QUANTA · BOTH SIDES OF THE COIN · DRAFT · FOR RED-PEN

A claim that binds

We ran two experiments and only remembered one. The first proved a stamped claim changes how an agent works. The second proved our stamps are mostly wrong. Two halves of one coin — and we have been spending only the good half.

Authored by Fable in-loop · grounded in EXP-TB Part B (reader) + A1 (writer) · draft for annotation, nothing built yet
01

The story

The reader half worked. Put a claim in a builder’s workspace carrying four things — what is claimed, the command that re-proves it, the files it rests on, and a stamp saying whether it is still fresh — and the builder changes behaviour without being told to. Redundant re-checking collapsed about threefold. Skepticism landed exactly where it should: stale-stamped claims were re-verified 6 times out of 6. Quality did not move.

The writer half failed. We planted 28 changes that genuinely made a banked claim false, and asked whether the claim’s own receipt would catch it. It caught 5. Twenty-three of twenty-eight falsifications left the receipt green.

Why: most of our claims are guarded by hermetic unit tests that stub out the very production code the claim is about. The instrument reports the basis is complete. It is complete, faithful, and claim-useless.

Put the halves together and there is only one sentence: agents will trust the stamp exactly as much as we hoped, and the stamp is currently right about one time in five.

That is worse than it sounds, because the same experiment measured the failure mode. A check command carrying untrustworthy freshness produced more pointless re-checking than plain prose — 12/24 against 0/24. We are not one step from the win. Ship delivery on top of today’s corpus and we ship the arm that measured worse than saying nothing.

02

The field we never recorded

A claim today carries: the text · the receipt command · the basis files · the freshness stamp. It does not carry the one property that decides whether any of those matter:

Does this receipt actually go red when this claim becomes false?

Call it binding. It is measurable. We already built the instrument that measures it. We have never once recorded the answer on a claim.

This is the same disease as everything else caught this week — an instrument that cannot fire, reporting success. Two vacuous gates. A download counter that logged nothing. A drift detector never wired. A map served and ignored. A claim whose receipt cannot go red is that same defect one layer up: knowledge that cannot be wrong, which is knowledge that cannot be trusted.

03

What I decided, and why

  1. Binding becomes a measured field, not an assumption.
    Every claim records how many planted falsifications its receipt actually caught. A claim never tested renders as UNPROVEN — never as trustworthy. We do not get to assume our proofs prove things.
  2. The reader is told.
    The surface already shows FRESH/STALE. It gains one column: BINDS 8/10. That is the whole design in one move — the proven reader mechanism, pointed at a signal that is finally honest. Trust becomes calibrated instead of hoped for.
  3. Measure before building anything.
    Slice 1 is a number, not a feature. If almost nothing binds, we learn it in days for the price of compute, and this dies at its first gate rather than after a month of plumbing.
  4. Flip the producer, and accept fewer claims.
    The discriminator is measured: the broad integration receipt traced 59 modules and bound its claim; the hermetic claims traced 2–5 files and did not. So we shed from certification runs and integration gates, not cheap unit tests. Twenty binding claims beat one hundred and fifty unbinding ones, because the reader effect is per-claim, not per-corpus.
  5. Binding is necessary, not sufficient.
    A claim can bind perfectly and still be worthless if no builder would ever have checked it. Binding is the floor for trust. Usefulness is a separate bar and stays a judgment call.
04

What could go wrong

The binding test is itself vacuous
A “plant” that does not really falsify the claim. Guard: every plant carries a falsification receipt proving the claim is now false, mechanically. Inherited from A1, not re-invented.
Goodhart
Work shaped to produce high-binding claims rather than useful ones. Guard: we measure binding on receipts that already exist for their own reasons, not on work invented to score.
Broad receipts are expensive
True and unavoidable — the tension the whole spec sits on. The answer is fewer claims, not cheaper proofs. A small binding corpus is the product.
Trust may not transfer
Agents might ignore the new column. Possible. Slice 4 measures it and can kill the arc.
05

The build — four slices, in order

  1. Score the corpus. Run the plant-a-rot battery over every banked claim; publish the binding distribution. Proves: whether binding claims exist at all today. Cost: days, no build.
  2. The binding column. Record measured binding; render it; UNPROVEN renders as UNPROVEN. Proves: the reader can see what is trustworthy. Cost: small.
  3. Flip the producer. Shed from certification, integration gates, indexing. Stop shedding from hermetic tests. Proves: binding claims can be a byproduct of real work. Cost: medium.
  4. The reader A/B, gated on binding. Same brief, two workers, binding-verified claims in one. Proves: the effect survives contact with production. Cost: one night.

Slice 1 gates everything. Slice 3 is the prerequisite for the continuous-certification ceiling the other fleet found: a claim that binds and is broad is one that can be held now rather than recently.

The kill, pre-committed: if agents re-check binding-verified fresh claims at the untrustworthy-arm rate (≈12/24), then trust does not transfer through the binding column, and the corpus concept needs a different mechanism entirely. That result stops the arc — and is worth knowing.

Your calls

1 · Do we bank untested claims at all?

Or refuse until they have been binding-tested? My recommendation: bank them, rendered honestly as UNPROVEN. Refusing empties the shelf today and we lose the reader arm entirely. An honest “we have not tested this” is still information.

2 · Is a 20-claim binding corpus acceptable in place of today’s 150?

My recommendation: yes. The measured reader effect is per-claim. A hundred and fifty claims that cannot go red are not a shelf — they are furniture.

3 · Score first, or build the column first?

My recommendation: score first. It is the only step that can kill the arc cheaply, and its result changes what the column should say.

Canonical source: docs/ai/SPEC-both-sides-of-the-coin.md — machinery, the mechanical definition of binding, and the anti-vacuity control on the instrument itself live there.