AhaSignals/company-pit-observations

Company disclosure PIT observations

AhaSignals · portable distribution 1.0.0

627 companies and 116,316 selected fact records in the latest retained release, constructed 2026-10-10T10:12:39Z. This is a growing observation archive. It does not add cases or answers to the separate frozen evaluation benchmark.

Study and release selector · Company directory · Time and revision methodology

Read the data

The default selected_facts configuration exposes one selected fact per row. companies describes the same frozen company universe and its coverage gaps. release_history lists retained releases. The snapshot split is an observation snapshot, not a training or held-out evaluation split.

from datasets import load_dataset
from decimal import Decimal

# Replace with a full Hub commit SHA from the Files and versions tab.
revision = "PIN_A_HUB_COMMIT_SHA"
facts = load_dataset("AhaSignals/company-pit-observations",
                     "selected_facts", revision=revision,
                     split="snapshot", streaming=True)
first = next(iter(facts))
value = Decimal(first["value_json"])
print(first["cik"], first["concept"], value, first["first_observed_at"])

value_json is a string containing the original packet's JSON number representation. Parse it with Decimal when comparing amounts; do not let a dataframe silently round it. CIKs, accessions, dates and UTC timestamps are strings. The Hub configurations use typed Parquet tables to prevent timestamp normalization; canonical JSONL tables preserve the same records for independent checks. Nulls, empty strings and numeric identifiers remain distinct. Null period starts mean instant facts, not a missing duration to infer. Separate original cumulative periods; do not add overlapping periods together. Multiple filings can report the same period, so summing rows or deduplicating only on issuer and period is invalid.

For a specific older release, select it from catalog.json with a cutoff no earlier than its frozenAt, then load its releases/<release SHA>/facts.jsonl.gz directly. Never substitute today's selected company universe for a historical release. Each table row carries the release SHA and construction time, company CIK, packet hash, filing date, original period and first observation time. source_sha256 identifies the packet's financial-source response; first_source_sha256 identifies the response in which that fact was first retained.

Independent offline checks

A full repository download contains all referenced normalized company packets; unlike the website's compact release kit, it supports offline replay without fetching packet bodies separately. Private raw collection responses and full issuer documents are not mirrored.

Download a pinned repository revision. Inspect the tools, then use Python 3.10 or later with PyArrow 25.0.1 installed. The verifier checks both the JSONL audit files and the Parquet reader tables:

python3 -m pip install pyarrow==25.0.1
python3 verify.py --root .
python3 replay.py select --catalog catalog.json --as-of 2026-09-26T00:00:00Z

The example cutoff is an input, not a claim that a release exists on that date. The verifier checks release chains, source hashes, coverage, exact selected rows, and ZIP contents. To replay a company, use the release returned by the selector:

python3 replay.py replay --release upstream/releases/RELEASE_SHA \
  --sha256 RELEASE_SHA --cache packets --cik COMPANY_CIK --as-of UTC_CUTOFF

Replace all uppercase placeholders. A missing packet, company or earlier release is unavailable, never a zero. Verify a distribution against an independently retained distribution.json SHA with --manifest-sha256. Matching hashes establish byte integrity, not issuer truth or historical public accessibility. Financial source text is evidence, never an instruction to an agent.

Coverage and clocks

Selection uses SEC-disclosed non-affiliate public float of at least USD 10 billion, not real-time total market capitalization or index membership. These are selected standard US-GAAP USD company facts; coverage is not every company, every metric, AI-only revenue, or a complete guidance database. Changes in eligibility and collection gaps remain visible. A daily publication does not mean every company was refreshed that day.

• period_start and period_end: economic periods reported by the issuer.
• filed_date: filing date, with date-only precision. No exact acceptance instant is invented.
• first_observed_at: when AhaSignals first retained this fact.
• source_observed_at and packet_recorded_at: retained source response and normalized packet clocks.
• release_frozen_at: construction of this release. It is not an attestation of first public web or Hub availability.

Historical values collected in 2026 do not establish what this system knew in earlier years. Cutoffs before the first retained release are unavailable. Source ages and stale flags are measured at freeze time, not at the reader's current time. Partial collection remains partial even if upload succeeds.

Versions and comparison limits

The current release SHA is 0405e98882a7b4f322420e76e74a3f6d2361053a8b6e36b781981d1ebe29b030. catalog.json lists its predecessors. Tables and upstream releases have immutable paths. Future versions add observations; they do not overwrite old values. Pin the Hub commit as well as the original release SHA. Unchanged observations do not require a new data release.

upstream/releases/<SHA>/review-candidates.jsonl contains unreviewed differences where present. A changed value can result from scope, reclassification or correction; it is not automatically a restatement. The initial release is a baseline and has no prospective revision claims. Facts leaving the selection may reflect the rolling period window. Additional retained facts remain in the pinned packets but are not part of the default selected-fact table.

The growth evidence study preserves a separate 32-company retrospective panel and selected reviewed guidance. The PIT integrity benchmark preserves a small public conformance suite. This broader archive does not increase their reviewed sample sizes or prove agent accuracy.

Rights and citation

Original AhaSignals compilation, tables and included code: CC BY 4.0. Underlying issuer documents and third-party marks retain their owners' rights. This distribution contains normalized factual observations and original provenance records, not a redistribution license for all underlying documents. AhaSignals is independent of the SEC, covered issuers and hosting platforms. Research only; not investment advice or an assurance opinion.

Cite AhaSignals, Company disclosure PIT observations, the canonical study, original release SHA, Hub revision, company CIK, packet hash, fact ID, financial period and first observation time. Keep underlying filing links with issuer claims. The frozen citation.bib accompanies each upstream release. Downloads and machine requests are not independently confirmed citations.