Evidence-to-Artifact Flow: One Evidence Object Across Briefs, Outreach, Captions, Paid Reuse, and Reports

The five artifacts an AI creator program ships

Most teams describe the work as “running a campaign.” The system, in practice, emits five distinct artifact classes. Each one is a surface on which the same evidence object should appear without contradiction.

• The brief — What the creator is being asked to make. Includes the product claim, the audience, the angle, the call to action, the rights scope, the metric the post is expected to move.
• The outreach line — What the creator is being told before they decide. Includes the pitch, the proof of fit, the offer, the timing, the prior performance the recommendation is based on.
• The caption review — What the post is being approved against. Includes the claim wording, the disclosure, the brand voice constraints, the legal carve-outs, the evidence that backs each claim in the caption.
• The paid cutdown decision — Whether the organic content gets reused as paid. Includes the rights scope, the platform allow-list, the audience targeting fit, the realized organic performance, the residual term.
• The report row — What the campaign told the brand. Includes the outcome, the attribution method, the evidence pointer that justified the spend, and the confidence on the recommendation that drove the post.

The bug in most AI creator stacks is that each of these is generated against its own context window. The brief calls a SKU “the bestseller”; six weeks later the report row says “number three in the category”; the paid cutdown calls it “new.” All three are AI-generated; all three are pointed at the same product; none of them share a record.

What propagates and what does not

Not every field on the evidence object belongs in every artifact. The propagation rule is strict: each artifact reads only the fields it is allowed to express, and writes back only the fields that grade or expire it.

• Brief reads — claim, source class, source URL, expiry, rights scope, approval state, the recommendation contract’s target metric and direction.
• Outreach reads — claim (in soft phrasing), creator-fit evidence, prior post performance pointer, offer terms, confidence on this creator archetype.
• Caption review reads — claim wording, disclosure rules, legal carve-outs, source-class precedence (regulated claims override marketing claims), expiry.
• Paid reuse reads — rights scope, platform allow-list, residual term, realized organic performance, source class for any on-screen claim.
• Report reads — recommendation contract (full), evidence pointer, realized movement, attribution method, confidence after grading.

Writes go the other way. Caption review writes back disclosure approval state. Paid reuse writes back the residual countdown. Report writes back realized movement and confidence delta. The brief and outreach are read-mostly; the rest amend the record.

The propagation walkthrough

Take one product claim, ship it through the whole pipeline, and watch the evidence object stay in the center. The walkthrough below uses a regulated skincare claim because regulated claims expose the bug fastest.

Step 1 — The claim becomes an evidence object

A clinical study supports a 28-day texture-smoothing claim. Source class: regulated lab study. Source URL: study PDF in the rights-managed claim library. Approval state: legal-approved. Rights scope: organic plus paid social, 12 months. Expiry: 2026-11-30. Owner: the brand’s regulatory lead. This is the evidence object. Nothing downstream invents this; everything downstream points to it.

Step 2 — The brief expresses the claim

The AI brief generator pulls the approved wording, the disclosure requirement, and the expiry. The brief renders the claim verbatim in the section the creator must read before posting. The brief stores a pointer back to the evidence object, not a copy of the wording. If the wording is amended, the brief reflows on next view.

Step 3 — The outreach uses a softer slice

The outreach drafter cannot express regulated claims directly. It reads the claim’s category, not its wording, and writes a pitch line that signals what the brief will require without making the claim itself. The evidence pointer is attached to the outreach record so legal can audit the linkage if a creator asks why the brief is strict on phrasing.

Step 4 — The caption review enforces the wording

The creator submits a caption that paraphrases the claim into something slightly stronger. The AI caption review reads the evidence object, sees source class “regulated,” and rejects any phrasing that does not match the approved wording within a defined tolerance. The reviewer approves the corrected caption; the approval state is written back to the evidence object.

Step 5 — The paid cutdown checks the rights and the claim simultaneously

The post does well organically. The paid system proposes a cutdown. It reads the same evidence object: rights scope says paid social is allowed within 12 months; expiry has not been reached; the approved wording is still the wording on-screen. The cutdown ships, with a pointer to the same evidence object, and the residual countdown is written back so the paid system knows when to stop running it.

Step 6 — The report row defends the spend

The brand asks why the paid spend was approved against a regulated claim. The report row points to the evidence object. The evidence object points to the lab study, the legal approval, the rights scope, the expiry, and the wording the caption was forced to use. The defense takes one query, not a folder search through Slack and email.

What happens when the evidence object changes

The interesting case is not the happy path. It is what happens when the record changes mid-campaign. Three change classes are common; each one fans out through the artifacts predictably.

• Wording amendment — Legal narrows the approved wording. The brief reflows on next view; outreach is unaffected because it never expressed the wording; pending captions are rejected against the new wording; running paid cutdowns are flagged for review; report rows are stamped with the wording version active at post time.
• Expiry — The evidence object expires. Brief generation halts; outreach pauses; pending caption reviews are rejected; running paid cutdowns are taken out of rotation; report rows annotate the claim as withdrawn so the historical record stays honest.
• Rights downgrade — Rights scope is reduced from paid plus organic to organic only. The brief and outreach are unaffected; the caption review is unaffected; the paid system removes the cutdown; the report row notes the spend ceiling change without retroactively erasing prior valid spend.

Each one is a single write. The fan-out is automatic because every artifact carries a pointer, not a copy. Programs that copy the wording into each artifact have to chase down every artifact by hand when the record changes, and they always miss one.

Why pointers beat copies

Pointers are not a styling choice. They are the difference between an AI program that the brand can defend and one that ships unbounded liability. Copies look the same on the day they are written. They diverge silently over the next six weeks as the claim is amended, the rights are renegotiated, the expiry approaches, and new reviewers join the workflow.

• A copy cannot expire — Once a claim is rendered into a brief as a string, no expiry can reach it. The brief keeps showing the old wording on every refresh until a human re-generates it.
• A copy cannot be downgraded — Rights scope changes on the evidence object do not reach into the artifacts that already copied the claim. The paid system keeps running the cutdown until somebody notices.
• A copy cannot be audited — When five artifacts each have their own copy of the claim, no audit can ask ‘which evidence justified this?’ The answer is reconstructed, not retrieved.

How the flow pairs with the rest of the system

Evidence-to-artifact flow is the runtime expression of the record. The record itself is defined by the Campaign Evidence Object schema and lives inside the campaign evidence control plane. The recommendation that drove each artifact is graded by its expected-metric contract and approved through the approval gates pipeline.

The artifact end of the flow connects to the operating surfaces the team already uses: the AI brief generator workflow for briefs, the outreach pre-flight simulation and outreach artifact provenance for outreach, the AI creative QA workflow for caption review, the AI-generated creator ad variations for paid reuse, and influencer campaign reporting for the report row.

Pair with AI creator marketing source of truth so the program treats this propagation as the spine, not a feature, and with the accurate, approved, measurable operating model so the three failure modes the flow prevents are explicit at every gate.

Where Storika fits

Storika is the operator surface where the flow runs. Every artifact the platform generates carries a pointer to the underlying evidence object. Amend the wording, narrow the rights, advance the expiry, downgrade the approval — briefs reflow, outreach pauses, caption reviews reopen, paid cutdowns rotate out, and report rows annotate, all from a single write.

The team works at AI speed without losing the ability to defend any single artifact in a single query. That is the bar a $100M brand expects, and the bar that distinguishes a creator program from a content factory.

Mistakes to avoid

Mistake 1: Rendering claims as strings into briefs

Once the claim is a string, the evidence object cannot reach it. Render at view time from a pointer, never bake the wording into the brief document.

Mistake 2: Letting outreach drafters express regulated claims

Outreach should signal category, not claim wording. Regulated source classes must be unable to be quoted in the outreach surface.

Mistake 3: Approving captions without a pointer write

A caption approval that does not write the approval state back to the evidence object cannot be propagated. The paid system has to re-approve from scratch.

Mistake 4: Treating paid reuse as a separate rights decision

Paid rights live on the same record as organic rights. A program that decides paid reuse from a spreadsheet outside the evidence object will run cutdowns past expiry. Pair with usage rights pricing.

Mistake 5: Reporting on artifacts instead of evidence

Report rows that point to the post instead of the evidence cannot defend the spend. The post is the expression; the evidence is the justification. Reports should resolve to evidence on click.

FAQ

One record, five surfaces