The problem Calibra solves

The models for predicting flight outcomes used by insurance companies, airports, and airlines require significant amounts of upfront capital for models that have no method of new model discovery or adaptation. Subpar models that are expensive to replace and difficult to compare and find the true quality are used frequently in the industry.

This is part of a broader problem where institutions rely on existing prediction models because they do not have clear, objective comparisons with other prediction models and because the cost of switching models is too high to justify the uncertainty.

Calibra’s Solution

Public competition is the best solution for problems like this, and Calibra offers a venue for public judging and new model discovery that is objective and secure while simultaneously aggregating predictions for the user, which is more accurate than a single model in the long run.

The user interested in determining the probability of flight events begins by funding a batch of flights with a bounty. At fund-time, the user states the time window during which the user is interested in receiving probabilities, and a set of random timestamps during the time window is generated to use in judging following the conclusion of all flights in the batch. A public encryption key and private decryption key is generated for the batch.

Following batch funding and before the end of the prediction window, anyone can post a bond equal to sqrt(bounty) to join a batch and compete for the bounty. Predictors can commit probabilities on-chain for each of the events in the batch at any time during the prediction window, encrypted using the batch’s encryption key. The funder account is the only one able to decrypt the submissions through a wallet signature.

Following the conclusion of all flights, the predictors are required to reveal their true submissions in order to be eligible to claim the bounty and keep their bond. True submissions at the randomly selected timestamps are revealed for public auditing and each predictor receives a score based on their predictions. The most accurate predictors split the bounty and are refunded their bond while the least accurate predictors see their bonds slashed.

Challenges I ran into

The hardest problem was submission privacy: I needed predictors to commit forecasts on-chain without revealing them to other predictors (or the public) during the prediction window, while still guaranteeing that submissions couldn’t be changed later. I solved this with a commit–reveal scheme in CalibraProtocol where providers first post a hash commitment keccak256(batchIdHash, root, salt) during the window (commit()), and only after the window ends they reveal the root + salt (revealCommits()), which the contract verifies matches the original commitment before accepting it.

To prevent people from “reading” each other’s forecasts off-chain, the actual prediction payload is not posted in plaintext during the commit phase; instead, providers upload the encrypted payload off-chain and only post an encrypted URI hash on-chain (emitted as encryptedUriHash in the Committed event). The funder/operator publishes a batch-specific encryption public key at batch creation (stored as funderEncryptPubKeyHash), so predictors can encrypt their submissions for the funder while keeping them private from everyone else; later, during the reveal phase, providers publish a publicUri and the contract stores a hash of it (publicUriHash) so the revealed reference is still tamper-evident. This combination gave me the properties I needed: privacy during the window, immutability via commitments, and verifiable reveals after the deadline.

Some Problems to Solve in Future Iterations

Currently, Calibra’s infrastructure is not made for high-frequency submissions due to the resulting high gas costs. The judging process is done off-chain by the operator and put on chain for public auditing, but this could be improved with an optimistic method of judging. Calibra could also be improved for institutional use by increasing the granularity of the batches - specifically, allowing for different flights within the same batch to have different probabilities that the predictors must submit would be beneficial. While increasing the bounty size does increase the number of models willing to predict for a batch, there is not a clear enough association between the number of flights in a batch and the corresponding bounty or objective measures of quality in relation to the size of the bounty.

Use of AI tools and agents

Calibra does not currently use AI tools or autonomous agents within the protocol itself.

The system is model-agnostic: it provides infrastructure for submitting probabilistic forecasts and evaluating their accuracy after real-world outcomes occur. Forecasts may be generated by humans, statistical models, or external machine learning systems, but the protocol does not require or run any AI internally.