Version 0.2.13 (2026-11-04)
This repository is part of Fiskwise, with its project hub. See MIGRATION.md for repository and local config migration notes.
Teach your AI coding agent to run a biomass-dynamic stock assessment — from catch series to catch advice — for any stock, on your own machine.
spict-wise is a knowledge pack that turns a general coding agent (Claude Code, Codex, Cursor, Gemini CLI, Mistral, …) into a competent assistant for SPiCT assessments — the Surplus Production model in Continuous Time of Pedersen & Berg (2017). It is distilled from the reg-spict (golden redfish) and reb-spict (beaked redfish) assessments of the Deep-water species and cartilaginous fish group at the Institute of Marine Research.
Tell your agent which stock you want assessed, and it will:
- Ask you a structured questionnaire and confirm a plan.
- Scaffold a project folder with wise-writing's standard layout (
docs/,R/,data/,figures/,ai/,logs/,config/). - Compile the data — the catch series and the abundance index — and record where every number came from.
- Fit, diagnose, and report the model, then produce the working-group chapter and the advice sheet.
It is not a trained model. Your data never leaves your machine and never enters any model's weights. The agent simply reads this repo while it helps you. See Confidentiality.
spict-wise owns the SPiCT assessment layer. wise-writing owns the shared project layout, scientific writing style, document conventions, and rendering pipeline. The dependency is real: scaffold.R calls wise-writing's scaffold_document() and overlays the SPiCT content, so there is exactly one copy of shared infrastructure. Install wise-writing on demand before scaffolding; installing spict-wise alone does not require it.
For length- and age-structured models of the same stocks, see gadget-wise (gadget3).
Use fiskwise-workflow for requests spanning packs and fiskwise-install, fiskwise-update,
and fiskwise-uninstall for managed lifecycle. Standalone spict-wise skills remain available.
For managed R tasks, always use r-code-style.
install https://github.com/DeepWaterIMR/wise-writing
- Install spict-wise:
The agent runs
install https://github.com/DeepWaterIMR/spict-wisespict-install: clone spict-wise → install its skills globally → check the SPiCT toolchain. - Before scaffolding, ensure wise-writing is installed; the agent will request it if needed.
- Start an assessment:
The agent runs the questionnaire, plans, scaffolds a project, and drives the workflow.
Let's set up the 2026 SPiCT assessment for beaked redfish in ICES subareas 1 and 2, using spict-wise.
Manual install: clone the repo and open your agent in the folder, or point its instructions at
<clone>/skills/. Nothing here needs a network connection at use time.
| Step | Skill | Produces |
|---|---|---|
| Ask about or review a SPiCT fit in any project | spict-model |
Answers and reviews; no scaffold needed |
| Scaffold a project (questionnaire + plan) | spict-new-assessment |
The project folder and config.yaml |
| 1. Compile the catches and the index | spict-compile-data |
The two series the model is fitted to |
| 2. Fit, diagnose, project | spict-fit-model |
The assessment data report and the summary object |
| 3. Explore alternatives (benchmark) | spict-explore |
Candidate fits, comparisons, the Shiny explorer |
| 4. The working-group chapter | spict-assessment-report |
A Word chapter in the group's template |
| 5. The advice sheet | spict-advice-sheet |
The catch advice |
Long renders go through spict-render — code review first, then a screen session with a log in logs/.
The model is fitted once. Step 2 saves a small summary object; the chapter and the advice sheet read it rather than refitting, which is what makes all three documents quote the same numbers.
AGENTS.md/CLAUDE.md— the agent contract.skills/— the ninespict-*skills, installed globally.knowledge/— the shared source of truth: the model conventions, catch data, indices, diagnostics, advice, the hard constraints, project structure, the wise-writing contract, packages, and rendering.project-template/— the SPiCT layer overlaid onto wise-writing's skeleton.scaffold.R—scaffold_spict(), which stamps a project fromconfig.yaml.examples/redfish/— a worked example.config/validate-scaffold.R— the structural check.
SPiCT is a state-space model fitted to two short series, and most of the ways it goes wrong are silent. These are the pack's conventions, documented in knowledge/spict.md:
- Catches in tonnes throughout; the index scaled by its own mean; index timing a fraction of the year matching when the survey runs.
stdevfacvectors must average to 1 — they multiply an estimated observation error, so a vector that does not silently rescales it.lognfixed at the Schaefer value: a production-function shape estimated from one catch series and one index is not identified.logalphaandlogbetapriors deactivated, so they do not fight an informative prior onlogsdb.- Reference points are relative. SPiCT estimates F
MSYand BMSY; Blimand MSY Btriggerare conventions the advice framework supplies. - Priors carried over from another stock are placeholders, not defaults.
Three hard constraints override tidiness: exploratory-status callouts stay, official-track advice numbers do not change under a cosmetic edit, and advice-history rows are extended rather than fitted.
This repo is public and contains instructions only — never data:
- No data and no private paths are ever committed here. Catch series, index files, and model output live only in the project folder the agent generates for you, where
data/,logs/,figures/, andai/review/are git-ignored. - Norwegian sales-note data carry confidentiality constraints, and catch data by country or vessel can be commercially sensitive. Generated projects default to a local Git repository with no remote; publishing one follows a review of whether the series it contains can be published.
- wise-writing — scientific writing style, document conventions, and the shared project skeleton. Required before scaffolding; install it on demand.
- index-wise — sdmTMB spatiotemporal survey indices. Produces the index a SPiCT assessment consumes.
- BAIT — IMR Biotic data access, privacy, maps, life history.
- wise-presentations — Quarto RevealJS slides.
Install whichever you need alongside; the skills stay out of each other's way.
This repo speeds up the data compilation, the fitting, and the writing. It does not decide what is true, and it does not produce advice that is ready to publish. A scaffolded project is a place to start: the priors are placeholders until someone calibrates them, the diagnostics need judging rather than displaying, and every number in an advice sheet is read by managers. Expert judgment is required throughout, and the humans named as authors are accountable for all of it.
Pedersen, M.W. & Berg, C.W. (2017). A stochastic surplus production model in continuous time. Fish and Fisheries 18, 226–243. doi:10.1111/faf.12174.
Mildenberger, T.K., Berg, C.W., Pedersen, M.W., Kokkalis, A. & Nielsen, J.R. (2020). Time-variant productivity in biomass dynamic models on seasonal and long-term scales. ICES Journal of Marine Science 77, 174–187. doi:10.1093/icesjms/fsz154.
See CONTRIBUTING.md. Run Rscript config/validate-scaffold.R before opening a pull request.
GPL-3, matching spict itself. See LICENSE. The Word reference documents and the citation style come from wise-writing at scaffold time and keep their own terms.