Post Ranking Association

The committed-input recount includes 84/84 posts and 57 author groups. Route: OpenCode Zen free chat via CLI - NOT JEV. Real Jev is queued, not run.

Source: analysis.json. The added label control is not preregistered. These are conditional estimates on a small, selected corpus; see the report for feature errors and limits.

Run these commands from the repository root. Point RUN_DIR outside the repository and use an existing OpenCode CLI installation. The route requires no API key. Protocol registration must already be committed before acquisition.

RUN_DIR=/path/to/external/viral-predictor
OPENCODE_BIN=/path/to/existing/opencode

nice -n 10 python3 experiments/viral-predictor/cli.py collect \
  --opencode "$OPENCODE_BIN" --scratch "$RUN_DIR/oc" \
  --ledger "$RUN_DIR/calls.jsonl" --output "$RUN_DIR/features.json"

Calls run sequentially, first through space-bunny-free, then mimo-v2.6-flash-free only after a transport or schema failure. Each call has a 180-second deadline. The shared-host memory guard waits for usage below 75% before launch and aborts active calls above 85%; its wait is recorded separately from latency. Refusals, invalid cost, and nonzero reported cost stop the campaign.

The append-only ledger retains the 150-call ceiling across restarts, including interrupted attempts. Reuse the same ledger to resume; completed cells are replayed. Six predetermined public posts exercise model-authored explanations. The completed corpus run is retained; no new acquisition is needed for a recount. The current corpus pin differs only in unused source-credit metadata, verified by corpus-binding.json. Original receipt and registration hashes remain unchanged; future acquisition under the re-pinned protocol needs a new ledger. Raw recordings and runtime state remain under the external paths. Output files must be new paths; existing output is never overwritten.

Supply an external JSON list containing objects with exactly id and text. The output includes every input ID: successful rows contain score, category, up to three supported reasons, explanation provenance, and nearest reference IDs; feature failures have a null score and an explicit failure status.

nice -n 10 python3 experiments/viral-predictor/cli.py score \
  --posts "$RUN_DIR/drafts.json" --output "$RUN_DIR/ranked.json" \
  --features experiments/viral-predictor/results/features.json \
  --opencode "$OPENCODE_BIN" --scratch "$RUN_DIR/oc" \
  --ledger "$RUN_DIR/calls.jsonl"

Input texts, private IDs, raw calls, and ranked output stay outside the repository. Text-only drafts supply no attachment metadata, so has_visual is false with that uncertainty recorded. The model's explanation receives only validated checks and category. Its vocabulary and reason codes are bounded; rejected explanations fall back to labelled deterministic wording.

For each query, the five nearest training posts minimize Boolean Hamming distance plus two for a different category. Source order breaks distance ties; engagement counts never choose the neighbors. Published post and neighbor references are SHA-256 hashes of text. The original tie keys are restored only in memory, preserving the registered algorithm and seeded control ordering.

Let y = log1p(likes). The raw prediction combines the neighbor mean with a category mean shrunk toward the overall training mean:

category_estimate = (sum(category training y) + 5 * mean(training y))
                    / (category training count + 5)
raw_prediction = 0.6 * mean(neighbor y) + 0.4 * category_estimate
score = 100 * (training targets below raw_prediction
               + 0.5 * training targets equal to raw_prediction) / training count

All fitting and percentile calibration use training rows only. Floating-point ties use relative and absolute tolerances of 1e-12. With fewer than five training posts, all available references are used. An unseen category falls back to the training global mean. A score expresses position against that training corpus; it is neither a success probability nor a forecast of likes.

The public sanitized feature file contains opaque hashes, checks, category, outcomes, author-group hashes, and preassigned folds needed for evaluation. This command recomputes correlations and intervals without a model call or the external raw ledger:

nice -n 10 python3 experiments/viral-predictor/publication.py recount

To verify original raw receipts and reproduce acquisition accounting, use the CLI with the matching external ledger. Sanitized receipt fields are retained in the committed feature artifact:

nice -n 10 python3 experiments/viral-predictor/cli.py evaluate \
  --features "$RUN_DIR/features.json" --ledger "$RUN_DIR/calls.jsonl" \
  --output "$RUN_DIR/evaluation.json"

Every author's posts occupy the same fold, assigned as int(sha256(author UTF-8), 16) % 5. Predictions for each fold use only the other folds. Spearman correlation uses average ranks for ties. The followers baseline uses the same eligible rows; the shuffled control permutes held-out scores once with seed 20260925 and is descriptive, not a p-value. A separately labelled, unregistered shuffled-label control permutes training likes within each fold using seed 20260925 + fold, refits once, and scores the unchanged held-out likes. It is a diagnostic, not a significance test.

The 95% interval uses 2,000 paired author-cluster bootstrap resamples with seed 20260925, preserving score/target pairs and all posts from each selected author. Models are not refitted. The interval is conditional on these fitted models and corpus; it does not capture repeated feature-call variability or selection bias. Constant-rank resamples have no defined correlation and are excluded with their counts reported. Missing or invalid feature rows are reported in coverage, never converted into negative check answers. A fold without available training rows is explicitly unavailable.

This is a small, selected corpus with dependent posts and limited authors. The earlier sample's saved correlation used different folds and features, so it is context rather than a controlled comparison.

The same checks are queued in pending_jev.jsonl as opaque text hashes with current source and contract pins. No live Jev call is part of this measurement. The updated queue uses durable ledger schema 2; older ledgers fail closed and must not be reused or discarded to bypass budgets. To generate a fresh queue:

nice -n 10 python3 experiments/viral-predictor/cli.py queue \
  --output "$RUN_DIR/pending_jev.jsonl"

After quota is available, the following command can drain the checked-in queue using an already configured free-client credential file. Keep the results external and retain the same results file across invocations:

JEV_PROVIDER=zen JEV_MAX_USD=0 JEV_MAX_CALLS=84 \
nice -n 10 python3 experiments/viral-predictor/jev_queue.py drain \
  --queue experiments/viral-predictor/pending_jev.jsonl \
  --source samples/post-virality/fixtures/posts.jsonl \
  --results "$RUN_DIR/viral-jev-drain.jsonl" --max-calls 84 --batch-calls 4

The drain uses the existing strict free-client guards, memory admission, and a 30-second HTTP timeout. Its durable call ceiling survives restarts; the batch limit caps one invocation. A refusal, error, or unresolved reservation stops automatic retries. Results from these future calls must be labelled separately from the free-chat measurement.

nice -n 10 python3 -m unittest discover -s tests -p 'test_viral*.py'
nice -n 10 scripts/check.sh

The synthetic scoring tests exercise target independence, author isolation, neighbor ordering, known tied-rank statistics, coverage, and cluster-bootstrap behavior. Model and queue tests use local fixtures and mocked clients.