Grain — Identity Tolerance Spectrum

Grain is INDB's identity verification module. It answers one question across four tiers of noise tolerance:

Is this the same origin — and how sure are we, given distortion?

Where Echo finds meaning through resonance and signing delivers binary axioms, Grain unifies both into a single axis: from barcode-exact to SynthID-statistical.

Inhale identity. Tolerate noise. Verify the grain.

The Problem

INDB stores memory, not frozen archives. Real memory tolerates:

Partial quotes and rewrites (literary analysis)
Surface change (signature on glass vs paper)
Export damage (7 of 10 signed events survive backup loss)
Long-run statistical fingerprints (token patterns over hundreds of events)

Ed25519 alone is barcode-tier: verified or rejected, ~0% tolerance. That is correct for Axioms — but insufficient for the full epistemological pipeline.

Grain fills the gap between Axiom (exact) and Echo (semantic resonance).

The Noise-Tolerance Axis

From rigid binary data to flexible statistical patterns:

[Exact] ──────> [Redundant] ──────> [Resonant] ──────> [Statistical]
 ~0–5%              ~30%              ~30–50%           floating (p-value)
 Barcode            QR code           Signature         SynthID
 Ed25519            N-of-M cluster    Echo × rep        n-gram fingerprint

Tier	Analogy	Mechanism	Verdict style
exact	EAN-13 barcode	Ed25519 / `key_id`	`verified` or `rejected`
redundant	QR (Reed–Solomon H)	N-of-M signed events survive damage	`verified` / `likely_same_origin`
resonant	Handwritten signature	Echo resonance × source reputation + FAR/FRR	`likely_same_origin`
statistical	SynthID watermark	Token n-gram distribution vs reference cluster	p-value + length gate

Relationship to Other Modules

Module	Layer	Question
Signing	Storage	Is this cryptographically proven?
Grain	Identity	Same origin despite noise?
Echo	Meaning	What resonates with this?
Prism	Interpretation	What does it mean to this observer?
Recovery	Restore	Can we rebuild signed memory by `key_id`?

Grain uses Echo internally for the resonant tier but does not replace it. Grain wraps signing for the exact tier but does not weaken Axiom requirements.

API

List tiers

curl https://api.indb.tech/api/v2/grain/tiers

Verify identity

curl -X POST https://api.indb.tech/api/v2/grain/verify \
  -H "Content-Type: application/json" \
  -d '{
    "reference_event_id": "<seed-event-uuid>",
    "candidate": {
      "raw_data_anchor": ["partial", "rewrite", "summer", "wine"],
      "location": "books/test"
    },
    "tier": "auto",
    "tolerance": 0.5
  }'

Response:

{
  "status": "success",
  "data": {
    "tier_used": "resonant",
    "match_score": 0.62,
    "p_value": null,
    "signal_to_noise": 1.63,
    "verdict": "likely_same_origin",
    "false_accept_risk": 0.08,
    "false_reject_risk": 0.12,
    "insufficient_signal": false,
    "tokens_observed": 4,
    "min_tokens_required": 0,
    "key_id": null,
    "insight": "Resonance 0.71 × reputation 0.50 ≥ threshold 0.50 — likely same origin.",
    "meta": {
      "resonance": 0.71,
      "reputation": 0.5,
      "threshold": 0.5,
      "reference_id": "...",
      "candidate_id": "..."
    }
  }
}

Request fields

Field	Required	Description
`candidate`	yes	Event dict or partial event (`raw_data_anchor`, optional `signature`)
`reference_event_id`	resonant / statistical	Primary seed event
`reference_event_ids`	redundant	Cluster of signed events for N-of-M check
`expected_key_id`	exact / redundant	SHA-256 public identity hash
`tier`	no	`auto` (default), `exact`, `redundant`, `resonant`, `statistical`
`tolerance`	no	0.0 = strict (high false reject) → 1.0 = permissive (high false accept)
`min_tokens`	no	Minimum tokens for statistical tier (default: 20)
`min_events`	no	Minimum events for redundant tier (default: 3)
`damage_ratio`	no	Max tolerated damage for redundant tier (default: 0.30)

Verdicts

Verdict	Meaning
`verified`	Axiom-grade match (exact or full redundant cluster)
`likely_same_origin`	Probabilistic match — resonant or statistical
`insufficient_signal`	Too little data (e.g. 20 tokens for SynthID-like detection)
`rejected`	Origin not confirmed at chosen tier

Tier Details

1. Exact (~0–5%)

Uses core/signing.py — Ed25519 over canonical event fields.

curl -X POST https://api.indb.tech/api/v2/grain/verify \
  -H "Content-Type: application/json" \
  -d '{
    "candidate": { "...signed event..." },
    "tier": "exact",
    "expected_key_id": "<64-char-hex>"
  }'

Binary threshold: signature valid or not. Use for Axioms and recovery proofs.

2. Redundant (~30%)

QR-like: a cluster of signed events tolerates partial export loss.

curl -X POST https://api.indb.tech/api/v2/grain/verify \
  -H "Content-Type: application/json" \
  -d '{
    "reference_event_ids": ["id-1", "id-2", "id-3", "id-4"],
    "tier": "redundant",
    "damage_ratio": 0.30,
    "expected_key_id": "<key_id>"
  }'

Survival ratio = valid signatures / total. Default requires ≥70% valid (30% damage tolerance).

3. Resonant (~30–50%)

Signature biometrics: same hand, different surface.

match_score = resonance × (0.5 + reputation × 0.5)
threshold   = 0.70 − tolerance × 0.40

Low tolerance (0.0): threshold ≈ 0.70 — strict, more false rejects
High tolerance (1.0): threshold ≈ 0.30 — permissive, more false accepts

Use for: partial rewrites, fused events, cross-location echoes, Reality Guard signal vs crowd.

4. Statistical (floating p-value)

SynthID-like: compares candidate token n-grams to a reference cluster distribution.

Length gate: short candidates cannot be verified — same as SynthID in 20 words vs 500 words.

Reference cluster: ≥ min_tokens (default 20) across ≥ 2 events
Candidate:         ≥ max(5, min_tokens / 4) tokens

Optional watermark embedding for source connectors:

from core.grain.statistical import embed_watermark_bias

tokens = embed_watermark_bias(
    tokens=["memory", "truth", "signal"],
    secret=b"source-watermark-secret",
    strength=0.15,
)

Call from services/sources/*/transformer.py at ingest time — not applied automatically.

Auto Tier Selection

When tier: "auto":

Candidate has signature → exact
Reference cluster ≥ min_events with signatures → redundant
Reference + candidate token counts meet statistical gates → statistical
Otherwise → resonant (requires reference_event_id)

Use Cases

Recovery after partial export

Export lost 2 of 5 signed events. Redundant tier: 3/5 = 60% < 70% → rejected at default; lower damage_ratio to 0.40 → likely_same_origin.

Literary partial quote

Original event ingested from Dostoevsky. Candidate is a partial rewrite. Resonant tier with moderate tolerance confirms thematic origin without exact token match.

Reality Guard

Crowd events (low reputation) → resonant tier, high false_accept_risk
Internal heartbeat (high reputation + exact signature) → exact tier, verified

Source connector provenance

Long-running Hermes/Moltbook ingest builds statistical fingerprint per key_id. Later partial repost detected via statistical tier.

Architecture

routes/grain.py
    └── GrainEngine (core/grain/engine.py)
            ├── exact.py      → core/signing.py
            ├── redundant.py  → exact × N
            ├── resonant.py   → core/echo.py + source reputation
            └── statistical.py → n-gram chi-squared

Philosophy

Any verification is signal detection in noise.

QR: digital signal — 100% or 0% at the cliff edge
Signature / SynthID: analog signal — delta distance from reference, FAR/FRR tradeoff

INDB's Axiom layer stays exact. Grain makes the layers above it honest about uncertainty — without pretending memory is a barcode.