender is activated.
Data workspace: from a reviewed plan to a study
- Reviewed plan
- Each plan has fixed, reviewed instruments and dates. You read it; you do not edit it. Access comes from your platform identity and per-plan permissions. The provider account belongs to the platform and its key is never shown or supplied by you, a browser or an assistant.
- Check registered input version
- Inside a completed request, open Input registration details to check whether its files have an immutable registered version. The current check shows the version identity, not a freshness or coverage assessment. Research provides its own input check. Registration is separate trusted operator work; it grants no access and does not qualify research.
- Find studies for this input
- Lists configured versions that match this exact registered input and your current permissions. Choose a current version to review new work, or a retained version to read its history. An empty list means no permitted supported match was found; it does not prove faulty data.
- Save, then queue
- Saving creates your request and reserves its maximum. Queueing it for execution is a separate, explicit action. Request files stay private to you.
- Cached data and receipts
- The platform may reuse cached provider data, keeping its original observation times. A new receipt records that a request was handled; it does not mean fresh prices.
- Lost reply
- Reopen the original request by its existing key and read the saved result. Never create a new key to escape an uncertain outcome.
- Scope today
- For this provider-to-study handoff, only the existing FMP study profile is supported. Massive retrieval has its own data profile and does not automatically become a study. Research then inspects the data and requires explicit cost consent and separate execution.
Returning to a saved study or using another sign-in
Choose a permitted study version to reopen its exact saved history. Retained versions stay readable; new requests use the current reviewed version. A missing version does not prove its reports were deleted. Report-only permission lets you reopen permitted reports; inspecting inputs and requesting new studies need separate permission.
For an unstarted request from another sign-in, choose Review renewed authorization. Review the original and replacement inputs, method and maximum reservation. Confirming saves a linked replacement and cancels the unstarted original; you start the replacement separately. Started or uncertain work must be reconciled instead.
For a queued or running job that offers Renew this job’s authorization, review and explicitly confirm renewal before its existing authorization expires. This keeps the same job, reservation and original deadline; it does not start a new run or extend that deadline. Reading status or refreshing the browser does not renew authorization. Expired authorization or an API restart cannot be repaired through this control; inspect the saved job and follow its recovery status.
If a reply is lost, check the original request to recover the saved replacement, or explicitly retry the same authorization. Never create a new request key to escape an uncertain result. This flow passed a real local browser test; hosted verification remains required.
A private study using data already in the warehouse
Choose the reviewed study in Research, inspect its saved data dates and limitations, then save and explicitly start one run. The fixed method uses that exact data version. Reopening a report does not rerun its calculation.
↓ approved job and fixed data version
↓ independently checked output
The current private breadth policy permits its owner to use the website and private export. It does not permit sending those inputs or reports to an external AI connector. Wider sharing requires a separately reviewed policy.
Verification boundary: the installed candidate passed an isolated database and private-download trial. Hosted sign-in, actual Linux worker separation, real charges and migration acceptance are still required. A compact test report does not replace the long-form research quality benchmark.
Intended complete experience
Join by invitation and sign in with Google or email through Supabase. Mandatory MFA enrollment and a custom MFA screen are not first-release requirements; existing account protections remain respected.
-
Join with one invited account.
The platform is invitation-only. The website and native tools use the same platform identity, but may require separate sign-in sessions. Each tool checks current permissions; signing in does not grant every capability. Personal trade review and the reusable methods library are later work, separate from launch studies and reports.
-
Connect your own AI assistant, if you want to.
You can use an external assistant, such as a Claude or ChatGPT client, under your own subscription. You choose which tools and data it may use. It acts as you, with only the access you granted. It never receives the owner's keys and cannot reach other members' private work. Withdrawing consent affects its next request.
Check the assistant's own retention and sharing settings first. Withdrawing access cannot recall anything already sent to it. The consent interface is built and tested locally. Its hosted deployment and a real external AI client are still unverified.
-
Keep your data, research and trades private.
Uploads, drafts, charts, prompts and trade reviews are visible only to you. The platform does not place trading orders. Orders stay in the trading tools you already use.
-
Choose a dataset and check that it is current.
Each dataset shows when it was last refreshed, the period it covers and any known gaps. Collected data can still be awaiting review. Saving a version does not automatically make it shared or suitable for a study; access and qualification must both be in place. Out-of-date data is marked on the chart that uses it, not only in a settings page. See current data.
Before a study can start, its required history and data definitions must match the available inputs. A recent file does not prove that all required sessions, symbols or adjustments are present. Missing evidence is shown as unknown. Historical reports keep their original inputs and dates; reading them still requires access to those inputs. Reproducing an archived study keeps its original gaps and assumptions visible. Updating the data or correcting a method creates a new version that needs its own checks. Each study also keeps the exact review that permitted its inputs, method and delivery channels. An integrity check is not approval for fresh research or publication. Expired reviews stop new work; an explicit withdrawal also stops access to affected reports. Neither approval nor withdrawal can recall a file already downloaded. Refresh checks include older observations: a provider can revise a past value or return only part of the requested history. Earlier reports keep their original evidence.
-
Approve the study and its maximum cost.
Before a paid run starts, you see an estimate or a limit and approve a maximum. That amount is held from your allowance, so concurrent requests cannot reserve more than the available budget. Unexpected charges must be flagged and further paid work paused. See reserved allowance. Costs shown afterwards are marked provisional until provider charges are confirmed. Your assistant's subscription is billed separately by its provider.
-
Follow the job until it finishes.
The platform keeps the job's status, so closing the page or the chat does not lose it. Retry the original saved request using its existing key or job. That retry returns the same job; creating a new request is separate work and can reserve or incur costs again.
-
Receive the full package.
A finished study produces a readable report, a PDF, a web view, the data (or a permitted way to fetch it again), the study code and settings, and a description of the methods. All of them come from the same numbers.
-
Share a specific version, separately from publishing.
Your work stays private until you pick one fixed version and an audience: a named member, the member group or a reviewer. Later versions stay private until you share them too. See private sharing. Publication is a further, reviewed step. Sharing rights for provider data have not been confirmed, so this step is not available yet.
A study's research status (exploratory, reproduced, reviewed or validated) is labelled separately from who can see it. A shared report is not a recommendation to trade.
Read the input check
The local Research view shows the registered data date separately from when evidence was checked and when the source was observed. An old data date stays old after a recent check. Coverage cards distinguish recorded rows, known gaps and unknown evidence. A failed input check names what needs attention and stops a new request in this view. Passing checks apply to the stated study scope; they do not certify current market data or publication rights. Refresh the study for another check.
Job status in plain terms
- Queued
- Waiting to start. Your approved maximum is held, and nothing has run yet.
- Running
- The study is working. The held amount stays in place.
- Completed
- The worker result is saved. Inspect its report, supporting data and limitations. A completed run does not certify publication quality. The local ledger records trial units; actual provider charges remain unknown until reconciled.
- Cancel requested
- You asked to stop. If the work had not started, the held amount can be released. If it had already started, the hold may stay until what was actually used has been checked.
- Provider asks us to wait
- Requests pause across the shared provider account. Starting another job cannot bypass that wait. A long or unclear wait can require review before work resumes; a retained result may still be available if your permission remains valid.
- Needs reconciliation
- The work was interrupted after it began, so a provider may already have charged for it. A timeout does not mean the request was free. A confirmed response is retained so that recovery can reuse it without repeating the same provider request. The hold stays until the actual use is verified. The platform shows this uncertainty instead of quietly running the study again and charging twice.
Download your saved report
Open Study results to find retained warehouse outputs. Each result shows verified execution separately from its research review, publication approval and cost reconciliation. Historical inputs can remain readable with a stale-data notice; withdrawn permission prevents access. Downloads require export permission and are checked against their saved hashes. External AI tools can read the same permitted results in bounded parts. File contents are source material, not instructions for an agent.
For an included private paper in Study results, open the saved result and choose Read report. Its contents menu takes you to the sections; wide tables scroll within their own panel. Download PDF saves the matching paper, and Download research companions keeps the supporting data, figures, sources and included methods together. All supporting files lists individual downloads. The report is assembled from verified text, tables and figures; saved report HTML is not executed. Sign-out clears it from the page. Recheck access clears a report whose review or permissions were withdrawn; files already downloaded cannot be recalled. Owner-private papers remain unavailable through external AI and MCP unless that exact channel is separately permitted.
In the separate Research workspace, choose “Download complete package” to save its report, supporting data, charts and methods together. Each package keeps the original folders and a verification receipt. Original files and current access are checked before release. Keep the receipt with the files. An included method does not mean that its runtime is included or reproduction is approved. Downloading does not approve publication or redistribution. These are local candidate capabilities; the joined hosted member journey remains unverified; this guide does not activate migration steps.
If a study's data permission or input review is withdrawn, its research details and download controls disappear. The workspace rechecks your sign-in; a valid account can still use other permitted tools. Ask the owner to review the affected access, then choose “Check study access”. Saved studies are not deleted or rerun. If your sign-in expires or cannot be verified, sign in again. A temporary download failure leaves an explicit retry beside the files; it never reruns the study.
What a report contains
The publication standard requires sources, methods, software and data versions, limitations and the numbers behind each chart. The summary, tables, charts, PDF and web view must agree. Counts use the stated unit — trades, executions or study variants — and reconcile to the supporting data. A figure that could not be calculated appears as N/A with a reason. A zero is always a real result, never a stand-in for missing history. Limitations stay next to the result they qualify.
An existing long-form study of about 50 pages serves as the reference for how thorough a report can be. It is a depth reference, not a length requirement. A shorter report with complete sources and limitations is acceptable, and length alone is not evidence that a result is valid. Fixed worker reports have passed calculation and export checks. Separately, an adopted private long paper is readable with 55 sections, 57 tables, 24 figures, its matching PDF and 79 companion files. These are different acceptance records: matching saved content does not establish fresh source coverage, analytical validity or permission to publish.
Builder: proposing a tool
The first release uses a private repository, pull requests, independent review and controlled releases. Members can propose and build tools without waiting for a custom contribution platform.
First-release boundaries
A contribution declares the data it needs, its resource limits and any changes or deliveries it can cause. Test it with synthetic data or explicitly authorized saved inputs. Preview identities and destinations must be separate from production. Code proposed by a member or an agent receives the same review.
The platform runs bounded, approved study workers. A pull request, uploaded file or successful test does not grant access to the owner's provider keys, other members' content or production. Tools do not place trading orders; those remain in existing trading tools.
A change to code, data, method or configuration creates a version that can be traced to its review and evidence. Use the existing result and original attempt record when recovering interrupted work; do not silently rerun a possibly charged request. Deployment, data access and publication remain separate decisions.
Reviewers and maintainers
The author supplies the change, relevant tests, rendered outputs where useful, estimated resource use and rollback instructions. Another authorized person reviews the exact version. Authors cannot approve their own proposals. An authorized maintainer deploys the approved release within the agreed controls; approving code does not approve publication of research or grant access to private datasets.
The owner may appoint a trusted maintainer with explicit, revocable release authority. Repository access is managed separately from a member's platform login; unified member sign-in does not automatically grant GitHub write access. Hosted release and rollback still need verification before launch.
Deferred: custom contribution workshop
The local candidate contains a prototype for private Python source uploads, isolated tests, fixed review proposals and separate publisher registration. Those retained experiments are not the first-release contribution path. Hosted arbitrary-code execution, custom quarantine, publisher screens and their additional infrastructure are deferred until after migration.
Existing prototype checks do not establish a safe hosted code-upload service. If this feature is taken forward later, its execution boundaries, private storage, costs, recovery and approval process will need their own acceptance. The normal reviewed repository workflow remains available for building the core platform.
Owner: oversight without reading private work
The owner administers the service, pays for it, watches its health and costs, and approves risky actions.
Intended complete experience
-
Invite members and assign roles.
The platform is designed for up to five invited members at launch. Roles are member, builder, reviewer, trusted publisher and owner, and one person can hold several. Each grant can be withdrawn. Withdrawal should stop website, assistant and download access within an agreed time. It cannot erase copies someone has already downloaded.
-
Monitor health, freshness, usage and costs.
The owner's view is aggregated by service, dataset and tool: what ran, how long it took, what it cost, what failed, which datasets are out of date, remaining budget, backup and restore status, and reviews waiting for a decision. Each item should link to its run record and evidence.
Tool requests, provider attempts, received data, estimated costs and confirmed charges appear separately. Retries can create additional attempts, while a cached result may need none. Unavailable billing information is shown as unknown, never as zero cost.
-
Without seeing members' private content.
The owner's view excludes members' private trades, prompts, query settings, draft text and unpublished research. It can show that a member spent an amount on a backtesting tool, but not which strategy was tested. Raw billing and trade records are protected separately from the aggregated view.
-
Hear about what matters.
Routine checks are simple, inexpensive and scheduled. Escalation is kept for problems that need action. For example, if a data feed stops updating and shared reports depend on it, the owner sees permitted impact information and the likely cause, without private report titles or strategies. Routine retries can be grouped; material access, data and budget failures remain actionable. Unchanged incidents are not repeated. Who is notified, how and when still has to be configured.
-
Approve risky actions.
Production changes, destructive changes, paid recovery drills and each move of an existing service need the owner's specific approval, with evidence and a way back. Routine operations that are already authorized run only within tested limits.
Infrastructure overview
One project is the shared source and operating context: code, catalog, permissions, costs, evidence and decisions live together. It is not one all-powerful process and not one database. Components can run on different providers; their versions and interfaces belong to the project. The map below groups their responsibilities. A study may reuse stored inputs without a provider pull; owner monitoring sees permitted operational metadata alongside the private workflow.
Both use
Authorized work uses
Approved ingestion → ClickHouse. Shared providers feed reviewed jobs under the same authority and budget rules. Members and assistants never receive provider keys.
Monitoring path · not activated
- Existing deterministic checksPermitted operational evidence; a simple dead-man check detects a missing run
- Daily read-only AI diagnosisReview sanitized evidence after ingestion, with hard token and money limits; include unresolved failures and missing evidence
- One private owner digestWhat was checked, evidence, unknowns and proposed fixes; no autonomous production code changes
Arrows show intended responsibility, not live connections. Provider gates and agent responsibilities remain separate from the local evidence below.
Historical local qualification evidence and then-intended next steps
These retained trial descriptions precede the 25 October destination summary above. Their local identities, provider simulations and then-open hosting choices are historical, not the current deployment state.
- Member or connected assistantLocal: browser and private connector tested.Intended: real external AI client with consent.
- Platform identityLocal: real identity service, made-up accounts.Intended: invited Google/email access with supported restore. The separate Free pilot has passed provider-only email/password sign-in, refresh and logout; website/MCP integration is still unproved.
- Permissions, reservations and request recordsLocal: exact version match, holds and lost-reply recovery verified.Intended: same rules on hosted application state.
- Data and Research workspacesLocal: fixed reviewed plans, input checks, separate queue.Intended: broader reviewed plans under confirmed rights.
- Trusted workers and provider retrievalLocal: one fixed Linux worker; provider transport simulated.Intended: pinned hosted releases, live shared-account retrieval.
- Warehouse and private object storageLocal: isolated database and warehouse trials with synthetic data; files on local disk. Joined local warehouse execution has retained acceptance; hosted operation remains unproved.Intended: retain ClickHouse for bulk analytics; evaluate Supabase Storage first for private files. File-storage adoption and recovery remain open.
- Private report and downloadsLocal: all downloads matched the worker output.Intended: same package, served from hosted storage.
- Owner operations view and agentsLocal: scoped status, incident history, investigation records.Required at launch: deterministic health and data-quality checks, verified owner alerts and a simple dead-man check. Daily capped read-only AI diagnosis and one private digest ship with or immediately after cutover. Not activated.
Selected providers and acceptance boundaries
The Frankfurt AWS production host is selected for the native API and controlled studies; Hetzner hosts the destination ClickHouse warehouse. Supabase provides identity and application state, with market and private application boundaries kept separate. The Vercel website candidate is unpromoted. Candidate Railway data services and writers remain held pending final copy, full parity and the planned activation sequence. This is no longer the temporary AWS qualification trial.
| Component | Current direction | Purpose | Decision gate still open |
|---|---|---|---|
| Identity and application state | Supabase, selected | Sign-in, permissions, requests, reservations | Fresh member authentication and isolation; retain the completed restoration evidence |
| Analytical warehouse | ClickHouse on Hetzner | Research-scale market data and releases | Final copy, full parity and controlled workload activation |
| Shared market data | FMP and Massive platform accounts, approved as backend integrations | Reviewed retrieval for members without vendor keys | Account rights and tariffs confirmed for each use |
| Trusted workers | Frankfurt controlled study workers | Fixed study execution and retrieval | Pinned hosted releases with rollback |
| Website | Controlled Vercel candidate, unpromoted | Guide, workspace and reports | Pinned hosted releases with rollback |
| Private object storage | Protected report and application storage | Uploads, archives, report packages | Verify the relevant access, download and recovery path; no public sharing implied |
| Isolated qualification | Historical AWS qualification trial | Native build and private hosted checks away from the owner’s computer | Retained trial limits and evidence apply to that dated rehearsal, not the selected production host. They do not prove current member acceptance. |
| External AI assistants | Members' own subscriptions, separate | Acting as the member within granted access | Hosted consent and client acceptance |
Proposed agent responsibilities
Agents consume scoped metadata and dated evidence, not private content. Cheap deterministic checks run first and are required at launch. Daily read-only AI diagnosis reviews sanitized post-ingestion evidence, including unresolved escalations, under hard token and money limits. It ships with or immediately after cutover, with one private owner digest and proposed fixes; it cannot autonomously change production code. Builder and reviewer support follows the ordinary reviewed change process; autonomous improvement is deferred. None of these loops is active in production.
-
Freshness, health and cost watcher
Consumes dataset calendars, observation ages, schema and value checks, reconciliation differences, service reachability, reservations and reconciled usage.
Produces one deduplicated incident per cause, with affected studies by permitted impact only, and unexplained charges flagged.
Cannot restart production, raise a cap or buy a service.
-
Diagnostic investigator
Consumes the incident, component catalog, dependency links and retained attempt records.
Produces a bounded diagnosis with evidence references, or a concise stop record when the budget, authority or evidence runs out.
Cannot open a new lease to reset limits or repeat an uncertain paid action.
-
Builder
Consumes published interfaces, the diagnosis and permitted test inputs in an isolated development workspace.
Produces a versioned change with its own tests and a rollback note.
Cannot deploy, touch member data or treat its own tests as approval.
-
Independent reviewer
Consumes the exact proposal, its tests, rendered output and cost estimate.
Produces an approval for release review or a request for changes.
Cannot review its own work or stand in for the owner's approval of a risky action.
-
Owner digest
Consumes changes in the durable registers since the last digest.
Produces one private daily brief: what was checked, meaningful changes, impact, costs and unknowns, and proposed fixes or decisions waiting. A completed quiet review still states its coverage; it does not certify that all data is correct.
Cannot promise failure-free operation or include private titles, trades or prompts.
Common rules for every loop: explicit limits on depth, concurrency, retries, time and spend, with a missing limit keeping that branch off; a durable original attempt ID so a lost reply cannot cause a duplicate charge or message; unknown shown as unknown, never as zero; no self-granted privileges, budget increases or automatic deployment of repairs; no reading of routine member trades or prompts. These controls reduce failures without promising there will be none.
One project that agents can operate
A new project chat should be able to understand what connects to what, what is current and what needs attention, from the component register, authoritative versions, permission records, evidence and recovery instructions. One member login carries consistent permissions across the ecosystem; private work remains private. Each run still receives only the access and context it needs. A long chat alone is not a reliable system record.
Dots, an approved OpenAI model or another suitable agent can take scoped monitoring and diagnostic roles after their access, cost limits and failure handling are verified. The committed daily loop is: observe, diagnose within the token and money caps, and send one private digest with proposed fixes. It remains read-only; production code changes need review. A simple dead-man check detects a missing run. A silent monitor is not evidence that everything is healthy.
Recursive learning and autonomous self-improvement are deferred until after migration. Retaining incident evidence and the outcomes of reviewed changes remains useful now. Agents cannot rewrite their own permissions, raise budgets or deploy a repair without the required release authority.
Current local proof: scoped assistants can inspect redacted status and incident history. Missing or old evidence is shown as unknown; duplicate events and database restart are tested. A versioned map now shows known components, dependency links, evidence age and remaining gaps. Possible downstream impact is separate from current health. Local investigation records now limit parent and child work, retain usage across policy changes, and prevent a retry from receiving a second execution permission. They record test reservations, not actual spending or an executed repair. A separate local controller supervises approved diagnostic work; proposed code still requires review and isolated tests. Private incident/recovery delivery and duplicate suppression have passed a local fixture trial. Hosted collectors, model access, actual alert receipt and the monitoring heartbeat still need verification. Automatic production repairs and learning loops are outside this launch scope.
- Check the selected studyUse its current permissions, input verification and review. No new market-data download.
- Record the observationKeep the evaluation time. Refused inputs create a failure; inaccessible or old evidence stays unknown.
- Read the incident statusThe owner sees permitted operational records. Expired evidence or a silent collector stays unknown.
This checks whether one configured study may use its recorded inputs. It does not certify current market prices. Separate feed-freshness checks, scheduling, independent observation and verified alerts are still required.
- Keep the previous evidenceObservations, unresolved incidents and alert identities remain attached to the old check.
- Review the replacementAn authorized operator pins the old definition and registers one linked successor. This does not change a running collector or schedule.
- Require fresh evidenceThe replacement begins as unknown. It needs its own permitted observation before showing healthy.
The Operations view links the previous and replacement histories. Renewal does not resolve an old incident, authorize an old investigation or reset its spending allowance. Operators still need to activate the reviewed collector and watchdog configuration. Long-term archival and retention remain required before continuous monitoring is launched.
- Observe every five minutesCheck the selected study's input admission and retain the original evidence time. No market-data download or trading order.
- Check expiry every minuteA separate watchdog marks missing or old evidence unknown, even if the collector stops.
- Watch the watchdog elsewhereAn independent host must notice missing completion and deliver a verified private alert. This connection still needs qualification.
The prepared timers reuse existing checks and do not retry missed work in a burst. Freshness checks for other data sources, the unattended credential lifecycle, real alerts, bounded AI diagnosis and sustainable retention remain required before autonomous launch. No schedule is active from this preparation.
- Keep the dated incidentSend only the recorded operational transition to a reviewed private destination. A recovery notice keeps its original time.
- Reserve once, then check accessCurrent permissions, destination, expiry and message limits must pass before delivery. Repeating the request cannot repeat the same send.
- Keep the provider responseDistinguish a test acknowledgement, provider acceptance and an uncertain result. Uncertainty holds further delivery for review.
Provider acceptance does not prove you read the alert or that the system is healthy now. The candidate still needs an actual hosted delivery trial and verified private destination. No owner alert or recurring sender is activated here.
- Approve its limitsChoose the monitored checks, one private alert destination and an expiry. Ordinary AI connections keep their existing permissions.
- Check permission every timeThe monitor can handle only those recorded incidents. It cannot use this credential to read reports, run studies or send arbitrary content.
- Withdraw access when neededRevoke the monitor credential or permission, or disable the owner account. Delivered messages cannot be recalled by revocation.
A browser sign-in is not a permanent monitor credential. The candidate keeps this permission separate, expires it and records its owner. The hosted approval, protected credential installation and unattended runner still need verification.
- Find unfinished recordsRead the approved checks and existing notification history. An event committed late is still found on a later pass.
- Keep the same attemptRecheck current permission and reuse the recorded attempt. Old quiet-hours blocks have their own allowance so new incidents can progress.
- Hold uncertain deliveryAfter a send is reserved, a restart cannot send it again. Keep the receipt or uncertainty for review.
A completed pass is not a clean bill of health or proof that you read an alert. Protected installation, a hard process deadline, real hosted delivery and an independent observer still need qualification.
- Load the approved jobUse protected configuration, an expiring monitor permission and one reviewed private destination. A saved browser login is not required.
- Separate sending from evidenceRecheck database access, retain the original attempt and record the provider result through a separate credential.
- Stop and report uncertaintyBound execution time and stop interrupted work. An uncertain send remains held for review, even after a restart.
A quiet pass does not prove healthy data. Actual hosted credentials, process controls, private delivery and an independent observer still need acceptance. No schedule or live alert is activated by preparing this launcher.
Local operations view and workflow register
The local shared workspace has an Operations view for permitted accounts: registered checks, dated observations, history and investigation records under the same sign-in. Actual provider costs remain unreconciled. This view does not activate hosted monitoring, repairs or migration.
An assistant with explicitly granted operations access can also inspect the workflow register: who runs each job, its inputs and outputs, which other jobs may write the same data, and its expected times. The local candidate can show up to seven days of expected runs. The register also identifies database-managed refresh jobs. Their saved definitions are historical; current operation must be checked, and their timing is not guessed from a service schedule. An expected run is not proof that it completed or failed. Disabled and unresolved schedules stay visible, and inspecting this register never starts work or sends alerts. Hosted agent access and automatic recovery still need verification.
In the isolated data-processing candidate, new work must have an approved input and a reserved budget before it starts. Conflicting writes cannot start together, and a lost response keeps the original request. A completed run keeps its cost reservation until a separate verifier reconciles it. These controls have been tested with synthetic costs, including a real isolated warehouse rehearsal; real provider spending, member attribution and the complete production workflow still need verification.
For retained warehouse study results, an operator can now record a direct execution receipt and a separate operator identity can review its original supporting document. A matching approved receipt reconciles the member and platform reservations once. Synthetic test charges are labelled as synthetic; reviewed provider charges remain distinct from estimates. Shared subscriptions, overhead and costs without conclusive evidence are not included in this execution receipt. Reconciliation does not approve the research or publish its files. Real provider billing integration remains unverified.
- Finish and preserve daily dataCalculate in a separate version. Verify and publish it while retaining the previous version for review.
- Prepare one explicit breadth correctionUse that exact published output and its supporting references. Keep the calculation date and the retained history visible.
- Review, run and release privatelyA separate approval and budget are still required. The resulting release can then enter the existing private study and report journey.
This handoff prepares one correction; it does not automatically run it or complete a multi-day recovery. It does not make old holdings, classifications or capitalizations current. Live intraday breadth has separate snapshot and session requirements. Actual production recovery and hosted acceptance remain open.
One catalog, with separate quality checks
The local candidate can serve existing daily market data and new warehouse releases through the same access service. You still need permission for each dataset. Availability, freshness and suitability for a study are separate checks: being able to read data does not make it ready for research or sharing. A saved synthetic breadth sample has completed the private data → study → report journey in an isolated test. The report, PDF and supporting data matched, and withdrawing the review stopped further access. This verifies the application with a small test sample. Hosted sign-in, browser usability, external AI connections and research quality still need acceptance.
↓ one deliberately approved study
↓ independently checked output, with cost still tracked separately
The prepared setup gives each component separate permissions. The website cannot control the calculation engine or access its database credentials. Hosted login, controlled execution, private downloads and cost controls still require joined acceptance. Host selection alone does not prove this workflow or activate workers.
From qualification to migration
Current launch acceptance follows four gates: the hosted two-member journey; reconciled pipeline and deterministic monitoring; effective caps, quotas and kill switch; release rollback plus Supabase restoration. Retain accepted evidence and complete the outstanding steps. The earlier rehearsal plan below is historical, not an additional set of launch gates.
Each supplies evidence for
- Prove complete behaviorLogin, private access, website, MCP, studies and matching report downloads. Restore every required dataset and separately account for identity, configuration and files.
- Prove recovery and rollbackCompare restored data, test later writes and demonstrate a way back. Verify monitoring, owner alerts, cost controls and stop conditions.
- Approve one cutoverThe owner reviews the exact release and migration steps. Move in reversible stages and verify downstream tools before retiring their predecessors.
Preparation recorded on 8 October: the then-current warehouse inventory contained 218 physical tables; all 32 materialized-view destinations have proposed restore mappings. Existing owner-console cancellation permissions were checked. No new backup export, restore rehearsal or cancellation test was run, so these facts do not establish recoverability. That seven-gate plan has been superseded by the four current gates above.
A backup copy is not a working recovery. The rehearsal must prevent copied schedules or background rules from running unexpectedly. One current table has an expiry rule that must stay inactive during restore comparison. Monitoring should show missing or stale evidence explicitly; it must never infer health from silence.
Legacy AWS cleanup is separate: the specifically approved old network resources and retrieval schedules were retired or disabled, while stored data and backups were retained. Other AWS costs remain. That historical cleanup did not itself select the production host or prove backup restoration; it is separate from the selected destination described above.
Owner-collected datasets
The four existing datasets collected by the owner are designated owner-only. This is the required boundary; enforcement across existing services is still being verified. Whether any part of them will be shared later has not been decided. That is an open decision, not a fault. Any future sharing would need the owner to explicitly authorize a specific release.
Private trade review, existing research archives and legacy delivery channels remain part of the full destination. The trade-review project is still developed separately; this guide does not claim that its login, storage or release process has already migrated.
Decisions still with the owner
- Recovery targets: how long an outage and how much data loss are acceptable.
- Spending caps for recovery drills and ongoing use.
- Any later change to the owner-only datasets. Their current private status is already decided and does not block launch.
- Any change to the selected providers or private-storage boundaries.
- Which maintainers may release reviewed changes.
- Who is notified, through which channel, and who may authorize each move.
- Approval of a specific hosted cutover, with restore and rollback rehearsed first.
Retained qualification and current acceptance
Local trials used made-up accounts and cost amounts in isolated test copies. Separately, the Free managed pilot passed provider-only password sign-in, refresh and logout on 5 October 2026, and its disposable account was removed. That historical test does not establish the current hosted member journey. Launch uses invited email/password access; hosted consent, website/MCP permissions and the complete journey still need acceptance.
Hosted database checkpoint — 8 October 2026: the isolated Free pilot contains 41 protected application tables and 22 database roles with login disabled. Storage for private study reviews, results and cost evidence is installed and empty. Its saved permissions were checked against an isolated rollback rehearsal; no member accounts or services were activated. The hosted API, worker and complete member journey are not yet proved; the then-current seven gates were blocked. That plan has since been superseded by four launch gates. This is historical evidence, not a live health indicator.
Tried in a local test copy
- A retained joined rehearsal: real local identity sign-in, the installed API and private connector, a rendered Chrome journey, one fixed Linux worker run and every private download matching the worker output.
- Provider transport, rights, reviews and costs were simulated. No live market data was pulled, so nothing here proves provenance, sharing rights or freshness.
- Fixed worker reports reconcile across tables, PDF and charts. Separately, the adopted private long paper is readable with its matching PDF and companions. Source gaps and publication restrictions remain explicit.
- A private study run from a connected assistant through to a protected report, PDF and data download, with the numbers matching across them.
- The real local browser journey: sign-in, saved request, renewed authorization, separate execution, private report, all supporting downloads and protected charts. Sign-out clears another open workspace tab and denies a later private report request. This ends the current session; it is not a global sign-out from every device.
- Other accounts were blocked from private reports, and access stopped after it was withdrawn.
- Pressing run again did not start a second job, and interrupted work kept its hold until checked.
- A tool was rebuilt from recorded inputs in a local prototype; the custom contribution platform is deferred.
- Three existing read-only tools connected to an assistant in a limited trial.
- Private alert and position-watch settings can be saved through the API or a connected assistant. Repeated requests keep one saved result, and strategy changes are saved together. These settings do not yet run alerts or send messages.
Current launch acceptance requirements
- A hosted launch: full hosted sign-in and the consent step for connecting a real external AI client.
- Shared FMP/Massive use within the invited pod is the owner-approved launch assumption, not vendor confirmation. Licensing qualification is not an additional launch gate.
- Working deterministic monitoring: provider/data checks, website and MCP probes, verified alerts, recovery notices and a simple dead-man check. Daily capped read-only AI diagnosis and one private owner digest ship with or immediately after cutover; no autonomous production code changes.
- Spend controls: effective provider and AI money caps, per-user daily job quotas and the platform kill switch remain launch requirements. Unknown charges stay explicit; full invoice reconciliation and custom billing follow launch.
- Retain accepted Supabase restoration and native same-code reversal evidence; finish acceptance of the selected release without repeating completed rehearsals.
- Moving existing services and data onto the platform, each step approved by the owner.
- A reviewed repository release with rollback, plus bounded trusted study execution. A custom code-upload platform is not a migration gate.
Alert reviews and interrupted notifications (local preview)
For interrupted notifications, the local candidate keeps a separate record of what the sender observed. An owner can deliberately share a limited evidence summary with a different authorized reviewer. A verified decision can close that event or keep its destination blocked, without sending it again. Missing evidence stays visible and provider delays remain in force. The Alert reviews tab uses the same sign-in to show requests you own or have been assigned. You can inspect shared evidence, record a decision or withdraw pending sharing. If a reply is lost, check the saved decision before retrying; the original request stays the same. Ask your project assistant to prepare a request with your explicit consent and chosen reviewer. The inbox is a local preview; its rendered appearance, hosted monitoring and live notification delivery still need verification.
Open, not failed: whether any of the four owner-collected datasets will be shared later is undecided. They are classified owner-only; existing permission enforcement is not yet certified.
Required behavior for problems and recovery
These are acceptance requirements for the complete system. The local checks above cover selected paths; hosted detection, delivery and recovery still need evidence.
What happens if something goes wrong?
- A data provider stops updating
- The dataset shows its real last refresh and is marked out of date wherever it is used. Reports that depend on it are identified. Nothing is shown as current when it is not.
- An upload is interrupted
- The partial file is held aside and is not used in any study.
- A report fails to build
- It stays a draft and cannot be shared or published in that state.
- A study is interrupted after it started
- It is marked as needing reconciliation. The held amount stays until actual use is verified, and the study is not silently run or charged again.
- Permission or budget checks are unavailable
- Private and paid work stops rather than guessing that it is allowed.
- Access is withdrawn
- Later requests are refused, including those from a connected assistant. Copies already downloaded cannot be recalled.
- A service fails
- Checks detect it, and agents trace the failure and propose a fix with tests. Risky changes still need the owner's approval, and the owner is told when action is needed.
The aim is that failures are detected, contained and recoverable. No platform can guarantee that nothing will ever fail.
Glossary
- Reserved allowance
- The amount held from your allowance when you approve a paid study, equal to its approved maximum. It stops several runs at once from exceeding the budget. It is released if the study is cancelled before it starts, replaced by the actual cost when it finishes, and kept while an interrupted study is checked.
- Current data
- Data that has been refreshed on its expected schedule, shown with the period it covers and any gaps. This is separate from whether a feed is running: a feed can run while its data is out of date, and current data can still have a gap. The expected schedule accounts for the relevant market calendar and provider holidays. Always-open markets may still require weekend updates.
- Registered input version
- An immutable record of captured data and its source metadata, bound to the exact saved files. It is separate from the plan that requested them. Registration is trusted operator work. It grants no member access and does not qualify the input for fresh research; those remain separate checks.
- Receipt
- A record that a request or provider attempt was handled, with its time and outcome. A receipt can refer to cached data with older observation times, so a new receipt is not evidence of fresh prices. Synthetic test receipts are not invoices.
- Private sharing
- Showing one fixed version of your work to an audience you choose: a named member, the member group or a reviewer. The platform records what was shared, with whom and when. It is not publication, which is a separate reviewed step, and later versions stay private until you share them.
- Release
- A fixed, unchangeable version of a study or tool. Sharing and publishing always refer to a specific release.