A dark factory for time series collection

A microphone on Mars and the flight muscles of a fruit fly both produce time series. Training and evaluating time series forecasting models across such different processes calls for recordings with varied time scales and relationships between measured variables. Foundation models are one use for this kind of breadth.

The internet is full of such observations. Scientific instruments deposit recordings in public archives, while experiments leave behind supplementary files and spreadsheets. Operational systems produce measurements for their own purposes. Much of this material sits outside the collections people usually reach for when training a forecasting model.

Finding it does not give you a usable dataset. The files follow the conventions of a particular instrument or experiment. A timestamp might live in a separate header, a binary format might need a custom reader, and the meaning of a column might only become clear after reading the accompanying paper.

A diverse training collection needs machinery that can work through those differences. With zeroshot, we can turn that work into an ongoing process that looks for sources, investigates their recordings and checks whether they qualify.

Define what qualifies

A usable series needs at least 1,000 observed timestamps in a coherent run, with a known clock and documented units and quality handling. We prefer related variables on a shared clock and require permission for the intended use. Interpolation cannot supply missing observations, and extra channels cannot make a recording longer.

Turn the research task into a loop

Each source can require a different small program. Coding agents can investigate documentation and implement a reader for the format they find. The graph gives that flexible work a fixed order and a set of checks before a source enters the accepted inventory.

Discovery starts with parallel proposals. An explorer looks for public sources of original observations. A synthesizer connects measurement mechanisms across domains, while a challenger looks for less familiar alternatives and obstacles the other proposals may have missed. The selector compares their findings and commits to one acquisition plan.

A probe then retrieves and decodes the selected recording within the allowed budget. When a format is difficult, it can investigate credible readers and documentation during that attempt. Its findings become fixed inputs for independent review.

The evidence reviewer reproduces the count and clock facts with a separate implementation. The method reviewer checks recording semantics and whether the handling of gaps, units and quality is justified. Originality review compares the acquired information with published source families and any sources already accepted by the workflow. An agent proposing an interesting source cannot accept it on that basis alone.

Once an investigation has closed, zeroshot routes the work back to discovery. A rejected source leaves its reason in the record, so later searches can use what the machinery has already learned. If no proposal is ready, an operational audit checks the attempt before the next handoff.

The return path is what makes this a dark factory. Routine discovery and checking can proceed without a person dispatching each task or carrying findings between agents. Budgets and stop conditions bound the loop. A scientific rejection can lead to another source, while an unresolved operational failure stops the run.

Accepted sources leave an acquisition recipe and a record of the observations that passed review. Temporary recordings remain available for the independent checks, then get deleted after the result and its provenance are durable. The retained recipe supports acquiring the data again for training or evaluation.

What a diverse collection looks like

These recordings illustrate the range of sources the machinery can work with. Each brings its own interpretation of time and its own measurement semantics.

Beyond time series collection

Zeroshot is well suited to ongoing tasks that need LLM judgment and should run without someone supervising every step. The same structure can support other research and processing work whenever its results can be checked against evidence.

In a single growing agent session, the same assumptions can influence what the agent investigates and how it judges its own output. If an early mistake shapes later choices, that bias can compound across the work.

Zeroshot makes acceptance a separate step with its own session and evidence requirements. Reviewers can challenge a proposed result before it becomes an input to later decisions, while the graph keeps routine execution moving within its operating limits.