Evidence before capital.

FactorHelm keeps the record of an investment decision: the sources a thesis rested on, the company's real economic exposure, the tests it passed and failed, and who approved it. The record is fixed and signed before money moves, then measured against what happened.

For research, investment-operations and model-risk teams starting to use AI-generated research.

Read the CrowdStrike replay Request access

The seal of one decision record, drawn from its SHA-256 hash, which is printed around the rim. Change one character of the record and both change. Try it.

AI writes investment theses faster than anyone can check them.

A research agent can draft fifty ideas before the morning meeting. It can't tell you which sources it used, whether they existed at the time, how much the company actually earns from the story, or what would prove the idea wrong.

Those are the questions an investment committee, a risk team or an auditor asks months later. FactorHelm answers them at the time, in a record nobody can quietly edit afterwards — including us.

One record per decision

This is the decision-time record for a historical replay: CrowdStrike, 23 February 2026.

Two of its four gate inputs were not reconstructed, so the record says so. A missing input is never shown as a pass, and a replay is never shown as a live call.

See the replay step by step, with a movable information cutoff

CrowdStrike (CRWD)

Thesis: Negative overreaction to an AI product launch with little revenue overlap. Expect partial reversion relative to SPY.
Information cutoff: 23 Feb 2026, 16:00 New York (21:00 UTC). Nothing published later can enter the record.
Method: Event-study rule, pre-registered: day 0 is the largest abnormal session within 5 days of the recorded date; graded only if |z| >= 2; thesis holds if 20-session reversion >= 0.5.
Execution authority: None. A research decision is not permission to trade.

Gates at the cutoff
Gate	Input	Requirement	Result
Market shock	abnormal -8.8%, z -3.7	\|z\| >= 2	Pass
Economic exposure	disruption overlap 0.15	low revenue overlap with the narrative	Pass
Profile written after the event. A prospective call needs a profile dated before the catalyst.
Source corroboration	—	>= 2 independent publishers	Missing
Not reconstructed for this replay.
Narrative velocity	—	velocity z >= 2	Missing
Not reconstructed for this replay.

Decision: qualified for grading

Qualifies under the event-study rule. Would not qualify as a strict call: two gate inputs were not reconstructed.

Recordfh-replay-crwd-2026-02-23SHA-2561149a5b87e710220ed9ae056283a76a950b2705fb4dd7bf084733866861e0df5

Outcome after 20 sessions, to 23 Mar 2026

Abnormal drift after the shock day: +9.1% at 5 sessions, +20.9% at 20 sessions. Reversion 2.37 against a threshold of 0.5, so the thesis held under the event-study rule. The outcome is a new entry that points to the decision record's hash; it does not edit it.

Points to1149a5b87e710220…SHA-25697dd63ac1dd7b891e458a930a0d59bdaa30adc803bc4b028fc364b5b70e2a60e

Recompute both hashes in your browser

From a thesis to a signed record

Observe

News, filings and creator claims about 1,022 public companies are collected and timestamped. Each story is measured two ways: how fast it is spreading, and how much of the company's revenue it actually touches.

Signals and Insights
Validate

A separate validation path tries to break the thesis. It rebuilds the data as it stood at the cutoff, compares the idea with simple baselines, and keeps every test it ran, including the failures.

Validation path
Decide

A named person approves, conditions or rejects the decision for a stated use and time window. Agents can gather evidence and argue against a thesis. They cannot approve one.

Decision record
Record

The record is hashed into a daily manifest, signed with a cloud key and stored where it can't be edited. Outcomes are added later as new entries. They never rewrite what was known.

Portfolio and scorecard

How each step works, and what each product does

What we can show today, and what we can't yet

Established

Outcome grading is frozen in a written contract: 20 NYSE sessions, adjusted closes, SPY as the benchmark, every price stored with a hash.¹
Strict calls fail closed without a canonical hash. Daily manifests are signed with an AWS KMS key and kept in S3 Object Lock, which no one can shorten or delete.²
1,022 companies are tracked. 151 have a human-reviewed revenue-exposure profile; 107 more are provisional and cannot trigger a strict call.³
Our first strict call, on Alphabet in July 2026, was invalidated for data quality. It stays on the public record, marked invalid.⁴

Not yet established

A graded live track record. No strict call had been graded as of 30 September 2026.⁵
A hit rate. None will be published before 30 independent catalyst families have resolved.⁶
Independent reproduction. Our validation path is owned by the same people who build the research, so we don't call it independent.
That any of this makes money. We don't claim it, and the site won't until a graded record supports it.

The scorecard, in its current state

What we won't claim

That a historical replay is a track record.
That a research approval is permission to trade.
That more agents or more data make a thesis true.
That our own checks are independent certification.

The full list

A six-week pilot

One universe, one catalyst family, one to three of your people.

You get a decision record for every candidate in scope, a weekly review of false positives and missed events, and an audit bundle you can hand to risk or compliance. There is no performance promise, no execution, and your data is not used to train anything.

See the pilot

Sources

Outcome grading contract ne-grade-v1, NarrativeEdge methodology. See Methodology.
Provenance v1, NarrativeEdge. See Methodology and Verify.
Signals dashboard user guide, production universe section, September 2026.
NarrativeEdge technical strategy v1.2, 14 July 2026. The call was invalidated because the company match came from mentions of unrelated peers.
FactorHelm execution baseline, 30 September 2026. The live scorecard is authoritative.
Publication rule shared by every surface: Signals, public scorecard, email and API.