# Reproducing the local FAROS equity report This benchmark compares model plus native coding harness inside CORAL. It measures adaptive research on the reconstructed Yahoo-backed S&P 500 task. It does not establish held-out trading alpha. Binance results are archived separately. ## Registered comparisons The authoritative core mapping is `REPETITION_PRIORITY_AMENDMENT_03.json`: origination physical runs 1/2/3/4/5 and refinement physical runs 2/3/7/5/6, in logical repetition order. Each allocation gives each of Astra, Terra, Luna and Sonnet sixteen charged candidate attempts. Refinement physical run 1 is diagnostic; run 4 stopped during controls before research and is preserved separately. Failed candidate attempts consume budget and retain the incumbent. Transfer compares no notes, relevant source notes and unrelated notes on the target task, twice per system. Source research is a separate allocation. The assisted final Luna endpoint in no-notes repetition 2 is excluded from the unassisted comparison; earlier eligible checkpoints remain. Read its budget-correction receipt for the exact intervention. Only the completed Astra island-sharing comparison is retained, explicitly exploratory. Other collaboration and framework comparisons are deferred by user instruction. The original four-axis Overall formula is preserved, but cannot produce a primary score without primary collaboration evidence. Missing scores must remain missing. ## Regenerate the report Run from this checkout using the existing local environment: ```sh .deps/faros-equity-runtime/venv/bin/python reports/render_faros_equity_report.py ``` This reads saved native evidence; it does not invoke research models or graders. It writes `reports/faros-web/index.html` and `equity-pilot-results.json`, and updates the existing report under the parent trading workspace. The renderer verifies hashes of saved usage evidence before including it. While research is live, output is a snapshot, not a final result. `faros_native_evidence.py` checks admission/result identity, finite scores, reconciled accounting and an unopened seal. `summarize_faros_matrix.py` constructs matched comparisons; `summarize_faros_equity_pilot.py` supplies the detailed refinement view. `faros_observed_usage.py` deduplicates recorded provider responses. `faros_strategy_history.py` retrieves frozen code from native Git commits. Means and pointwise 95% Student t intervals use independent matched repetitions, not strategy variants. The shaded learning bands are uncertainty in the mean across those runs, not uncertainty about future market returns. Early and late checkpoints may have different sample sizes during execution. ## Evidence to preserve - `configs/faros-equity-v1/`: registered tasks, budgets and runtime paths. Later runtime amendments and launch receipts explain configuration versions; do not assume an earlier amendment hash is the final launch hash. - `results/faros-equity-v1/`: native append-only ledgers, scored reports, Git repositories, runtime state and provider evidence. These are necessary for regeneration; the HTML alone is insufficient. - `SURFACE_BUILD_01.json` and `FINAL_SURFACE_BYTES_01.json`: frozen core surface manifest and current byte verification. The latter verifies bytes, not the economic correctness of features. - `MARKET_DATA_AUDIT_01.json`, `MEMBERSHIP_AUDIT_01.json`, and the source audit receipts: data acquisition and limitations. Histories are a reconstructed vendor vintage, not the friend's original raw panel. - `TRANSFER_SURFACES_01.json` and `TRANSFER_SURFACE_VALIDATION_01.json`: source/target construction and validation. - Per-run `CONTROLS`, `LAUNCH`, `HARNESS_AUDIT`, `COMPLETION` and `USAGE` receipts: control results, actual configuration, observed model identity, completion and resource evidence. Not every early run has every receipt naming variant. - `FINAL_FROZEN_ARTIFACT_INVENTORY_01.json`: complete inventory of ended submissions across the selected allocations, including Git blob identities and scored-report references. No selected allocation has a pending admission. - Registered grader and CORAL wheels under `.deps/`, plus runtime installation receipts: preserve exact bytes. Do not replace them with current package versions when reproducing an old result. Native histories can contain subscription credentials or sensitive transcripts elsewhere in their directory trees. This is a local preservation map, not a publishable archive. The report download contains selected evidence rather than a full environment export. ## Verification and execution boundaries Lightweight reporting tests can run without grading: ```sh uv run --no-sync pytest -q tests/test_faros_observed_usage.py tests/test_faros_matrix.py tests/test_faros_equity_reporting.py ``` The required repository-wide command is `uv run pytest -q`, followed by `git diff --check` before a local commit. Run the full suite only after the native manager is terminal: it can invoke graders, and this machine allows one grader globally. `TEST_STATUS_09.json` records the final full suite: 108 failed, 3179 passed, 20 skipped and 31 setup errors. The one new failure was an outdated plotting fixture; after correction, all 13 focused FAROS reporting tests passed. The other failure identities match the previous baseline. The repository suite is not green. Report regeneration and artifact verification are distinct from evaluator replay. No independent replay is claimed by this guide. Replaying requires the frozen strategy, exact grader/runtime, surface and economics in a separate audit output location; it must not append or reset existing experimental ledgers. Do not reopen the seal or launch new research to reproduce a report. For another machine, preserve artifacts and resolve environment paths explicitly. Registered Mac absolute paths are provenance, not portable commands. Model research itself is stochastic: reproducing saved arithmetic is different from obtaining the same strategy in a new subscription session. ## Three-capability reporting amendment The user requested a three-capability Overall after observing the pilot results. The current display excludes collaboration and equally weights the mean origination, refinement and transfer effects, each normalized by the unchanged public reference scale. This is a reporting amendment, not the original four-axis registration. Original `composite` fields remain archived; `report_overall` holds the displayed calculation. Overall uses all eligible observations within each capability, with equal capability weights rather than pooled run weights. Its conservative 95% interval sums the bounds of three Bonferroni-adjusted Student t intervals (98.333% each), then applies the same weighting and scaling. No independence between capabilities is assumed; within-capability Student t assumptions still apply. With fewer than two eligible observations in any capability, the point estimate is provisional and the interval is unavailable. Collaboration comparisons are planned for a later release. ## Fixed per-capability reference units (current scoring) `FIXED_AXIS_SCALES_01.json` supersedes the shared-scale three-capability display above. The original pilot registration and older scores remain historical evidence. The new rule was selected after observing the pilot, but its denominators use only the separate 2021–2022 public controls, not any model outcomes. Origination uses the best fixed full-panel public control over cash. Refinement uses that same control over the reversal seed. Transfer selects the best fixed control on public source issuers, then uses its public target-issuer return over cash. This last anchor is a simple information-guided control reference, not an empirical reference LLM notes effect. These units address the different starting opportunities, without claiming equal task difficulty. `AXIS_CALIBRATION_REGISTRATION_01.json` was written before the new source and target public evaluations; their individual results and daily paths are preserved under `axis-calibration-source` and `axis-calibration-target`. The original surface and candidate ledgers were not edited. Use `reports/freeze_faros_axis_scales.py` to reproduce the frozen denominators from the recorded control summaries. It fails rather than changing an existing frozen result. Nonpositive references yield an undefined scale, not an alternative selected after seeing the outcome. Overall averages the three independently normalized mean effects. Its conservative interval weights each axis interval by its own denominator. Future models use the same frozen values; changes to data, tasks or anchors require a new version. Normalized scores are shown below raw capability effects. These are benchmark units, not return percentages. ## Current report scope: two capabilities The user deferred transfer from the visible report. `fixed-reference-core-v1` averages only normalized origination and refinement, 50% each. Their frozen denominators are unchanged. For Overall, two 97.5% marginal Student t intervals produce a conservative 95% interval using Bonferroni; both capabilities have five eligible observations per model. Transfer does not enter scores, plots, usage tables, or examples in this release. Its complete data, configurations, calibration outputs, and native evidence remain archived for later work. No additional transfer runs were launched.