Source-rich work
Projects, writing, research, founder material, public talks, publications, repositories, or case notes.
Design partner alpha
Marrow is for people whose professional identity is source-rich, nuanced, and hard for generic AI to represent accurately. The alpha is a focused proof sprint around real work, real evidence, and real voice.
Who this is for.
The strongest partners have high-value professional identities where generic AI loses the evidence, nuance, or voice that makes the work specific.
Nuanced representation
Work that generic AI tends to flatten, overstate, or describe without the right caveats.
Trust-sensitive output
Profiles, bios, proposals, applications, or agent answers where evidence and restraint matter.
What the alpha includes.
The sprint is narrow on purpose. Marrow stores and retrieves the source-backed corpus; products and agents decide what to generate or render.
Build a small source corpus
Use selected files, writing samples, project notes, and public links rather than a large archive.
source packet
- CV, profile, or about page
- 2-5 writing samples
- 2-5 project notes
- public links or publications
- one target output promptGenerate evidence-backed outputs
Create one professional output and one writing-style output, then compare them against generic AI output.
Marrow retrieves:
- source notes
- claims and entities
- style observations
- provenance
- correction boundariesReview evidence and corrections
Inspect which records supported the output, then accept, narrow, reject, or correct claims.
review result
accepted claim
narrowed claim
unsupported claim
missing nuance
durable correctionWhat is out of scope.
This is not a finished dashboard. The alpha does not include a public lookup profile, a general assistant, or autonomous write-back. Those may become product surfaces later, but the current sprint tests whether the source-backed corpus is useful and trustworthy.
No public profile lookup
Private source material stays private unless the partner chooses what to share.
No autonomous writes
Corrections become durable records only after review and acceptance.
No archive import
The sprint starts with a small source packet instead of a large historical dump.
No dashboard promise
The first proof path is CLI/API-first so the data contract is real before UI work.
How outputs are evaluated.
The useful evidence is not only whether an output sounds polished. The useful evidence is whether the person trusts it more because they can inspect why it said what it said.
Accuracy
Important claims are correct, scoped, and backed by source records.
Specificity
The output preserves the work's concrete projects, artifacts, judgment, and outcomes.
Voice
The writing uses style evidence without inventing new facts.
Corrections
Accepted edits become durable source-backed updates instead of hidden prompt memory.
Trust questions we want answered.
These questions decide whether Marrow should become a deeper product relationship, not just a useful demo.
Traceability
Does seeing evidence change whether the partner trusts the output?
Privacy
Which claims should be public, private, cited, summarized, or withheld?
Evolution
Do partners want corrections and new evidence to improve the corpus over time?
Attestation
Which claims would benefit from a coworker, client, collaborator, or institution confirming them?
The strongest signal is trust.
A good design partner should be able to say: this represents me better than generic AI, I understand why it said that, and I want to keep improving the corpus.
Request a design partner slot