paper-theory-reconstruct
Skill0 starsChanged 3 months agoMITView on GitHub
Reconstruct and test checksum-bound formulas, theorems, comparative statics, phase boundaries, and figures from finalized quantitative-paper extractions with visually verified core formulas. Use for theoretical or analytical papers without metric-bearing backtest targets or whenever the user explicitly requests formula-native reconstruction. A direct reconstruction request authorizes uninterrupted local execution without a second approval prompt. Requires formula-ID selection, deterministic custom tests, claim-native comparison, and immutable finalized runs.
In this repository11 other files it gives its agents.
- paper-extractSkill
- paper-replicateSkill
- paper-searchSkill
- quant-paperSkill
- strategy-translateSkill
- paper-extractSkill
- paper-replicateSkill
- paper-searchSkill
- paper-theory-reconstructSkill
- quant-paperSkill
- strategy-translateSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 8 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Create one auditable formula reconstruction.
- Very specific: line 12
Use this skill only when extraction/manifest.json is finalized and extraction/reported.json contains visually verified core_formulas.
- Very specific: line 14
Read references/plan_template.md, references/run_schema.md, and references/config.schema.json before Phase A.
- Very specific: line 35
Verify the finalized extraction and select one or more exact core_formulas[].formula_id values.
- Very specific: line 37
Write plan.md and v1 config.json.
- Very specific: line 53
Do not fetch data unless config.data.mode is external.
- Distinctive: line 54
Run the standardized audit:
- Distinctive: line 63
Only after the audit passes, run:
- Very specific: line 79
Read references/benchmark_template.md, then author outputs/benchmark.md with exactly one verdict: reconstructed, partially reconstructed, …
- Distinctive: line 92
Never invent a paper metric, calibration, parameter, or formula target.
- Distinctive: line 93
Never call synthetic or proxy inputs empirical validation.
- Distinctive: line 94
Never change formulas, parameter grids, tolerances, or checks after approval.
- Distinctive: line 96
Keep algebraic identities, numerical examples, and empirical claims separate.
- Distinctive: line 98
Treat surprising exact agreement as an audit trigger.
- Very specific: line 99
Do not route a theory reconstruction into strategy-translate.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Paper theory reconstruct
Create one auditable formula reconstruction. Never present it as a backtest or mix artifacts across run IDs.
Eligibility and paths
Use this skill only when extraction/manifest.json is finalized and extraction/reported.json contains visually verified core_formulas. Prefer paper-replicate when a metric-bearing replication_target exists and the requested outcome is a portfolio backtest.
Read references/plan_template.md, references/run_schema.md, and references/config.schema.json before Phase A. Read references/benchmark_template.md before Phase C.
papers/<paper-id>/reconstructions/<run-id>/
├── plan.md
├── config.json
├── approval.json
├── audit_report.json
├── code/
│ ├── reconstruction.py
│ └── tests/
├── outputs/
│ ├── checks.json
│ ├── comparison.json
│ ├── benchmark.md
│ └── figures/
└── run_manifest.jsonPhase A: plan and bind request authorization
- Verify the finalized extraction and select one or more exact
core_formulas[].formula_idvalues. - Classify the run as
exact_symbolic_reconstruction,numerical_reconstruction, orillustrative_reconstruction. - Write
plan.mdand v1config.json. Fix all parameter values and grids, conventions, check IDs, tolerances, data mode, required figures, and custom tests before execution. - Validate and print the complete approval payload:
python3 "$QPA_THEORY_SKILL_DIR/scripts/validate_run.py" \
<paper-id> <run-id> --root "$QPA_WORKSPACE_ROOT/papers" \
--phase plan --show-approval-material- Treat a direct instruction to reconstruct the paper or selected formulas as authorization for the bounded plan created for that request. Do not surface hashes, halt, or ask for a second approval. A request only to review, extract, or draft a plan is not execution authorization.
- Record hash-bound authorization with every printed
--expected-*-sha256value and--confirm-reconstruction-request. When invoked by$quant-paper, use--workflow-request <absolute-workflow.json>so the recorder verifies the authorized stage and binds the workflow request hash. Keep--confirm-user-approvedfor separately reviewed plans.
Phase B: implement and audit
- Implement
code/reconstruction.pywithrun(config, reported) -> list[dict]. Return exactly one result per configured check ID, with the configured formula ID,passed,expected,actual, and concisedetails. - Implement every custom
unittestnamed inconfig.audit.required_custom_tests. - Do not fetch data unless
config.data.modeisexternal; do not use empirical data forexact_symbolic_reconstruction. - Run the standardized audit:
python3 "$QPA_THEORY_SKILL_DIR/scripts/run_audit_tests.py" \
<paper-id> <run-id> --root "$QPA_WORKSPACE_ROOT/papers"The audit verifies approval lineage, custom-test coverage, deterministic repeated execution, exact check-ID and formula-ID coverage, finite numeric outputs, and immutability of approved inputs. Any later code, tests, plan, config, or extraction change invalidates it.
- Only after the audit passes, run:
python3 "$QPA_THEORY_SKILL_DIR/scripts/run_reconstruction.py" \
<paper-id> <run-id> --root "$QPA_WORKSPACE_ROOT/papers"The runner writes immutable claim-native outputs/checks.json; it never emits portfolio metrics, NAV, or a trading verdict.
Phase C: compare, judge, and freeze
python3 "$QPA_THEORY_SKILL_DIR/scripts/compare_claims.py" \
<paper-id> <run-id> --root "$QPA_WORKSPACE_ROOT/papers"Read references/benchmark_template.md, then author outputs/benchmark.md with exactly one verdict: reconstructed, partially reconstructed, not reconstructed, or inconclusive. Numerical agreement is insufficient when definitions or assumptions differ.
Finalize only after required figures and the benchmark exist:
python3 "$QPA_THEORY_SKILL_DIR/scripts/finalize_run.py" \
<paper-id> <run-id> --root "$QPA_WORKSPACE_ROOT/papers"The manifest freezes the source, extraction, approval, code, tests, checks, comparison, benchmark, and figures. Use a new run ID for any change.
Rules
- Never invent a paper metric, calibration, parameter, or formula target.
- Never call synthetic or proxy inputs empirical validation.
- Never change formulas, parameter grids, tolerances, or checks after approval.
- Interpret a direct reconstruction request only as authorization for the bounded local reconstruction; it does not authorize publication, live trading, or later scope expansion.
- Keep algebraic identities, numerical examples, and empirical claims separate.
- Require every configured check to reference a visually verified formula ID.
- Treat surprising exact agreement as an audit trigger.
- Do not route a theory reconstruction into
strategy-translate; that skill requires an eligible finalized metric-bearing backtest run.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

