DD: PROMY Pipeline Specification (Draft)#
- Created:
2026-03-29
- Updated:
2026-03-30 (SEED/FEED/GROW/REAP lifecycle adopted)
- Origin:
Informal PROMY discussion during site reorganization session, refined after first PROMY:SEED run on e7He
- Status:
Draft — SEED/FEED/GROW tested on e7He; REAP prompt drafted
The Problem#
FORGE produces raw model content inside append-only llogs: axioms, theorems, symbols, predicates, proofs, test results, and repairs. This content is correct but unstructured — it lives in a 10,000-line session transcript, interleaved with prompts, failed attempts, corrections, and meta-discussion.
SISYF needs clean, structured model source files to assemble audience-depth views.
The gap between FORGE output and SISYF input is PROMY’s job. Currently this gap is bridged informally (or not at all), which is why FORGE-generated models (e7He, e7Ch, e7Tr) lack the structure that older models (PET, JUB) have.
The SEED/FEED/GROW/REAP Pipeline#
PROMY’s four stages follow the natural growth cycle used throughout the POST system (SEED → FEED → GROW → REAP). They also map onto the ZION framework (Zoning → Investigating → Organizing → Navigating):
Stage |
Growth |
ZION |
What PROMY does |
|---|---|---|---|
SEED |
Plant |
Zoning |
Identify and extract axioms, theorems, properties verbatim from FORGE llogs into standalone files. Zone the territory: what exists, how many statements, which labels. |
FEED |
Nourish |
Investigating |
Fill gaps in the standard file set. Investigate which symbols were used, which submodels exist, which logics apply, which aims are open. Feed the record so the next stage has everything it needs. |
GROW |
Integrate |
Organizing |
Apply HELL evidence (corrections, repairs, new findings) to update axiom and theorem statements. Organize beyond the model’s horizon: link to diverse views, cross-reference logs, integrate across the system. This is the read-write dangerous step. |
REAP |
Harvest |
Navigating |
Audit the full session. Close the books. Produce tailored overviews for three audiences (beginner, producer, expert) so each gets a landing page calibrated to their level. The model is now ready for SISYF. |
Each stage is a named lifecycle command (PROMY:SEED,
PROMY:FEED, PROMY:GROW, PROMY:REAP) with an explicit
human review gate before transitioning to the next.
PROMY:SEED — Zone the Territory#
Pull axioms, theorems, symbols, predicates, and definitions out of
FORGE llogs into standalone .rst files with proper labels. Also
create the initial aa.rst (Any Aims) and llog.rst (local
pointer to HELL logs). Write the PROMY llog in HELL.
This is a read-only operation on the llog (extracting, not modifying). Low risk.
Input: FORGE llog (e.g., sa3_2026m03d28/llog.rst)
Output: axioms.rst, theorems.rst, models.rst,
logics.rst, aa.rst, llog.rst + PROMY llog in HELL
Transition: Human reviews extracted content for completeness.
When satisfied: PROMY:FEED
PROMY:FEED — Fill the Gaps#
Ensure the model directory has the complete standard file set (see
below). Investigate which symbols, predicates, and cross-references
exist. Fill in any files that SEED left incomplete. Check that all
:ref: labels resolve. Run make html to confirm zero new
warnings.
Input: Files from SEED + existing model files (symbols.rst, predicates.rst, etc.)
Output: Complete model directory matching the standard layout, with all gaps filled and all cross-references checked
Transition: Human reviews directory structure and file
completeness. When satisfied: PROMY:GROW
PROMY:GROW — Integrate Evidence#
Apply HELL evidence (corrections, repairs from TEMPER rounds, HELL/con and HELL/pro findings) to update axiom and theorem statements. Link to diverse views across the system. Organize cross-model connections. This is the read-write dangerous step. Requires git commit before and after. No unsupervised modifications.
Input: Complete model files from FEED + HELL evidence
Output: Updated model source files with integrated corrections
Transition: Human reviews all diffs. When satisfied:
PROMY:REAP
Standard Model File Set#
After a full PROMY pass (all four stages), every model directory should contain:
model/<name>/
index.rst Main page + toctree
axioms.rst All axioms with :ref: labels
theorems.rst All theorems with :ref: labels
symbols.rst Symbol table
predicates.rst Predicate specifications
models.rst Submodel architecture
logics.rst Formal disciplines used
overview-beginner.rst Plain-language overview
overview-producer.rst Builder-oriented overview
overview-expert.rst Formal/mathematical overview
aa.rst Any Aims tracking (POST convention)
llog.rst Local pointer to HELL llogs
Every model has submodels (even if only “m0 prerequisites + m1–m7
stages”). Every model uses specific logics (even if inherited from
PET/JUB — spell them out). “Not applicable” is never acceptable
for models.rst or logics.rst — the information exists, it
only needs to be documented.
Which files each stage produces:
File |
SEED |
FEED |
GROW |
REAP |
|---|---|---|---|---|
axioms.rst |
create |
check |
update |
|
theorems.rst |
create |
check |
update |
|
models.rst |
create |
check |
||
logics.rst |
create |
check |
||
symbols.rst |
fill gaps |
|||
predicates.rst |
fill gaps |
|||
aa.rst |
create |
update |
update |
close |
llog.rst |
create |
update |
update |
finalize |
overview-beginner.rst |
create |
|||
overview-producer.rst |
create |
|||
overview-expert.rst |
create |
|||
index.rst |
update toctree |
check |
finalize |
LLog Storage#
PROMY llogs record the Prometheus cycle: what toxins (errors, flaws, inconsistencies) were found, what was purified, and what regrew. These records are critical because:
The purification is dangerous. PROMY modifies source data. If a modification introduces a new error, the llog is the only way to trace what happened and why.
The cycle repeats. New FORGE sessions produce new raw content. New HELL evidence exposes new flaws. Each PROMY pass must know what previous passes did to avoid re-introducing fixed errors or overwriting intentional changes.
The liver regrows. Models grow. What was clean after one PROMY pass may need re-purification after the next FORGE session. The llog tracks which version of the model each purification targeted.
Storage location: All PROMY llogs live in HELL, under a dedicated directory:
hell/llog/promy/
promy_e7he_2026m03d29.rst
promy_e7ch_2026m04d01.rst
...
The naming convention is promy_<model>_<date>.rst. The promy_
prefix groups all PROMY logs together when listing the directory. The
model name identifies which model was purified. The date identifies
when.
Why HELL? HELL is “Historically Experienced Lessons Learned” — an adversarial audit record. PROMY logs are exactly that: a record of what was wrong, what was done about it, and whether the fix held. They belong with the other adversarial records, not inside a compiler’s implementation directory.
Future migration: When the broader llog migration happens (moving
FORGE and VV llogs into hell/llog/ as well), PROMY llogs will
already be in the right place. The directory structure will become:
hell/llog/
forge/
forge_sa2_2026m03d27/
forge_sa3_2026m03d28/
promy/
promy_e7he_2026m03d29.rst
vv/
vv_jub_oov2_2026m03d21/
...
Safety Model#
SEED and FEED are low-risk (read-only with respect to existing content; they create new files from llog content or fill gaps). Can run with lightweight oversight.
GROW is high-risk (modifies source data). Requires:
git commitbefore starting (clean baseline)Human review of every diff before accepting
git commitafter completion (auditable endpoint)No batch mode — each modification is reviewed individually
REAP is low-risk (creates overview pages and audit entries). The content is new writing, not modification of existing formal statements.
PROMY must never be invoked by a blanket regeneration command. Each step is a separate operation with a separate human gate.
Open Questions#
Should the
llog.rstfile in each model directory be a literal RST include of the HELL llog, or a hand-maintained pointer page with:doc:links? (Current practice: pointer page.)How should PROMY handle cross-model dependencies (e.g., e7He axioms that reference e7Day axioms)? PROMY is defined as intra-model, but extraction may surface cross-model issues.
Should there be an explicit
make htmlcheck between FEED and GROW to check all labels resolve before attempting modifications?How do the three REAP overview pages relate to SISYF’s audience- depth views? Are they the same thing (in which case SISYF produces them), or are they model-local summaries that SISYF then integrates across models?