As-Of Information-Set Pinning

also known as Snapshot-Pinned Analysis, Point-in-Time Information Set, As-Of Pinning

Bind every input a dated decision may read, including the model snapshot, retrieval corpus, reference data and table version, to the information set that existed at its as-of timestamp, and record the pin.

Context

An agent makes decisions that are dated: a credit file is underwritten on a given day, a trade is placed against the prices quoted that morning, a sanctions screen is run against the list published that week. Every source the agent reads keeps moving after the decision is made. Fundamentals are restated, sanctions entries are added and revoked, bureau files are re-scored, a document index is re-crawled, and the served model is replaced by a newer snapshot whose training data already covers the period being reasoned about. Meanwhile the decision itself is expected to stand still: an examiner, a disputing customer or a backtest will come back to it months later and ask how it was reached.

Problem

Re-running the decision later produces a different answer, and nothing in the record says whether the agent changed its mind or the world underneath it changed. Because each read resolved to whatever the source held at the moment of the re-run, the original inputs are unrecoverable, so a dispute cannot be settled and an audit cannot distinguish a defensible call from a mistake. The same gap corrupts measurement in the other direction: an agent evaluated on historical periods reads restated figures and reasons with weights trained on text published after the decision date, so it scores itself on facts it could not have had and the backtest overstates what the deployed system would have done.

Forces

Live operation wants the freshest value of every source, while a dated decision wants the value that stood at its as-of timestamp, and one read path cannot satisfy both.
Upstream sources mutate silently and in place: restated fundamentals, revised sanctions entries and re-scored bureau files replace the earlier value rather than adding a version beside it.
Lookahead leaks through the weights as well as the data, because a model trained on unrestricted corpora embeds information published after the period it is asked to reason about, and a chronologically consistent alternative gives up accuracy to stay inside a training cutoff.
Pinning is only as durable as retention: a version identifier recorded in a decision outlives the snapshot it names once the source's retention interval expires, leaving a pin that resolves to nothing.
A recorded pin that the read path does not enforce is decoration, since a fallback to the current value produces a plausible answer and no error.

When to use

A decision is dated and can be questioned later by an examiner, a customer dispute or a regulator.
Upstream sources are restated, revised or re-crawled in place rather than versioned beside their earlier values.
An agent is scored on historical periods, where reading current values or a model trained past the decision date inflates measured performance.
Two runs of the same case must be comparable, so the input set has to be held constant across them.

When not to use

The task is about the present state of the world, where reading a frozen view would give a stale and wrong answer.
No source can be versioned or copied locally, so a recorded pin would never resolve and would only give false assurance.
The dispute window is shorter than the cost of retaining addressable history, and a plain input log is sufficient evidence.
The decision has no legal, financial or measurement consequence that would ever require re-derivation.

Example

A lender's agent declines a loan application in March. In September the applicant disputes the decision, the agent re-runs the case, and this time it approves: the credit bureau has re-scored the file and the served model has been upgraded since March. Nobody can tell whether the March decision was wrong or the inputs simply moved. Had the March run pinned its bureau vintage, policy document version and dated model snapshot, the re-run would reproduce the original decision and the examiner could see exactly what it was based on.

Diagram

flowchart TD A["Dated decision starts"] --> B["Fix as-of timestamp"] B --> C["Pin resolver"] C --> D["Model snapshot id"] C --> E["Corpus / index version"] C --> F["Reference-data vintage"] C --> G["Table version (AS OF)"] D --> H["Pin manifest stored with decision"] E --> H F --> H G --> H H --> I["Re-execution at dispute or backtest time"] I --> J{"Version still servable?"} J -- "yes" --> K["Same information set, difference is the agent's"] J -- "no" --> L["Run fails; no fallback to today's data"] A -. "unpinned path" .-> M["Live reads"] M --> N["Replay silently differs; cause unknown"]

Solution

Therefore:

Fix a single as-of timestamp when the decision starts and route every read through a resolver that turns a source plus that timestamp into an immutable version identifier: a dated model snapshot rather than a moving alias, a corpus or index version rather than a name, a vintage of reference and market data rather than a current quote, and a table version or AS OF timestamp on a store that keeps past states addressable. The resolved identifiers are written into the decision record as a pin manifest, so the decision carries the description of its own inputs rather than a pointer to whatever those inputs later became. Re-execution at dispute, examination or backtest time replays the manifest: the same snapshot, the same corpus, the same table versions, so a difference in outcome is attributable to the agent and not to the world. Retention on every pinned source is set to at least the window in which the decision can be questioned, and a read that cannot be served at the recorded version fails the run rather than falling back to the current value, which turns a silently mutating dependency into a visible error.

What it gives you

A decision can be re-derived months later from the inputs that actually produced it, so a dispute or an examination compares like with like.
A difference between the original outcome and a replay is attributable to the agent, because the information set is held constant across both runs.
Backtests stop scoring the agent on restated figures and post-dated text, so measured performance is closer to what the deployed system would have achieved.
Silently mutating dependencies surface as unresolvable-version failures instead of as quiet answer drift.

What it costs you

Every pinned source needs retention long enough to cover the dispute window, and storing addressable history for market data, indexes and model snapshots costs money that grows with the window.
A pinned model snapshot ages: it misses later fixes and capability gains, and pinning to a training cutoff trades accuracy for temporal validity.
Sources outside the operator's control often expose no as-of read at all, so the information set has to be copied and versioned locally before it can be pinned.
The pin is a second thing that can be wrong; a manifest recorded from a resolver that quietly fell back to the live value is worse than no manifest, because it looks like evidence.

What this pattern forbids. A dated decision must not read any source at its live current value: every read resolves only through the pin manifest for that decision's as-of timestamp, no source may be pinned for less than the window in which the decision can be questioned, and a source that cannot serve the recorded version fails the run rather than falling back to today's data.

And the patterns that stand alongside it, or against it —

complementsDeterminism-Tiered Replay Gate·— Classify an agent into a reproducibility tier by re-running identical inputs, require the strictest decision-determinism tier for regulated decisions, and gate deployment and validation-sample size on the measured tier.
complementsReplay / Time-Travel★★— Re-run a past agent trace from any step with modified inputs/prompts/tools to debug or branch.
complementsLineage Tracking★★— Track which prompt version, model version, and data sources produced each agent output.
complementsJournaled LLM Call★— Record the output of every non-deterministic step on first execution and replay that recorded value during crash-recovery instead of re-invoking the model.
complementsDurable Workflow Snapshot★— Capture workflow execution state as a snapshot in a pluggable storage provider so a paused run can resume across deployments, process restarts, and host crashes.
complementsServing-Stack Attestation★— Treat the served model, inference backend and generation defaults behind an endpoint as an unverified claim, and establish them by black-box behavioural probe recorded alongside every result.
complementsEval Harness★★— Run a held-out dataset against agent versions to detect regressions and measure improvement.
complementsMemo-As-Source Confusion✕— Anti-pattern: the agent cites its own past memos as ground truth instead of re-verifying them against the artifacts they describe, accumulating false confidence in stale summaries.
complementsNow-Anchoring·— Ground the agent's reasoning in the current absolute time without requiring tool calls, so every reply is implicitly time-aware.