Path Claim Plan 46 EN

Experiment 46 plan: a tiny auxiliary-search claim model

Corpus fitting has not started. Freeze the first FP32 plan before inspecting the new cost rows. The model will suggest when auxiliary file search may be worth trying. Ten candidate models are planned; that is not the number of evaluated tasks.

What do 24 and 2,400 count?

The initial difficulty evaluation has only 24 actual tasks and remains a small pilot, insufficient to establish generalization or production utility. Distinguish the later search experiment's 24 model configurations from its 2,948 actual questions. Repeated seeds, representations or choice orders do not create independent tasks.

Quantity	Next-experiment requirement	Purpose
Training rows	At least 2,400 qualified tasks	Fit coefficients
Development validation rows	At least 2,400 qualified tasks	Compare epochs, call penalties and thresholds
Final evaluation rows	2,402 protected tasks; acquisition/source review separate	Confirm after freezing model and execution conditions
Model configurations	Five penalties × two seeds = ten	Candidates evaluated on the same data

2,400 is a minimum for starting comparison, not sufficient evidence to declare success. Preserve existing connected duplicate groups and repository boundaries. Text/ID deduplication does not prove semantic independence or absence of pretraining contamination. Report benefit/harm/tie counts, per-repository results and rare failures separately; retain insufficient-evidence conclusions for small subgroups.

Hypothetical sample-size illustration: with zero failures in independent, same-distribution binomial trials, the one-sided 95% upper failure bound is 1−0.05^(1/n): about 11.73% for 24 trials and 0.125% for 2,400. These are not observed results or guarantees for our data. Repeated observation, correlated tasks and distribution shift invalidate direct use of this simple illustration. See NIST exact binomial confidence limits.

Inputs are the fixed 16 numeric features from baseline search. Raw queries, repository names, file paths and gold ranks never enter model inputs. Output is a ranking score about auxiliary-search value. The sigmoid of a cost-weighted loss is not an actual success probability. Execution policy and fallback to baseline belong to a separate path.

Readiness

Require at least 2,400 source-qualified successful cost rows in both training and validation, with at least 100 in each of five validation repositories. Keep all 7,335 training and 5,686 validation tasks in coverage. Missing data, invalid targets and source conditions do not become negative labels. Protect the 2,402 final tasks until model and threshold are sealed and final evaluation is separately prepared.

The experiment 45 readiness join yields 3,048 training and 107 validation rows after joining conditional source candidates with successful costs. Training count alone does not permit fitting. These are operational candidates for a narrow local numeric scope, not proof of all historical issue rights or public-release permission.

Objective and candidate count

Five call penalties {0, 0.25, 1, 4, 16} and two seeds {1729, 2718} produce ten FP32 candidates. Keep features, learning rate, regularization, 100 epochs and batch64 fixed. Minimum validation loss chooses the epoch, with earlier epoch on ties. Do not select only a favorable seed.

Target is baseline pages − helper pages − penalty > 0, weighted by the absolute difference. Penalties are page-equivalent experimental values, not money or token prices. Divide every weight by the fixed constant 5,016 derived from 100,000-candidate/page20 bounds. Constant scaling preserves the trainer's normalized weighted loss and update interpretation. Do not clip large examples after inspecting their outcomes.

Validation threshold selection includes disabled baseline and minimizes page cost under a 90% helper cap. Keep equal-score groups intact. Apply fixed call-count, penalty, threshold and epoch tie-breaks. Compare baseline-only, always-helper and a baseline-gap heuristic with the same cap too.

Required evidence

Both seeds must reduce validation pages by at least 5% versus baseline and 1% versus always-helper. Pooled and equally weighted repository Hit10 must not decline, and no validation repository may worsen total pages by more than 5%. Record failure rather than adaptively searching this same data family. These are pilot advancement gates.

There are 16 FP32 coefficients, with Go-owned contiguous training columns. Disclose float64 feature or accumulation arithmetic separately from coefficient storage. Separate head bytes from feature/search pipeline costs. The complete runner's isolated RSS target is at most 256MiB; report exceedance without hiding work or removing rows.

Implementation preparation and remaining work

The existing Go learner supports dimension16, while old three-repository wrappers do not fit the new cohort. internal/pathclaim prepares the numeric dataset, fixed conditions and model format. Synthetic test fitting does not mean real corpus fitting is complete. The future runner must execute all ten trials, capture every epoch, seal actual input/source/runner hashes, measure resources and review publication notices.

After useful FP32 signal, separately compare FP32/INT8/ternary PTQ/ternary training with identical inputs and budgets. Packing 16 coefficients alone need not accelerate the whole pipeline. File search still needs actual agent completion and total usage evidence; model up/down routing, dependency-aware decomposition and repository selection require their own golden sets. First-target pages cannot replace the full project objective.