Workflow
This repository supports technical and scientific AI discussions through a provenance-preserving evidence maturation pipeline. Workflow state, evidence state, implementation state, and published output are separate concerns.
Repository areas
.cache/ - local evidence workspace
.cache/ is ignored by Git. Use these local subfolders:
.cache/src/ for fetched, copied, or captured source material;
.cache/md/ for extraction output and working Markdown;
.cache/data/ for raw and intermediate datasets;
.cache/results/ for provisional machine results;
.cache/reports/ for provisional generated reports.
Files here may be incomplete, large, licensed for local use only, or generated. They are inputs to review, not durable claims and not publication inputs.
memory/ - episodic investigations
An investigation is a bounded record of work on a question. It should identify:
the question, scope, and date;
sources with stable identifiers or URLs and retrieval dates;
methods, tools, queries, and transformations;
observations separately from interpretation;
findings, confidence, limitations, contradictions, and open questions;
links to promoted knowledge, when promotion occurs.
Memory may preserve negative results and uncertainty. It is evidence of an investigation, not automatically a reusable truth statement.
knowledge_base/ - curated reusable knowledge
Knowledge-base content is reviewed and structured for reuse. Each material claim must remain traceable to committed investigation records or directly to stable sources. Curated content should state its scope, confidence, freshness or review date, and known disagreements. When evidence changes, revise the curated content without erasing the investigation history.
src/ and src/specification/
src/ contains the local CLI, scripts, automation, result generation, and
report generation. Do not store collected evidence or generated results there.
src/specification/ contains durable behavioral contracts for src/. For a
non-trivial automation change:
write or update the relevant specification;
create a dated plan when the work is complex;
implement against the specification;
validate with focused tests or reproducible commands;
record implementation facts in the plan packet.
Promotion gates
Use these gates before material advances:
| .cache/ | memory/ | Source identity, retrieval date, method, scope, and limitations are recorded; unsafe or unlicensed payloads stay local. |
| memory/ | knowledge_base/ | Claims are reviewed, evidence links resolve, contradictions and uncertainty are represented, and the result is structured for reuse. |
| knowledge_base/ | website | Content is approved, contains no local/private material, passes rendering checks, and preserves citations and provenance. |
Nothing is promoted solely because an automated process produced it. When a claim cannot pass a gate, keep it in the earlier layer and record why.
Plans
Complex work uses dated packets:
plans/YYYY.MM/
DD/
<title-slug>/
plan.md
implementation.md # create only after implementation starts
test.md # optional preserved verification plan.md should capture the problem, goals, scope, non-goals, dependencies,
risks, open decisions, phases, and exit criteria. Use stable open-point IDs such
as OP-001 when a decision must survive across sessions.
implementation.md is a factual log, not a second plan. It records what
landed, deviations, files changed, and remaining risks. Begin it with a status
line such as [##--] Phase 2/4 - ...; use [####] Done - ... when the planned
implementation is complete.
Keep these indexes synchronized:
plans/open.md lists packets with incomplete implementation or unresolved
planning decisions;
plans/closed.md lists completed implementations and settled planning-only
packets.
Website and generated reports
Published output must be reproducible from committed, approved inputs. Website
build tooling may read memory/ and knowledge_base/, but default navigation
should distinguish episodic investigations from curated knowledge. It must not
read .cache/ in CI.
Generated reports remain in .cache/reports/ until reviewed. If a report is
accepted as durable content, promote its source-backed content into memory/
or knowledge_base/; do not treat generated formatting as provenance.
Validation and ownership
Use the smallest meaningful validation: link/citation checks for evidence,
schema and focused tests for src/, document consistency checks for workflow
changes, and a production build plus Pages artifact check for website changes.
Record durable proof in the relevant packet's test.md or
implementation.md.
The maintainer owns Git history and external publication. Agents leave changes in the working tree unless explicitly asked to commit, push, deploy, publish, or modify another repository.