feed-protocol 0.2.0

FEED — Format for Enforced Evidence-based Digestion. A self-bootstrapping document protocol that forces downstream LLMs to ground answers in cited evidence.

FEED — Format for Enforced Evidence-based Digestion

Make any document tell the AI reading it: here's what matters, here's the evidence, cite it or say you can't.

FEED is a plain-text convention you embed in a document so that any LLM — Copilot, ChatGPT, Claude, Gemini, a local model — grounds its answers in your evidence and cites it by ID. No install, no plugin, no provider support, no prior knowledge of FEED required on either side. It works on every model today because the document teaches the rules inline.

It also has teeth: because evidence carries stable IDs and answers must cite them, you can mechanically verify that an answer is actually grounded.

┌─ Header ──────────────────────────────────────────────┐
│ <!-- FEED:DOC version="0.2" grounding="strict" -->     │
│ > AI INGESTION NOTICE … ground answers, cite [E###] …  │  ← teaches any LLM the rules
├─ Tier 0: Claims & Decisions (front-loaded) ───────────┤  ← small-context-safe
├─ Tier 1: Findings (narrative, references [E001]) ─────┤
├─ Tier 2: Evidence (atomic key/value facts, IDs) ──────┤  ← the source of truth
└────────────────────────────────────────────────────────┘

Why it exists

You generate AI reports and send them to people who paste them into their own AI. Every hop loses fidelity — the reader's AI skims headings and riffs. FEED fixes the author side of that loop: the document constrains how the downstream AI reads and answers. Nothing else does this at the document level (llms.txt is website-level and has no grounding contract; RAG/citation systems are all retrieval-side, controlled by the AI, not the author).

FEED is AI-to-AI (the library never needs its own API key)

The AI already in your loop — the one that wrote the report, your assistant, a pipeline step — is what produces FEED. The library just renders, validates, and verifies; it never calls an LLM of its own. Authoring is self-bootstrapping, the same way ingestion is:

Reading side:   the document carries a notice that teaches any AI to ground & cite
Authoring side: an authoring prompt + schema teach any AI to emit FEED, then
                feed.build() renders it deterministically — no key, no network

Primary flow: the AI emits FEED, you render it

from feed import AUTHORING_PROMPT, FEED_JSON_SCHEMA, build

# 1. Hand AUTHORING_PROMPT + FEED_JSON_SCHEMA to whatever AI is already in your loop.
#    It returns structured JSON (no FEED tooling needed on the AI's side).
# 2. Render that JSON into a validated FEED document — pure Python, no API key:
doc = build(ai_json, grounding="strict", author="N. Gul")
doc.write("report.md")

Or entirely from the shell:

feed prompt > authoring-kit.txt        # the prompt + schema to give any AI
feed build ai_output.json -o report.md # render the AI's JSON into FEED (no key)

Manual additions are just edits: open report.md and add/adjust evidence and claims by hand — it's plain markdown. feed validate checks it's still conformant.

feed tag draft.md is an optional convenience that calls Claude directly, for when you have a plain document and no AI already in the loop. It is not the primary path and is the only thing that needs an API key.

Quick start

Build one in Python (manual / programmatic)

from feed import FeedDocument

doc = FeedDocument("Q2 Pump Health Assessment", grounding="strict")
doc.add_evidence("E001", asset="XYZ-003", metric="vibration_rms",
                 value="12.4 mm/s",
                 threshold="11.2 mm/s (ISO 10816-3 Zone C)", confidence="high")
doc.add_claim("C1", "XYZ-003 needs intervention", evidence=["E001"],
              decision="Approve bearing replacement work order")

doc.write("report.md")    # clean markdown, FEED in HTML comments
doc.write("report.html")  # styled, opens in any browser

Your team opens report.md as a normal report. They upload it to whatever AI they use → it reads the notice, sees the evidence, and answers grounded.

Verify an answer was grounded (the teeth)

from feed import FeedDocument, verify

doc = FeedDocument.read("report.md")
report = verify(ai_answer_text, doc)
print(report.passed)   # False if it cited evidence that doesn't exist,
                       # or (strict mode) didn't cite anything

From the command line

feed prompt                                     # authoring kit for any AI (no key)
feed build ai_output.json -o report.md          # render an AI's JSON into FEED (no key)
feed validate report.md                         # is it well-formed FEED?
feed verify --doc report.md --answer answer.txt # is this answer grounded?
feed render report.md --to html -o report.html  # styled HTML
feed tag draft.md --grounding strict -o report.md  # OPTIONAL: auto-tag via Claude (needs key)

What's in this repo

Path	What it is
spec/feed-spec-v0.2.md	The protocol definition (the constitution)
feed/	The reference library — build, render, validate, verify, auto-tag
feed/verify.py	The citation verifier — FEED's defensible edge, ~40 lines
feed/cli.py	The feed command-line tool
examples/	A complete worked example: a pump condition report (.md + .html) and the script that builds it
templates/blank.feed.md	A hand-authoring starter
tests/	Round-trip, validation, and verification tests

Install

pip install -e .            # library + CLI, zero dependencies
pip install -e ".[tagger]"  # adds the auto-tagger (needs anthropic + ANTHROPIC_API_KEY)

The core — authoring kit, build, render, validate, verify — is pure Python with no dependencies and never calls an LLM. Only the optional tag convenience calls Claude directly; it defaults to Claude Opus 4.8.

The three primitives

• Evidence — atomic, ID'd, key/value facts. Never prose. The source of truth.
• Claim — a short statement grounded in evidence IDs, optionally a decision.
• Header — declares the grounding mode and carries the self-teaching notice.

Plus grounding modes (strict / standard / open) — the author's dial for how strict the reading AI must be. In strict, no evidence means "Not supported by this document."

Status

v0.2 — spec + reference library + verifier + CLI + auto-tagger + worked example. Deliberately out of scope for now: PDF/DOCX embedding, a hosted validator, and any provider-native "FEED mode".

MIT licensed. Spec and tooling are open — adoption is the point.