The Lost Matches

Reconstructing sports history without pretending the record is complete

R. Cole Peterson

rcolepeterson.com · rcolepeterson@gmail.com

The idea

The Lost Matches is a web experience for reconstructing historic sporting events that were never fully captured on film.

Instead of presenting a seamless documentary, it assembles the surviving fragments: official footage, fan photos, ticket stubs, audio, written memories, and editorial context. Visitors move through a chronological timeline, seeing both what has been recovered and what remains missing.

That incompleteness is visible by design. Every event carries a restored percentage, making it clear that the experience is a reconstruction—not a definitive record.

The current prototype includes:

• The Battle of the Sexes (1973): a fuller demonstration of the timeline and content system.
• Miracle on Ice (1980): a second event showing how the format translates to another sport.
• Thrilla in Manila (1975): a locked “coming soon” entry that signals how the platform could expand without overbuilding the prototype.

Where the project came from

This was not a brand-new idea. My team first pitched an earlier version to the BBC roughly four or five years ago, then explored it again around a potential tennis opportunity two or three years later. Neither moved forward.

I still thought the underlying product question was worth exploring, so I rebuilt the concept independently last week as a working prototype.

The current version is entirely hardcoded. There is no database, live contribution system, user authentication, or moderation workflow behind it yet. That was intentional: the goal was to make the idea tangible enough to test the experience and expose the harder product questions before investing in infrastructure.

The real opportunity

I do not see a large, obvious standalone consumer market for this in its current form, and I do not want to manufacture one for the sake of a case study.

The more credible opportunity is as a feature or editorial format inside an existing sports, broadcast, archive, museum, or rights-holder platform—somewhere that already has an audience, historical material, and a reason to deepen engagement with its archive.

The underlying problem is real but specific:

• Important events often survive as incomplete records.
• Valuable context remains scattered across official archives and personal collections.
• The people holding firsthand memories will not be available forever.
• Most archive experiences are built to publish finished material, not to collect and reconcile fragments over time.

The Lost Matches asks whether an incomplete archive can still be compelling—provided the product is honest about what it knows, what it does not know, and where each contribution came from.

What AI accelerated—and what it did not

AI made it faster to turn the concept into a functioning interface. It did not remove the need for product judgment, editorial scrutiny, or quality control.

The most revealing problems were not dramatic engineering failures. They were small moments where the generated version technically worked but was still wrong.

A blurry image that was not a CSS problem

The hero image looked soft regardless of how I adjusted the blur treatment. The problem was upstream: the source image was only 702 × 860 pixels and was being stretched across a full-screen container. No CSS adjustment could restore detail that did not exist.

The lesson was simple: inspect the input before continuing to tune the output.

A “Written Memory” card that was not a memory

One timeline card was labeled and styled as a personal recollection, but the copy was actually platform-written editorial context. Nothing was technically broken. The page rendered correctly.

But the design borrowed the emotional credibility of a real eyewitness for content that had no eyewitness behind it. That made a content-model problem look like a visual-design problem—and created a subtle trust failure.

The fix was to reserve first-person treatments for genuine first-person material and give editorial narration its own clearly labeled presentation.

A video interaction with two meanings

A timeline card included both a play control and a separate “Watch on YouTube” action. One kept the visitor inside the experience; the other sent them to another platform. The distinction was not clear enough.

Again, the feature worked. The issue was whether the interaction communicated its consequence before the user clicked.

None of these problems was difficult to fix once identified. The real work was noticing them instead of accepting “it renders” as evidence that the experience was right.

Key product decisions

Show uncertainty instead of hiding it

Each match displays a restored percentage rather than presenting the reconstruction as complete. The number makes uncertainty part of the interface, not something buried in a disclaimer.

In a production system, that percentage would need a transparent methodology. In this prototype, it establishes the intended product behavior: missing evidence should remain visible rather than being silently filled with generated content or presented as fact.

Keep provenance close to the content

The timeline distinguishes official footage, fan material, written memories, and platform context where they appear. The source is part of the experience because the meaning of a fragment depends on who supplied it and what kind of evidence it is.

A polished interface should not flatten those distinctions.

Reserve emotional design language for real voices

Pull quotes, quotation marks, and first-person framing create intimacy and credibility. I treated those choices as part of the trust model, not decoration.

Personal visual language belongs to genuine personal testimony. Editorial context needs to look and sound editorial.

Demonstrate scale without overbuilding

I chose to build two event pages and show a third as “coming soon.” Fully populating every event would have spread the work across several shallow examples. Removing the third would have weakened the signal that the system could apply across sports.

The right scope was determined by the purpose of the artifact: demonstrate a repeatable format, not imitate a finished content library.

Design for graceful failure

External media will eventually disappear, embeds will break, and contributed files will vary in quality. The prototype uses styled fallback states rather than leaving dead frames in the timeline.

A reconstruction platform should remain understandable even when one piece of evidence is unavailable.

Use one deliberate “wow” moment

As visitors scroll, the full-screen background changes to reflect the timeline moment currently in view. One moment can use looping video rather than a static image.

That interaction gives the archive a sense of presence without requiring every entry to become an expensive cinematic sequence. The effect supports the story rather than becoming the story.

Trust and uncertainty: what exists now

The prototype currently demonstrates:

• A visible completeness indicator instead of a false sense of “finished.”
• Clear separation between official material, contributed material, and platform narration.
• Source and contributor labels attached to timeline fragments.
• Graceful fallbacks when external media is unavailable.
• A corrected visual-language rule so editorial context does not masquerade as personal testimony.

These are useful foundations, but they are not yet a complete trust system.

The prototype does not currently handle:

• Two contributors giving conflicting accounts of the same moment.
• Claim-level states such as confirmed, disputed, unverified, or unknown.
• Corrections when published information later proves inaccurate.
• A transparent method for calculating archive completeness.
• Contributor identity, reputation, or corroboration.
• Rights, consent, and reuse permissions for submitted media.

That gap is the most interesting part of the concept. The difficult question is not simply, “Can people contribute?” It is: Can the system remain useful and honest when evidence is incomplete and contributors disagree?

What I would build next

1. A real evidence model

Move the hardcoded content into a database where each fragment has a source, owner, timestamp, rights status, evidence type, and verification state. Claims and media should be related without being treated as the same thing.

2. Accounts and contribution flows

Allow people to submit memories and media under a persistent identity, explain what they know firsthand, and provide enough provenance for moderators to assess the contribution.

3. A moderation queue

Nothing contributed should go directly onto the public record. Reviewers need tools to inspect sources, request clarification, check rights, and approve, reject, or hold material.

4. A dispute model

When accounts conflict, the system should preserve the disagreement rather than select a convenient version or merge incompatible details into false certainty. Competing accounts can coexist if their status and sourcing are clear.

5. Visible corroboration

Visitors should be able to see when a claim is supported by multiple independent contributions, an official record, or both. Trust should be inspectable rather than implied by confident copy.

6. A defensible completeness score

The restored percentage needs rules: what counts as a timeline moment, how different evidence types are weighted, and what changes when material is disputed or removed. Without that explanation, the number is useful as a concept but not yet authoritative.

Outcome

The earlier BBC pitch and later tennis-related opportunity did not become projects. This independent rebuild is the first time I have turned the idea into a working, publicly accessible product experience on my own terms.

The outcome is not a launched business. It is a tangible artifact that makes the opportunity—and its unresolved questions—specific enough to evaluate.

It demonstrates how I approach AI product work:

• Build the smallest credible version that makes the idea testable.
• Treat uncertainty and provenance as interface decisions.
• Review generated output for meaning, not just functionality.
• Be explicit about what is real, simulated, incomplete, or not built yet.
• Use the prototype to discover the next product questions instead of pretending they have already been solved.

Takeaway

AI dramatically shortened the distance between an old idea and a working prototype. But speed was not the most important part of the exercise.

The value came from deciding what the product should claim, where it should show doubt, how it should distinguish evidence from narration, and when the right move was to stop building.

The prototype proves the experience. The unresolved trust model defines the real product.