Claims, with re-run methods

Every number daimon publishes about itself ships with the method to reproduce it. The numbers below describe our own store — two months of dogfooding on one machine. Running the same commands on your install measures your store; that is the point. A claim you cannot re-run is a testimonial, and this page does not carry testimonials.

One definition rides with everything here: verified means daimon found this exact quote in your transcript — matched after folding case, whitespace, and formatting, with elided spans allowed. It is a mechanical check, not a truth claim.

The claims

Claim	Published number / guarantee	Re-run with
Fresh verbatim claims fail the quote check at a measurable, published rate	12.2% lifetime (450 of 3703 claims, 160 sessions, snapshot 2026-08-07) — dated row + caveats	snippet below
Every stored verbatim quote stays re-checkable after the fact	exit 0 proven / 1 mismatch found / 3 cannot prove	daimon audit quotes
A forgotten value is provably gone from every declared surface	same exit contract, hash-only reporting	daimon audit privacy
Deletion survives re-serializing the original transcript	tombstone suppresses the item on re-capture	daimon forget <id>, then daimon serialize <transcript>, then daimon recall for the value
A checkpoint's provenance is offline-checkable (opt-in)	ed25519 signature binding exact bytes to the source transcript	daimon verify-receipt

Re-running the downgrade rate

The rate is derived from your own checkpoint store — no network, no tooling beyond Python. Rotated pointer copies duplicate sessions on disk, so deduplication by session id is mandatory (a naive glob double-counts):

python3 - <<'EOF'
import json, glob, os
def items(o):
    if isinstance(o, dict):
        if "quote_verified" in o: yield o
        for v in o.values(): yield from items(v)
    elif isinstance(o, list):
        for v in o: yield from items(v)
seen=set(); checked=downgraded=0
for p in glob.glob(os.path.expanduser("~/.daimon/checkpoints/**/*.json"), recursive=True):
    try: cp=json.load(open(p, encoding="utf-8"))
    except Exception: continue
    if not isinstance(cp, dict): continue
    sid=cp.get("session_id")
    if not sid or sid in seen: continue
    seen.add(sid)
    for it in items(cp):
        checked+=1
        if it.get("quote_verified") is False: downgraded+=1
print(f"{downgraded}/{checked} downgraded across {len(seen)} sessions"
      + (f" = {downgraded/checked:.1%}" if checked else ""))
EOF

What it counts: every item that carries a quote_verified stamp — false means the quote failed re-verification at capture time and the item was demoted to inferred. The published 12.2% is a dated snapshot; the same command on the same store four sessions later already reads 11.9% (460/3869, 164 sessions). Your store will read your number, from day one.

What the audits prove

daimon audit quotes re-checks every stored verbatim quote against its source transcript, read-only. daimon audit privacy hashes every plaintext field on every declared surface and intersects with the deletion ledger. Both share one exit contract: 0 is proven clean, 1 names the residue by surface and hash, and 3 means a surface could not be read — never treat 3 as clean. That last code exists because "could not check" reported as "all clean" is how audit theater works.

What is deliberately not on this page

Anything measured once, on a sample too small to state, or not yet reproducible by a command you can run. When a number graduates to a method, it moves here with its date.