# Experiment 03: predictions before observations Protocol: `predictions-v1`. Exploratory, selected before the first study run. No kernel changes, policy tuning, randomness, or autonomous family selection. ## Question and selected rules Does continuing the line through the two most recent known samples help predict the next observation? When does holding the last value work better? Reuse the existing `prospective-calibration.gojo` fixture and `LocalCalibration.Forecast` readers. For each scheduled target: 1. Give the participant only the target's source time, not its value. 2. Select the two latest concrete receipts in its connected history. Missing receipts remain open and are skipped for fitting. With fewer than two known samples, record insufficient history and issue nothing. 3. Fit the selected affine frame and issue the forecast. Keep its selected receipts, frames, expectation slot, residue and local issue event. 4. Only then admit the scheduled observation, either a signed integer quantity or an open placeholder. Source time and local event order are separate axes. 5. Ask the existing assessment reader to compare the issue with that receipt. Record refusal for an open outcome; never substitute zero or a later sample. The participant receives one target at a time. The host owns the finite input tape and chooses this fixed policy; it supplies no errors, baseline predictions, case labels, regime labels or future values to the participant. The output readout and encoding are identity frames. A flat fitted line is legal and retains the unrepresented time input as residue. Expectation remains open; an exact conditional prediction does not imply confidence about the world. The structural spine is already implemented: ```text stream -> state(receipts, issues, assessments) two available receipts -> fit witness(time frame -> quantity frame) fit + readout + open expectation -> future issue witness later matching receipt + original issue -> signed mismatch witness before state -> request/witness -> after state -> timeline ``` Selecting two receipts is authored focus, not learned attention. Missing samples can leave an old line in use for several targets. Record the source-time gap from the newest training receipt and do not claim that such a line is current knowledge. No maximum-age cutoff, confidence update, pruning, imputation, feedback response or model-family switch is introduced in this experiment. ## Cases and split `PredictionProtocol.Cases` and the copied `cases.json` are the exact input tapes. Three development sequences check constant, rising and fractional continuation. Nine evaluation sequences use different values and cover a negative constant, unequal-time ramp, a reversal, alternating values, a step, gaps across a change, missing initial observations, a final blackout, and an entirely missing stream. These are hand-authored finite examples, not random samples of an application population. The policy is fixed before either split runs. Development cases validate the apparatus; evaluation cases are run only afterward. Once inspected, these evaluation tapes are no longer unused data for a revised policy. All cases and any apparatus failures stay in the run directory. ## Comparisons and measurements - An independent conventional exact rational implementation uses the same two-sample rule, information order and source times. It has its own history and no Gojo readers. Compare availability, selected times, predictions, errors, flat-frame residue indication and unresolved outcomes on every target. This is a numerical check, not a prediction contest or a throughput benchmark. - External hold-last issues the most recent known value before seeing the next observation. It needs only one known sample. It is explicitly an external baseline, not a newly implemented Gojo participant. - The no-prediction control receives the same samples and issues nothing. Its coverage is zero and error is unavailable, not zero. It has no claim to outperform a method merely by abstaining. - Report both methods' issued and scored counts over all scheduled targets. Compare mean absolute errors only on targets where both issued a prediction and an outcome is present. Also report hold-last's wider scored coverage. Empty denominators produce unavailable results. - Keep individual signed errors (`observed - predicted`), absolute errors, exact hits and worst absolute errors. Opposite misses cannot cancel the loss. - Record missing observations, forecasts awaiting an outcome, insufficient history, fit/issue/assessment refusals, newest-sample age, retained receipts, issues and assessments, and calls to the four Gojo APIs. Calls are a work proxy; they are not comparable CPU work or measurements of energy. All prediction arithmetic and signed mismatch reads on the participant path come from the existing connected Gojo frames. Integers represent authored sensor values and event coordinates. Exact fractions used by the reference and metric aggregator belong to the external evaluation boundary and never feed back. JavaScript floating point is confined to drawing the captured results. ## Records and validation Write this protocol, fixture and all tapes to a new directory before running. Retain each target's record immediately, including predictions before its outcome is admitted. Preserve partial event logs if an apparatus exception occurs. Save all completed traces, external comparisons and summaries; refuse to overwrite a nonempty run directory. One run per case, no random variation claimed. JSON is a diagnostic projection, not a restorable graph or exact replay contract; connected state and timeline remain inspectable in the runner. Tests check future-value isolation, prefix invariance, immutable old issues, missing-outcome refusal, unequal-time exact calculations, retained flat-frame residue, signed error without cancellation, empty-score denominators and a deliberately corrupted comparison record. Existing kernel tests remain the authority for the underlying forecast contract. Possible outcomes include losing to hold-last, extending the wrong trend, remaining unassessed during missing observations and having no usable history. None calls for changing the kernel or the policy to obtain a positive result. This experiment does not test multi-source consensus, learned model selection, agency, or superiority to conventional mathematics.