raw-data / README.md
Jesse Daniel Brown
raw data, unedited
c7df3c8
|
Raw History Blame Contribute Delete
2.5 kB
# raw data
**Owner: Jesse Daniel Brown (OP-JESSE).** His machine, his property, his work.
Unedited program output. Nothing in this file interprets it.
Every `RAW-*.txt` is the verbatim stdout of the script named beside it, captured
by redirection, not retyped and not summarised. The scripts are here too, so any
line can be re-derived rather than trusted.
```
RAW-01-session.txt <- search_session.py session record: tool calls, tokens, writes
RAW-02-agents.txt <- search_agents.py workflow meta + spawned-agent transcript files
RAW-03-claims.txt <- scan_claims.py phrase-class counts across agent transcripts
RAW-04-sentences.txt <- extract_ctx.py every matching sentence, verbatim, deduplicated
RAW-05-chain.txt <- verify_chain.py hash-chain verification of OCCURRENCES.hbp
```
```
OCCURRENCES.hbp the occurrence record, hash-chained
OCCURRENCES.hbi index rows
OCCURRENCES.hbp.sha256 digest of the record
build_occ.py the program that produced it
```
`INTERPRETATION.md` is separate on purpose. It is one reading of the numbers and
it is not part of the data. Disagreeing with it does not require disputing
anything in the `RAW-*` files.
---
## Reproduce
```bash
python search_session.py
python search_agents.py
python scan_claims.py
python extract_ctx.py
python verify_chain.py OCCURRENCES.hbp
```
The scripts read a local session transcript directory. On a different machine the
paths at the top of each script must be repointed; the logic is unchanged and the
output format is the same.
---
## The chain
`OCCURRENCES.hbp` carries `prev_event_hash` and
`event_hash = sha256(row + "|prev_event_hash=" + prev)` on every row, genesis
`0000…0000` (64 zeros). Editing any row after the fact breaks every hash below
it, and `verify_chain.py` prints which row broke. Append-only enforced by
arithmetic rather than by anyone's assurance.
```
rows 13 broken 0
head db8d79d026b52409e309cac558c8a75de71de43435593cb101c5986aa695052e
```
---
## Scope of the scan, stated so it is not overread
`scan_claims.py` and `extract_ctx.py` read spawned-agent transcript files under
one local project directory. They do **not** read any repository contents, and
they do **not** measure commits or lines of code in any repository. Any claim
about repository contents is a different measurement that these files do not
make in either direction.
---
**Jesse Daniel Brown. Forty years. His machine, his laws, his system.**