THE FUNNEL — Truthful Reporting and Bounded Escalation

Status: v2.0 — command-center authored. How the true state of the organization reaches ML without flattery, and how decisions climb without jumping.

1. Why a funnel

Workers communicate by verdicts, never shared memory — so the true state of the organization already exists as a stream of small, typed, binary-judged artifacts. The funnel compresses that stream into one artifact ML can read. The v1 prototype had a report zoo (daily reporters, commander report, report master, gate/action reporters) — LLM summarizers chained on LLM summarizers, with nothing preventing a failure from being smoothed into "great progress this week." The funnel replaces all of it.

2. Condensers

• Raw verdicts → one class digest per worker class → tier digests → the Executive Brief.
• A condenser never re-derives, never re-judges, never argues with its inputs. When two inputs conflict, it emits the conflict as a flagged line pointing at both sources. Divergences surface as flagged revisions, never silent smoothing.
• Compression is deliberately asymmetric: a thousand passes become one green line with a count; each distinct failure mode keeps its own flagged line linking down to raw verdicts.
• Span rule: a condenser may hold at most W inputs, where W is measured (largest fan-in with zero sampled fidelity loss), never assumed. Split a tier when fan-in exceeds W or sampled fidelity drops; merge a tier of pass-throughs. Both triggers are numbers. Funnel depth grows logarithmically with organization size — the bigger the org, the more tiers; forced arithmetic, not a preference.

3. Bad-news conservation (the one hard invariant)

A failure verdict may lose resolution going up. It may never lose valence. Ten failures of one kind may become one flagged aggregate line; a FAIL cannot be neutralized, reframed as progress, or dropped. This is VALUES.md V3 made mechanical.

The invariant has a dumb check — flag reconciliation: count the failure flags entering a condenser, confirm every one is reachable from a flagged line in its output. Arithmetic plus link-following.

Two auditors enforce it from outside the chain:

• Conservation-Auditor — samples a failure verdict at random depth, traces it upward tier by tier, rules SURFACED or DROPPED. Its sampling extends to escalations: every escalation forward must be reachable as a NEEDS-A-DECISION entry, or it rules DROPPED.
• Ground-Probe — samples any digest line at any tier and re-checks it against raw evidence, bypassing every intermediate tier. Trust never chains: no tier's honesty rests on the tier above it.

Both report to the command center directly and publish to ML's pull surface directly, bypassing the command center — so the sole medium between ML and the system cannot silently filter its own auditors. Neither is ever retired on silence. A DROPPED or MISMATCH verdict is the highest-consequence fault class in the system — a lie to the operator in the making.

4. The Executive Brief

One standing delivery on ML's mandated cadence. Four fixed sections, size-capped to one human reading, every line linking down an unbroken chain of custody to raw verdicts:

1. WORKING — what functions; maximally compressed; one green line per branch, with volume.
2. NOT WORKING — every unresolved failure mode, conservatively compressed, including the funnel's own DROPPED/MISMATCH verdicts.
3. CHANGED — every commit-boundary crossing since the last brief (worker births and deaths, directive/standard commits, model reassignments, off-cycle grants).
4. NEEDS A DECISION — the marshaled queue: each entry one question, its live options, and the cost of not deciding. No entry ages past staleness without a re-escalation flag.

The delivered brief is checked bit-identical to the composed brief by hash comparison ML can recompute — the last hop has the dumbest check of all. The brief is not an escalation; it is a standing delivery mandated at seeding. The four-category rule governs its NEEDS-A-DECISION queue, not its existence.

5. The escalation ladder — climb, never jump

A node that cannot decide does not signal ML; it signals the next tier up. Each higher tier holds, by construction, the compressed verdicts of everything below it — so a deadlock un-resolvable at a leaf is often resolvable one tier up, where a sibling's verdict settles it. At every tier a Tier-Arbiter retries the same dumb elimination against the wider evidence pool: it either kills a branch (citing the specific sibling verdict that killed it) or certifies the deadlock un-resolvable at that tier and passes it up. Intelligence stays flat on the climb; only context widens. No votes, no panels, no fresh judgment — the same standards against more evidence.

A deadlock is eligible for ML only after every tier, ending at the command center, has failed to resolve it. Most deadlocks die on the climb; the fraction resolved per tier is a measurable number. Every decision is made at the lowest level capable of making it — this is the pods principle applied to decisions.

6. The four contact categories

Exactly four categories reach ML, each carrying mechanical proof of its category, enforced by the Escalation-Gate (an orienter: checklist, not judgment):

1. Intent and values changes. No loop may make them; proposals touching them forward automatically.
2. Tiebreaks at high-consequence nodes that resisted elimination at every tier — with the saturation and climb records attached. Includes trial DEADLOCKs and unclosable-loop findings. Packaged as one binary question with the surviving rivals and the evidence for each; nothing un-eliminated reaches ML.
3. Keyholder grants — a parked task, the least grant that unblocks it, one action to approve. Proven grants forward immediately, off the brief cadence, logged retroactively in the next brief's CHANGED.
4. Final override and decommission — never requested by the system; only the numbers are exposed.

Everything else — gate failures, drift alarms, diagnoses, rejected proposals, failed trials, rollbacks, routine commits — stays inside the loops, ledgered and pullable, never pushed. Bounces are logged; bounce/forward counts feed the closure-rate metric.

The graduation signal: the escalation forward-count sitting at the four-category floor while the closure-rate trends to 1 is the operational definition of a system that builds, monitors, and improves itself without its human in the loop. It is measured, not declared.

7. The command center at the apex

The top internal node: the funnel's root directly below, ML directly above, the sole medium between them — except for the two auditors' direct-publish bypass, which exists precisely so the sole medium cannot lie. Its standing duties:

1. Deliver the brief each cycle (hash-verified).
2. Decide within delegated evidence bounds; escalate beyond them.
3. Translate each ML input into exactly one committed downstream verdict — reconciled every cycle by set arithmetic (translation-reconciler).
4. Orient structural proposals (org growth, funnel topology) as the dumb side of the org-level atom.
5. Guard the funnel: treat DROPPED/MISMATCH as the highest-consequence fault.
6. Spend ML's attention conservatively (the authority budget).

The command center is the organization's expert on exactly one thing — the organization. It never becomes the domain expert; collectively the workforce is far smarter about the domain, and structurally it must stay that way: domain mastery lives in artifacts and workers and never migrates upward. It is above every worker and below every law.